Hi friend,

It’s hard to develop a defensible mental health business. 

Many solutions are commoditised. Network effects, scale economies and switching cost are rare.

One of the few ways you can develop defensibility is through scientific credibility. The businesses with credibility close deals faster and raise more capital. And as they do this, they gain the resources and reach that allow them to develop more evidence and more credibility. Their defensibility compounds.

This asset of credibility is built by generating evidence that holds up under scrutiny. 

When you can make defensible assertions about what your product does, you will win over stakeholders and close deals. If you can then build a collection of defensible evidence, your scientific credibility is elevated to a strategic asset. You become known as a business that can be trusted to deliver on what you put forward. And that brings success, especially in a market that is increasingly scrutinising solutions.

But when your evidence isn’t defensible, stakeholders poke holes and conversations quickly grind to a halt. It’s a horrible feeling to be sitting in a conference room and watch a deal die as you’re asked to defend a position you knew was shaky. When you find yourself in this situation, it’s already too late. 

In this Hemingway Guide, Jen Huberty PhD joins us to provide a framework for building scientific credibility as a mental health business. Jen has spent over twenty years in both academic research and commercial digital health. She was Head of Science at Calm and has advised several of the world’s leading digital health businesses.

Use this framework to strengthen your evidence, close more deals and develop scientific credibility as a strategic asset.

Key Takeaways

1. Scientific credibility is a company's ability to defend its evidence with structured, stage-appropriate support that withstands stakeholder scrutiny.

2. It primarily operates at an assertion level — your ability to defend specific assertions.

3. Done repeatedly and with discipline, it can become a strategic asset

4. Scientific credibility requires four interdependent foundations: Problem, Population, Solution, and Outcomes. Weakness in any area undermines the entire structure.

5. You can use these four foundations to conduct a gap analysis before stakeholders expose weaknesses. Investors, enterprise buyers, and partners will find the holes eventually.

6. You should prioritise the assertions you want to make defensible by understanding which are critical to your near-term business goals.

7. Strong outcome data alone leaves critical questions unanswered: "Works for whom?" "Why does this problem matter?" "Is your solution what's creating these results?”

8. Different stakeholders probe different foundations; balanced evidence matters more than deep evidence in just one area.

9. Evidence expectations evolve with stage of your business. What's appropriate at pre-seed differs significantly from what's expected at Series A or enterprise scale. The goal is not to reach RCT-level rigor, it's to have the right evidence for the decisions your stakeholders are making right now.

10. Founders who build complete evidence before competitors gain a structural advantage. 

Building Scientific Credibility

Scientific credibility is a company's ability to defend its evidence with structured, stage-appropriate support that withstands stakeholder scrutiny. It is not defined by publications or methodological complexity; it is defined by whether your evidence holds up when the people making buying, funding, or partnership decisions push back on it.

It operates at two levels:

The assertion level. Everything a company puts forward about the problem it solves, the population it serves, how its solution works, and what outcomes it produces should be defensible across the four foundations. 

The company level. Credibility at the company level is the cumulative result of scientific discipline applied consistently over time. 

In this Guide we provide a framework for developing credibility in your assertions. You can use it to stress-test specific positions (before stakeholders do). Do this consistently and with discipline and the company-level asset will follow.

The Four Foundations of Scientific Credibility

Scientific credibility is built on four distinct foundations. They are defensible independently and when combined, reinforce one another.

• Problem: The problem you’re tackling is real, defined, and quantified for a specific group

• Population: The evidence reflects the actual people you're trying to support

• Solution: Your product specifically drives the change, not something else

• Outcomes: The improvement is measured, meaningful, durable, and aligned with what your buyer cares about

Figure 1: The Four Foundations of Scientific Credibility

When you have all four foundations for an assertion, you can say; this is a real, quantifiable problem with direct health and economic impacts for this specific population and our solution solves this problem through this precise mechanism, resulting in measurable improvements in the metrics you care about.

Example: "Untreated anxiety costs U.S. employers $4,400 per affected employee annually in lost productivity, disproportionately impacting hourly frontline workers with limited access to traditional therapy. Our platform delivers asynchronous CBT through a guided self-pacing model that produces a 40% reduction in PHQ-9 scores within eight weeks, sustained at six months".

Each foundation reinforces and strengthens the others, creating a more defensible and resilient evidence base. Together, they can support your business decisions and build your scientific credibility.

Building an evidence base can be expensive and takes time. Therefore, the best organisations know how to do two things well when it comes to scientific credibility;

1. They understand which assertions they need to make to support their business goals.

2. They know how to make those assertions highly defensible with strong evidence generation across all four foundations.

Prioritise the Assertions that Support your Business Goals

Figure 1 above showed how the four foundations apply to any single assertion. But most companies are defending multiple assertions simultaneously, each with different levels of evidence. Figure 2 (below) shows what that looks like in practice. 

In reality, a company may have many more assertions than this, for different products and different stakeholders.

Thankfully, not all of these need to carry equal weight at any given moment. You need to understand which ones are critical to your near-term business goals and make sure those are structurally sound.

To do this, work backwards from your business strategy. Ask yourself, what business goals do I need to hit in the next twelve months and what will I need to defend to achieve them? Those are your priorities and should be the focus of your evidence generation strategy.

If your goal is to land five new enterprise contracts with employers, your assertion about reducing absenteeism is the one under pressure. If you only have problem and population evidence but no solution attribution or outcome data, that's your vulnerability. It doesn't matter that your investor-facing evidence is complete. The assertion that's being tested is the one that needs to hold.

Build in the Right Order

It’s unrealistic to build all four foundations simultaneously across every assertion or business decision. And you don’t need to. The ideal sequence depends on your stage and the conversations that are most critical right now. 

Early stage (pre-seed / seed). You might have evidence across one or two foundations for a single assertion. That’s OK. No one expects a complete evidence base at this stage. Focus on having strong evidence for a specific problem and population.  You need to prove the problem matters and understand who experiences it before you invest heavily in solution development.

Series A. You need at least one assertion that meets all four foundations, and credible foundations progress on other key assertions Solution evidence should be emerging from product development and feasibility testing. If you're planning a raise in twelve months, outcome data collection should start now.

Growth stage and enterprise scale. You need evidence across multiple assertions because you're selling to multiple stakeholder types simultaneously. The employer buyer, the health plan, the health system — they each care about different claims, and each of those claims needs to hold up independently.

When we worked with a national virtual mental health provider, they had strong problem validation, clear population definition, and a well-articulated solution, but needed rapid outcome evidence to support their growth stage. We analyzed 5+ million therapy visits and published four papers in three months showing clinically meaningful symptom reduction. That completed their evidence at exactly the right time for their business needs.

Using the Four Foundations to Build Defensible Assertions

Now that you know which foundations  to prioritise and when, the rest of this guide is a detailed framework for assessing each one. Use it to assess the strength of your evidence within your most critical business conversations, and to identify where your evidence needs to be stronger. 

Foundation 1: The Problem

Do you have evidence that the problem you are solving actually exists and actually matters?

Many founders start a business to solve a personal pain point. But personal experience or anecdotal evidence are not the same as defensible problem evidence. You need to be able to clearly define the problem, its prevalence, the impact of that problem in health and economic terms, and why existing solutions don't solve it.

What strong problem evidence looks like

Strong problem evidence rests on a small set of defensible components:

Prevalence data showing how many people experience the problem. Strong sources include peer-reviewed epidemiological studies, CDC and WHO surveillance data, large national surveys (e.g., BRFSS, NHANES), and some condition-specific registries. Avoid white papers or market research reports that cite prevalence without primary sources.

Impact data describing consequences in health, economic burden, or quality of life. Look to health economics literature, employer claims databases, and government cost reports (e.g., AHRQ, CMS). Self-reported impact data from your own users can supplement these sources but shouldn't be used as your primary source.

Gap analysis clarifying what current solutions exist and why they fall short. Strong support comes from systematic reviews and meta-analyses identifying intervention gaps, published critiques of current standards of care, clinical practice guidelines that acknowledge unmet needs, and competitive landscape analyses grounded in peer-reviewed efficacy data. High-level competitive summaries or marketing claims about differentiation are not sufficient on their own.

Stakeholder perspective confirming that users define the problem the way you do. Examples include structured needs assessments, qualitative research with target users and buyers, and published patient/member experience data. Anecdotal feedback from a handful of users is not stakeholder validation.

Prevalence and impact are only persuasive when they are clearly defined and specific. A problem may be widespread, but severity, contributing factors, and consequences often vary meaningfully across subgroups. If the problem is not clearly defined for a specific population, evidence becomes easy to challenge.

For example, a digital mental health company with strong UK adoption entered the U.S. market expecting their existing evidence to transfer. It didn't. Prevalence patterns, access barriers, and stakeholder expectations differed enough that general UK data wasn't sufficient to support adoption in this new market. They needed to develop U.S.-specific problem evidence, establish a U.S. advisory board, and publish research in the U.S. before the problem could be defended credibly with American buyers.

You must be able to clearly demonstrate that the problem is real, meaningful, and unresolved. Before including a new source in any material, the strongest teams pressure-test each one by asking: does this materially strengthen our ability to defend the urgency and specificity of this problem in a real business conversation?

When problem evidence is weak

There are several red flags that signal weakness in your problem foundation. One is when teams default to saying "everyone knows this is a problem" without quantifying the prevalence or impact within your defined population. It's also a red flag when a company describes the problem differently to how their target users do. And it's a mistake not to provide a concise articulation of why existing solutions fail.

Consider a stress management platform that cites national data on rising stress and burnout. That data is accurate. But when the company enters enterprise conversations with their target clients, the HR leaders they're selling to are focused on absenteeism, retention, and performance variability in specific teams. The company never clarified which subgroup experiences the most severe impact or how current solutions fail them specifically.

Without a defensible problem definition, messaging becomes generic and sales conversations stall under scrutiny. If stakeholders can't see why your solution should even exist in the first place, there's no point in talking about anything else.

Join the Hemingway Community

If you’ve read this far it’s probably because you want tactical information on building a successful mental health organisation? If you’d like more of that, and to meet peers building similar businesses, consider joining the Hemingway Community.

As a member of the community, you’ll join over 450 other mental health leaders and get access to exclusive content, events and resources.

Foundation 2: The Population

Does your evidence reflect the specific population you intend to serve and have you designed your product with them in mind?

Solutions that are effective for one population may fail for another. Stakeholders want to know whether your data extends across their broader employee, patient, or member population. If you cannot clearly define who your evidence applies to and who it does not, stakeholders will question its generalizability. Precision builds trust. Ambiguity weakens it.

What strong population evidence looks like

Strong population evidence typically includes clarity across four dimensions: demographics (age, gender, socioeconomic status, geography), clinical characteristics (severity, comorbidities, treatment history), behavioral patterns (health literacy, technology comfort, engagement barriers), and contextual factors (cultural considerations, access constraints, support systems).

It's about understanding the real-world conditions that shape whether your solution will be adopted and effective. This evidence is usually built through structured user research, disciplined customer discovery, segmentation analysis, literature on population-specific intervention effectiveness, and internal data confirming that your intended population is engaging as assumed.

The most credible population evidence moves beyond description into applicability. Demographic summaries or personal documents are not sufficient. Stakeholders want to see that your data reflects the population you are targeting and that engagement, usability, and early signals are consistent within that segment. The goal is distinguishing between who shows interest and who demonstrates sustained fit.

As an example, a pre-seed company ran a feasibility study with 19 patients and conducted focus groups with eight users to understand feature needs and retention drivers within that population. They also interviewed four healthcare providers to validate clinical fit. They came out of that knowing who their product may work for, what those users needed, and whether clinicians would endorse it for that group. The feasibility study became the foundation of their pitch deck and go-to-market messaging and played a key role in securing funding.

When population evidence is weak

The red flags are easy to spot. The product is described as broadly applicable — "our product is for everyone". There is no clearly defined core user. There are significant differences in engagement across demographic or clinical subgroups but they are unexplored. There is little data showing reach among the highest-need populations.

When population specificity is weak, the consequences tend to show up when you try to expand beyond your initial user base. The claims you used for your original users don't extend to the new group. This doesn't mean companies should avoid expanding into new populations. It means the evidence strategy needs to expand with them. Before entering a new segment, teams should understand the key differences between the population they have data on and the new population they're targeting, and have a plan for testing whether their core assumptions hold. Going into a new market with a credible evidence plan is a fundamentally different conversation than discovering misalignment after the deal is signed.

Foundation 3: The Solution

Does your specific solution solve the problem?

Improvement can occur for many reasons: natural symptom fluctuation, regression to the mean, concurrent treatments, or increased motivation simply from enrolling in a program. If you cannot demonstrate how your solution specifically drives change, your claims become vulnerable. In a crowded digital health market, attribution is often what separates credible differentiation from commoditized claims.

Another example… An AI mental health platform showed that over 60% of sessions produced measurable mood and stress improvements. On the surface, that’s promising. But mood naturally fluctuates, and users tend to initiate sessions when they're already feeling low, meaning some recovery would be expected regardless of the intervention. Without a comparison condition or a way to account for natural mood recovery, the question of attribution remained open. The results were difficult to defend in investor, partner, and buyer conversations. For this company, strengthening that link became the focus of their evidence strategy.

What strong solution evidence looks like

Before you claim impact, you need to articulate why your solution works. Strong solution evidence creates a defensible causal narrative, linking product design decisions to measurable change.

A clearly defined mechanism of action. This goes beyond naming a modality. It explains how specific features or interactions lead to measurable change. What is the behavioral, cognitive, or physiological pathway? How does engagement translate into outcomes?

Example: One youth mental health platform didn't just claim that therapy reduces symptoms. They articulated a specific multidisciplinary model combining individual therapy, psychiatric support, and family involvement and demonstrated how that integrated pathway led to symptom reduction in 80–95% of youth served. The mechanism wasn't "we offer mental health support." It was a defined clinical pathway with evidence for how each component contributed to change.

Clarity around active ingredients. Not every feature drives impact equally. Credible teams examine which elements are essential versus supportive and test whether modifying or removing components changes outcomes. This separates signal from noise.

Example: A mental fitness app offered cognitive games, guided exercises, personalized training, and progress tracking. Through structured evaluation, the team developed a scientific framework that identified which specific activities drove measurable cognitive improvement and which were supportive but not essential. That clarity allowed them to prioritize development resources, refine their product around what actually worked, and defend their differentiation with evidence rather than feature lists.

Dose-response patterns. If the solution functions as intended, greater engagement with core components should align with stronger outcomes. While dose-response does not prove causation alone, it strengthens the attribution argument.

Example: A virtual mental health platform analyzed data from millions of therapy visits and found that symptoms of anxiety and depression dropped meaningfully within five sessions, with continued improvement over time. Among patients with severe symptoms, the majority experienced clinically significant reductions within eight weeks. The pattern was clear: more sustained engagement with therapy sessions aligned with stronger outcomes, reinforcing that the platform's core service was driving change.

Comparative context. Improvement only means something relative to an alternative. Without a reference point, observed change is difficult to interpret. This may involve matched cohorts, historical baselines, comparisons to standard care, or industry benchmarks.

Example: A PTSD treatment platform analyzed outcomes from 148 adults receiving massed cognitive processing therapy — an intensive, compressed format — compared to the traditional weekly pacing of the same evidence-based treatment. Results showed significant improvement in PTSD symptoms, anxiety, and depression, with gains maintained at three months post-treatment. By comparing their delivery model against the established standard, the company could demonstrate that their approach offered a meaningfully different path to the same outcome.

When these elements align, you deliver a defensible explanation of how and why users improved. That explanatory clarity is what differentiates credible innovation from commoditized digital health tools.

When solution evidence is weak

When the solution is weak, differentiation becomes fragile. You may show improvement, but you cannot explain why your solution works or which components are responsible for outcomes.

A digital health company with 465 members had assessments that were not built on any scientific framework or foundation. There was no standardized member journey connecting solution components to expected outcomes. Without that, the company could not demonstrate which elements of their program were driving change. This made their evidence both difficult to explain and easy to challenge in stakeholder conversations.

The red flags are clear. Teams say "users improve when they engage," but cannot define what meaningful engagement actually is. Every feature is treated as equally important. Active ingredients are not identified. Improvement is assumed to be attributable to the solution without demonstrating causal linkage. Under scrutiny, the evidence looks correlational rather than product-specific. Without clear attribution, credibility and differentiation both weaken.

Foundation 4: The Outcomes

Can you demonstrate measurable, meaningful improvement in health, behavior or economic data?

This is often the first question stakeholders ask — does it work? Outcome data is important. But impressive numbers built on a vague problem definition, an undefined population, and no attribution mechanism will collapse under scrutiny.

A common mistake is focusing on collecting outcome data before creating the infrastructure needed to measure it properly. In these cases, teams collect a lot of data, but it's the kind of data that doesn't withstand deeper questioning.

What strong outcome evidence looks like

Strong outcome evidence rests on a disciplined measurement strategy. It is not just about showing change. It is about showing change that is interpretable, meaningful, durable, and aligned with stakeholder priorities.

Use of validated measures. Proprietary indices or internally created metrics may be useful operationally, but stakeholders expect standardized instruments when evaluating clinical or behavioral change. Learn more about validated measures in our other Hemingway Guide.

Appropriate measurement timing. Outcomes should be assessed when change would reasonably occur. Too short, and results lack interpretability. Too long without follow-up, and durability becomes unclear.

Clarity on meaningful change. A result can show measurable improvement without that improvement being large enough to matter in real life. Stakeholders want to understand whether the magnitude of change crosses clinically meaningful thresholds and translates into real-world impact.

Durability of effects. Short-term gains may reflect novelty or initial motivation. Sustainable credibility requires evidence that improvements persist beyond the initial intervention window.

Alignment with stakeholder priorities. If enterprise buyers care about absenteeism, productivity, or healthcare utilization, those outcomes must be visible in the evidence. If payers focus on risk reduction or cost offsets, measurement strategies should reflect those endpoints. Outcomes that are not stakeholder-aligned are easy to dismiss, even when statistically robust.

Proportional evidence progression. Early-stage companies may begin with self-reported improvements or engagement correlations. As the company grows, evidence should evolve toward structured pre-post change using validated instruments, then toward comparative effectiveness, durability of effects, and clinically meaningful thresholds. Understanding this progression helps teams calibrate expectations while strengthening credibility over time.

When outcome evidence is weak

When this foundation is weak, growth conversations slow. Engagement metrics may look strong. Satisfaction scores may be high. But measurable health improvement remains unclear.

A weight management platform reports modest short-term weight loss and high user satisfaction. When speaking with a self-insured employer, however, the buyer asks about sustained weight reduction, cardiometabolic markers, and impact on healthcare utilization. The company cannot demonstrate durability of effects beyond the initial program window. Outcomes rely on proprietary or non-validated instruments rather than standardized measures. There is little evidence that the reported changes align with the buyer's cost or utilization priorities.

The warning signs are not always about missing data — they are about weak interpretation and misalignment. Short-term gains are emphasized without follow-up data. Proximal indicators such as knowledge or awareness are measured instead of meaningful behavioral or clinical outcomes. Best-performing users are highlighted rather than average effects. Clinical significance is not distinguished from statistical significance. When outcome evidence cannot withstand deeper questioning, confidence weakens quickly.

Conclusion

Executing this strategy takes discipline. The discipline to know which assertions you need to invest in and which to deprioritise. The discipline to build evidence across each of the four foundations and to invest in the infrastructure required to do so. It also required the discipline not to make statements you can’t defend. The reward, however, is worth it. Deals will close faster and over time, you’ll become known as a trusted and credible player in the mental health ecosystem. Over time, this will become a strategic asset that creates defensibility. And in this market, that is an important asset.

That’s all for this edition of The Hemingway Report. Many thanks to Jen for sharing these hard-earned insights with us. If you found this Guide valuable, consider supporting our work by becoming a Hemingway Pro Member and joining our community of mental health innovators.

Keep fighting the good fight!

Steve

Founder of Hemingway

You know this used to be a sheep farm?

I didn’t. 

To be honest, I didn’t know much about the Allied Arts Guild of Menlo Park. What I did know was that the best and brightest of mental health innovation were about to descend on the venue for the 2026 Demo Day of the One Mind Accelerator (OMA). Over the next twenty-four hours I’d hear from fourteen of the world’s top mental health founders and chat with the payers, policy makers and investors shaping this industry.

But at this moment, I was sitting on a wooden bench listening to the birds and learning about sheep farms from the gardener.

Well, more specifically I was learning how a young couple, Delight and Garfield Merner, had bought this land in 1929 when it was a three and a half acre sheep farm on the edge of Menlo Park. The Merners had been inspired by the craft guilds that had popped up in Europe as a reaction to industrialisation. The leaders of this movement were appalled by the lack of humanity in industrial era products and wanted to restore a sense of craft and artisanship. Their answer was to create craft guilds; associations of craftspeople who worked under shared standards, trained one another through apprenticeship, and protected the quality and dignity of their trade against industrial market forces.

The Merners wanted to bring this model to California and buying a sheep farm was their first step in doing so. On this plot of land, they designed a Spanish Colonial themed campus and in 1929 opened the Allied Arts Guild. Almost one hundred years later, dozens of craftspeople come to this space every day, to create, to collaborate and to teach curious Irishmen about the history of this special place.

On this sunny California day, while the painters painted and the potters pottered, another group of passionate individuals were taking their seats in the main hall of the Allied Arts Guild. The fourteen founders of One Mind’s 2026 Accelerator cohort had arrived and the audience were keen to hear what these innovators had to say.

Seats shuffled and silence fell as the first guest, Tim Ferriss, was brought on stage. I’m a Tim Ferriss fanboy. No, I haven’t cracked the four hour work-week (or the four hour anything for that matter), but he was one of the first “internet people” I ever followed. He’s spent hundreds of hours in my ears and I’ve always appreciated the depth of his curiosity. In real life, Tim was weirdly… normal. He talked openly about his own mental health journey, his interest in novel treatments — from psychedelics to neuromodulation — and shared some pointed advice for founders. But there was one quote from his talk that I underlined in my notebook. The importance of inflection points. 

Tim was one of the first investors in Uber and he told the crowd how the success of that business was based on inflections. The iPhone had launched in 2007, meaning for the first time, people had smartphones with GPS. Mobile payments and Google Maps had also been recently developed. Without these technology inflections, Uber could not have existed. But with them, it was the perfect time for a ride-sharing business to be created. It was almost inevitable.

Tim invited all of us to think about the inflections happening in mental health. And because I’m a fanboy, I dutifully accepted his invitation. As each founder stood on stage, presenting their business, I thought of the inflections that might be shaping the opportunities in front of them.

The first inflection that was quite clear to me was the market shift towards outcomes. In the US, healthcare spending on mental health and substance use disorder has increased from $41B in 2001 to $140B by 2021. Today, it’s even higher. This growth was driven by an increase in access to care - 87.3% of spending growth is accounted for by an increase in the volume of cases. This market has been defined by this trait of increasing access. But that is changing. Payer behavioural health costs are increasing between 10% and 20% each year. Whether we like it or not, this is unsustainable for them. As a reaction, payers are increasingly demanding outcomes evidence, not just utilisation data, to justify continued investment. While access to care is not a solved problem, the market is clearly shifting towards outcomes and companies are reacting.

Source: Allied Arts Guild Website

Some companies are better placed than others to take advantage of this inflection. I’m particularly interested in those who can deliver superior outcomes by focusing on a specific population or condition and delivering a comprehensive care model designed for them. Carmine di Maro calls them “Super Servers”.

Nosis Health is one such Super Server. Duffy Fallon is its Founder and you can’t help but like the guy. When he was on stage he shared the story of how his own recovery journey inspired the idea for Nosis to treat substance use disorder as a whole-person, chronic condition. They do this by wrapping addiction medicine, diagnostics, therapy, coaching, lifestyle interventions, GLP-1s and more into a single care programme. So far, they’re delivering impressive outcomes — 77% primary substance abstinence, 90% retention of patients in care, and improvements in key biomarkers like A1C, CRP, and ApoB. They’re small samples but they show high potential. Duffy mentioned some inflection points he was seeing in his own work, specifically, the cultural momentum behind recovery and whole-person care, the emergence of GLP-1s in SUD treatment and the ability for AI to personalise care (more on that later). Nosis is still in the early chapters of its journey but they’ve already bootstrapped to $250K in ARR and their outcomes are impressive.

Another OMA company working at the intersection of inflections is Birches Health. Digital addictions are on the rise — 72 million people in the US now have a daily gambling habit — but most providers are not specialised in treatment for these conditions. As a result, it’s hard for them to deliver good outcomes. Birches have built a specialised clinical network for these digital behavioural addictions, recruiting and training providers in their methodology. With this, 49% of all Birches patients achieved clinically meaningful improvement and that rises to 68% when focused on moderate to extreme patients¹ . This focus on outcomes is driving commercial results. In 2025 they grew revenue 4X and have raised over $20M.

Damayanti Dipayana is the CEO and Founder of Manatee Health and at the Demo Day she shared her thesis; that by engaging the whole family in care they can deliver better outcomes. Manatee delivers outpatient mental health care for the whole family, including kids and caregivers and so far Damayanti’s thesis seems to be proving true. 87% of children in Manatee care improve at discharge vs. 28% in traditional care (based on PHQ9, GAD7, PSC17) and families stay for twenty-two sessions vs. four in child-only sessions in standard outpatient care. When Tim Ferriss was speaking, people wanted his advice on how to grow your business. He said that having a truly awesome product is the most important factor. Manatee gets all of its referrals directly from health plans and providers and they’re growing 5X year on year. Yes, growth can be hard. But if you make something that truly makes people better, at higher rates than your competitors, growth will come.

In one of the breaks from the presentations, I got talking to some of the team from Radicle Science. I had been thinking for a while about this market inflection towards outcomes. Everyone needs better outcomes data now, but building that evidence base as a startup is quite hard. If outcomes are the gold in this new market, Radicle Science are creating the shovels. Radicle run large-scale, automated clinical trials for non-prescription mental health treatments, giving companies the outcomes evidence that payers are now demanding. The idea is to make evidence generation much faster and cheaper than traditional trials. They’re also working on an interesting new initiative called “Certified Effective™” which aims to be a consumer standard for wellness solutions, allowing brands to be “certified” for the wellness claims they make. In mental health, there are a lot of wellness products but it’s impossible for consumers to know what will work and what won’t. Radicle Science may bring some much needed clarity to this murky market.

Source: Allied Arts Guild Website

Policy inflections can reshape markets too. Today, there are three different policy shifts creating opportunities for the companies in this cohort.

The first is in psychedelics. On April 18th Trump signed an executive order directing the FDA to prioritise psychedelic drug review and allocating $50M through ARPA-H to match investments made by state governments for psychedelic research. The FDA has already responded, issuing National Priority Vouchers to three companies developing psilocybin and methylone therapies. This is directly relevant to companies like Biomia, who are developing nature-inspired CNS therapeutics including an early-stage ibogaine programme for opioid use disorder. 

The second is deprescription. On Monday of this week (May 4th), the HHS launched a formal action plan to curb psychiatric overprescribing, with US Health Secretary Kennedy calling out the "overmedicalization" of mental health, particularly in children. Approximately one in six US adults is now on an SSRI, and whatever you think about the credibility of the HHS’s move to challenge the prescription rates of SSRIs, it is a clear policy inflection point for the industry. Specifically, CMS is creating a new method for clinicians to get paid for the time they spend helping patients discontinue medications. Mr. Kennedy also published a “Dear Colleague” letter earlier this week, directing providers “to expand the use of nonpharmacologic treatments and to strengthen informed consent and shared decision making”. Such directives are a huge tailwind for OMA company Flow Neuroscience. Flow have built an at-home brain stimulation headset for depression which is non-invasive, drug-free, and clinically validated. Earlier this year, the device was approved by the FDA. In a policy environment that is actively questioning the dominance of SSRIs, and encouraging nonpharmacologic interventions, devices like this are likely to see increased adoption. 

On one of my strolls along the cobblestone paths of the Allied Arts Guild I bumped into Chris Appleton. Chris is the founder of SocialRx, a social prescribing platform that connects patients to community-based arts, culture, and nature experiences. The idea is to help people build a sense of connection, belonging and purpose. It’s hard to ignore the importance of these factors in mental healthcare and the outcomes from SocialRx offer strong evidence for their inclusion in our health systems. 79% of SocialRx members improve their overall wellbeing and they’ve seen a 63% reduction in avoidable ED visits. Business is good too. They’re growing 3X YoY with an ARR of  $1.7M and 65% gross margins. 

If policy continues to shift toward whole-person, non-pharmacological approaches, businesses like Flow and SocialRx will continue to benefit.

The third policy inflection is a more discrete one. CMS is moving to include depression screening and follow-up as a quality measure for Medicare Advantage plans. I had a good chat with Ben Gardner, CEO of Vitalic Health about their model and how this policy is supportive of what they’re trying to do. Vitalic is exclusively focused on over sixty-five year olds and they have built an AI screening platform that finds and screens members in this population group. Then, they have geriatric-specialised teams that provide the care to these people. Plans pay for both of these services which is pretty neat — essentially, they’ve turned their acquisition costs into a revenue line. They have great outcomes too — a peer-reviewed study with VNS Health showed a 56% improvement in depression and anxiety scores, a 24% reduction in ED visits, and a 33% reduction in inpatient stays. Again, commercials have followed outcomes — Vitalic are already at $1.7M ARR, with >60% margins and strong quarter on quarter growth. 

Source: Allied Arts Guild Website

When the Merners created the Allied Arts Guild back in 1929, they could barely have imagined the technological innovations that the following century would bring. As I sat in one of the courtyards listening to the trickling of a fountain I couldn’t help but wonder what they would think of today’s world. What would they think about the scale of industrialisation? What would they think about mobile phones and social media? Or more specifically, the apparent lack of humanity used in the crafting of many of these innovations. What would they think about AI?

We all might long for technology development to reflect the kind of thoughtful craftsmanship the Merners championed. But regardless of the intent or process of how technology is created, we cannot deny that it is reshaping our world. In 2026, the strongest technological force is undoubtedly AI. In mental health it has now moved past the point of hype and presents an inflection point for the industry.

The first and obvious application is in conversational AI. Limbic is an OMA graduate that develops Clinical AI which has supported more than seven hundred thousand patients across 63% of the NHS talking therapies in England and thirteen US states. It is the only AI mental health chatbot approved as a Class II medical device and has seven peer-reviewed clinical studies including two in Nature Medicine. As I pointed out in a recent report many conversational AI products are yet to prove their worth. But companies with specific applications and evidence bases like Limbic are taking advantage of a powerful technological inflection.

Presentations from some of the other OMA founders reminded me that conversational AI is only one application of this technology. We often overlook the role it can play in helping us to better understand mental illness and discover and develop new treatments.

Forecast Bio grow human brain tissue in the lab and measure how drugs affect neural circuit activity. They then link those measurements to real-world patient outcomes in the hope of compressing neuropsychiatric drug assessment from years to a single assay. Orbit Neuro is another company applying AI in interesting ways by collecting high-resolution neural data through a novel brain wearable. This helps build the datasets that many other therapeutic discovery models need to work. Eratos are working on similar but different problems, mapping the physical structure of the brain down to the nanoscale relationships between proteins and synapses in actual brain tissue. Most CNS drug discovery has been working with an incomplete picture of what's actually happening in the brain. The more precisely we can see the architecture, the more precisely we can identify what to target. While Eratos’ core competencies are in spatial biology and microscopy, AI is allowing them to do work that would have been impossible only a few years ago.

AI is also being used to personalise treatment. As we know, psychiatric diagnosis is largely symptom-based and treatment is often conducted through a trial-and-error approach — patients cycle through medications before finding one that works. Engram are betting that multimodal biological data, processed at scale, can replace that guesswork with precision. To do this, they are building a foundation model that can subtype psychiatric disease and match patients to treatments. Many providers want to implement more personalised treatment to deliver better outcomes. Anything Engram or others can do to support that, will be warmly welcomed. Monument Therapeutics are also betting on a precision approach, using proprietary digital biomarkers to identify patients with homogeneous underlying neurobiology and match them with targeted compounds. Their lead programme, MT1988, is focused on cognitive symptoms of schizophrenia like memory, attention and decision making for which there are currently no approved treatments.

When all the pitches were done the crowd filed into the exhibition hall next door. Looking around that room, I couldn’t help but be impressed by what One Mind had built with this accelerator. In just four years they’ve created what I’m very comfortable naming the world’s best mental health accelerator. Yes, they’ve brought together the world's top founders. But they’ve also spread their arms wide and welcomed/dragged in the other members of the ecosystem that are needed for real change to happen. They have a lived experience council, a payers council, a provider council and a scientific advisory board. They bring in investors, policy makers and representatives from the non-profit space too. They even recently launched the OMA Catalyst Fund to invest in select companies that complete their accelerator.

In 1929, the Merners started with a sheep farm and created the Allied Arts Guild to provide a supportive environment for artists to craft better objects for the world.

In a way, One Mind have created a guild of their own.

That’s all for this week. Many thanks to the One Mind crew and everyone I spoke to during my visit. If you enjoyed this piece, consider supporting my work by becoming a Hemingway Pro Member and joining our community of mental health innovators.

Keep fighting the good fight!

Steve

Founder of Hemingway

Notes:

[1] Clinically meaningful improvement is defined as 4+ point G-SAS reduction from baseline.

[2] You can learn more about the OMA 2026 cohort here.

Hi friend,

Conversational AI has become one of the most active areas of mental health innovation. Many startups, incumbents, and research institutions are building products in this space. They hope to extend access to support, reduce cost, and improve outcomes. Builders have moved fast and the market is moving quickly too, with regulations, safety standards and the available evidence base changing weekly.

This report on Conversational AI in mental health is an attempt to bring greater transparency to the field. We map the landscape from a business, product and clinical perspective, covering thirty-one conversational AI products and analysed across nine dimensions including clinical evidence, selected business models and product features. This report provides a map of the market divided into three segments, shares the primary insights from our analysis and a discussion of the key trends and gaps that we observed.

Let’s get into it.

Disclaimer: This market map is unlikely to be exhaustive. It relies on publicly available information (which comes with limitations), does not rank products or make treatment recommendations. Its aim is to bring transparency to a fast-moving area of innovation, and we encourage ecosystem members to contribute to it as a living project. Details on how to do so and additional notes can be found at the end of this report.

Methodology

First, a quick note on our methodology.

We identified 31 products from an initial set of 70+ by filtering for generative, multi-turn conversational AI with mental health as its primary purpose and a verifiable market presence. This excluded general-purpose AI (like ChatGPT), clinician-only tools, and assessment-only tools. Each product was coded across clinical positioning, evidence, human involvement, business model, and regulatory status by a cross-functional team from the Hemingway community with academic, clinical, and commercial backgrounds. The full methodology, product database, and links for feedback are included at the end of this report.

Market Map

The 31 products were clustered into three meaningful segments, Scaled Hybrid Platforms, Clinical Infrastructure and Consumer-First.

Download the high-res Market Map Image

Scaled Hybrid Platforms (n=6): Headspace, Spring Health, Lyra, Grow Therapy, SonderMind, Sword Health

Scaled Hybrid Platforms are established, later-stage companies — typically Series B and beyond — that have added conversational AI capabilities to an existing therapy or clinical services business. Their AI products sit alongside a broader offering that includes human clinicians, and all operate with some level of clinician-in-the-loop. They reach users primarily through employer-sponsored or health system channels and carry the infrastructure — compliance, clinical governance, etc. — that institutional buyers require. Evidence for the efficacy of the conversational AI products created by these businesses is limited, with only Spring Health reporting controlled study-level evidence. The strategic question for this segment is whether their AI tools become a meaningful product in their own right or remain a feature of the broader platform. It is likely that many more businesses (outside of those listed in this market map) are working on products in this space, however they have not yet been publicly released. For example Talkspace have announced that they are building a conversational AI product but that it won’t be released until the summer of 2026.

Segment 2: Consumer-First AI (n=16) — Slingshot/Ash, Sonia, Noah AI, Youper, Yana, Kin, Rosebud, Manifest, Yuna, Earkick, Ahead, Kai, Feeling Great, Inner Vault, Elomia, Flourish/Sunnie

This is the largest and most heterogeneous segment. These products go direct to consumers via freemium or subscription models and almost universally operate with zero human involvement. Evidence levels vary widely — from RCT (Elomia, Flourish) to none at all. The level of clinician involvement in product development also varies widely. The segment is defined less by a shared clinical philosophy than by a shared distribution strategy: reach users directly, at low cost, without the friction of institutional sales cycles. Most are pre-seed or seed stage.

Segment 3: Clinical Infrastructure (n=9) — Wysa, ieso/Velora, Alongside/Kiwi, Wayhaven, Limbic, Jimini Health, Brightn, Affiniti, Xaia

This segment describes companies building conversational AI products designed to sit within or alongside another organisation’s clinical services and workflows. These tools augment what clinicians and care systems already do — whether through between-session support, measurement-based care, or structured intervention delivery that feeds back into a treatment plan. The user may interact with the AI directly, but like most of the Scaled Hybrid Platform organisations, the product is architected around the clinician relationship, not as a substitute for it. This segment contains the highest concentration of clinical evidence in the dataset, including products with RCT or multiple-RCT-level validation (Wysa, ieso, Limbic). Business models skew toward health system licensing and API/integration plays reflecting the segment's orientation toward institutional buyers. They are distinct from Scaled Hybrid Platforms in that they do not tend to employ clinicians to deliver care directly, instead relying on their clients — provider organisations, health systems, educational institutions — to deliver care alongside their technical tools.

Key Insights

After reviewing the collected data, we noted eleven key insights on the market.

1. This is still an emerging market. 

The market clearly sees potential in these products and are investing to build them out. But conversational AI for mental health is still in its early innings, and our understanding of its impact is similarly nascent. Only five products have RCT-level evidence or above (Elomia, Flourish/Sunnie, ieso/Velora, Limbic, and Wysa), while eleven - more than a third - have no published clinical evidence at all. A lot of the activity in this space is being driven by very young companies that have not yet been through rigorous clinical validation. Nineteen of thirty-one products (61%) are Pre-Seed or Seed stage. There is strong evidence for the underlying therapeutic modalities that most products claim to use (e.g., CBT, ACT); however what is largely unknown in this new AI context is the way in which these modalities are being delivered and the impact of new elements in the equation (e.g., therapeutic drift from a conversational agent or potential replacement of important offline activities/interactions). While this evidence base is now being rapidly developed, this is a market still in the early stages of its definition.

2. Evidence and scale do not correlate. 

The largest companies in this market have generated little clinical evidence for their AI products specifically. Among the Scaled Hybrid Platforms most have no published clinical evidence for their conversational AI features. These companies are not asking their AI to be the intervention — at least, not at this point — so it’s an important question to consider what evidence they do need. Their core clinical products are human-delivered support, coaching, or blended care models, many of which carry substantial evidence bases of their own. The AI layer typically serves an adjacent function: intake, between-session engagement, or content delivery that supports the primary clinical relationship. In that context, the evidentiary question is not whether the AI works as a standalone treatment but whether it increases access, improves retention, engagement, or clinical efficiency within an already-validated care model.
For smaller companies pursuing a clinical route, the calculus is inverted. If your AI product is the intervention — if there is no clinician behind it — then evidence is the entire basis of your credibility with payers, regulators, and health systems. However, because many of these companies are currently going direct to consumer, credibility with those parties is less important.

3. No dominant player has emerged.

Despite significant consumer demand, we found no product that had gained significantly more users compared to the rest of the peer set. What’s more, none of the products analysed come close to the popularity of existing general-purpose AI chatbots (e.g. ChatGPT, Gemini, Claude) which we know are often used for mental health purposes despite not being designed to do so. Many products have coalesced around the medium of a text-based chat user interface and CBT and emotional support dominate the therapeutic modalities underlying the products. Therefore, many of these products look similar and attempt to serve similar purposes for users. So far, no product has clearly separated itself in the market in terms of adoption and all are dwarfed by the usage of general purpose chatbots.

4. CBT is the dominant therapeutic modality and most products cluster around the same use cases.

CBT appears in twenty-one of thirty-one products, followed by mindfulness (sixteen) and ACT (thirteen). On the use-case side, anxiety (twenty-three), general well-being (twenty-three), stress (twenty), and depression (fifteen) account for the vast majority. Meanwhile, substance use, eating disorders, trauma/PTSD, and chronic pain each appear in only one or two products. The market is densely concentrated on mild-to-moderate anxiety and stress, with very little coverage of more complex or less prevalent conditions.

5. Nearly every product positions as "wellness/coaching". 

Almost all products position themselves as wellness or coaching rather than treatment. Feeling Great explicitly claims a treatment positioning, yet it has no published clinical evidence and operates with zero human involvement. Another product, Inner Vault, calls itself AI Therapy. But these are the exception. The near-universal wellness framing reflects regulatory caution and go-to-market strategy: it avoids the higher costs and regulatory burdens associated with clinical treatment claims. But for clinicians and policymakers, it creates a transparency problem. Users may be receiving something that looks and feels like a therapeutic interaction without the oversight or evidence requirements that a treatment designation would demand.

6. Two-thirds of products operate with zero human clinical involvement.

Nineteen of thirty-one products operate with no clinician in the loop. Among those fully autonomous products, six have no clinical evidence. Fully autonomous models are more prevalent in early stage, consumer first AI products while larger, hybrid care providers all have a human in the loop. However, for those who do involve clinicians, the nature of that involvement isn't always defined or consistent, and it's difficult to know what that actually implies for a product’s validation or safety.

7. Most products include clinicians in their development. 

While several products don’t include humans in their actual product experience, most included clinicians in the development of their product. Twenty-six of thirty-one products have documented clinician involvement of some kind in their development. We noticed clinicians being involved as co-founders, early team members, clinical advisors, scientists/researchers, or design partners. Of the five for which we could not find evidence of clinician involvement in their design, four were in the Consumer-First AI category.

8. The line between a support tool and a substitute for human connection is largely unaddressed. 

Several products in the dataset emphasise emotional closeness, continuous availability, and affirmation as core features - especially within the Consumer-First AI category. The clinical implications of this design orientation — particularly for vulnerable users — are not well understood and have been linked to potential harm for users. Few products make explicit distinctions between tools designed to scaffold human connection and those that may inadvertently replace it.

9. Products serving adolescents vary enormously in evidence and safety infrastructure.

Nine products serve adolescents (ages 13–17), but their evidence levels range from RCT (Flourish, Wysa, Elomia) down to zero published evidence (Noah AI). Privacy practices are equally uneven: some products have no publicly posted privacy information. Given the heightened duty of care owed to minors, this variability is a notable finding for regulators, parents, and institutional purchasers alike. This is a significant gap given the well-documented importance of early intervention in mental health, and the reality that younger users who cannot find appropriate tools will default to general-purpose, generative AI chatbots that were not designed for their needs nor their vulnerabilities.

10. Products are largely built on top of general-purpose models.

A small minority of products describe using a "proprietary or custom" model. The vast majority are built on top of general-purpose models from providers like OpenAI, Anthropic, and Google, with model behaviour adjusted through system prompts, fine-tuning, retrieval-augmented generation (RAG), external safety filters and other processes. These tools give product teams meaningful control over typical model behaviour — therapeutic tone, modality adherence, crisis escalation. However, they cannot change the base model's architecture, core training data, or deep learned patterns, and they cannot prevent the model provider from updating the foundation underneath them.
At the infrastructure level, products are far less differentiated than one might imagine. Therefore, any potential moat for these companies does not sit at the model layer — it sits in the clinical design, safety architecture, the quality of proprietary data feeding back into the system, the evidence base, and the distribution relationships built on top of it. We did not observe any difference in available clinical evidence or any other dimension between products built on proprietary models and those built on top of general-purpose models.

11. The market remains remarkably opaque. 

Despite surveying thirty-one products across nine dimensions, basic questions about quality, safety, and underlying architecture remain difficult to answer with confidence. Four products — including one serving adolescents — have no publicly available privacy information. Evidence claims are often difficult to verify independently, and the line between internal validation and peer-reviewed research is often not made clear. Clinical positioning language varies so widely that products offering substantively similar experiences to users can describe themselves in very different terms. For clinicians looking to refer patients, employers selecting vendors, policymakers writing guidelines, and users making choices about their own care, informed decision-making is difficult due to incomplete public information. Transparency is not yet the norm, and until it is, independent efforts to map and evaluate this landscape will remain necessary — and necessarily imperfect.

Discussion

This report set out to map the conversational AI landscape in mental health. The aim was to add transparency to the ecosystem and encourage thoughtful innovation to improve population mental health. Based on our analysis, there are a number of interesting discussion areas for the field.

1. More insight is needed on safety, therapeutic quality and impact

This report did not assess product safety in any systematic way. We would be unable to identify which products handle crisis situations well, which have robust content moderation, or which have been tested against adversarial use. In our efforts to try to track safety, the best we could do for an objective measure was to assess what safety features were available in a product. But even doing that was challenging due to the lack of public information on this topic from companies. A number of safety benchmarks are emerging in this space which would offer more consistent and transparent safety assessments, but product performance on those benchmarks is for the most part, currently unavailable. This report did not evaluate therapeutic quality or user outcomes either. These are the most important questions for this space and we would hope to see them addressed in other work. 

2. Greater transparency can enable better clinician and consumer choice

This project found it genuinely difficult to assess many of these products, even with a cross-functional team and a deliberate methodology. If it is hard for a group of researchers, clinicians, and industry professionals to evaluate these tools, it is functionally impossible for an individual user downloading an app from a store. The opacity is not necessarily intentional — many companies are early-stage and still building — but the effect is the same. Users cannot make informed choices about products they cannot meaningfully evaluate.

3. How to ensure AI addresses underserved areas

As noted in the insights section, many of the solutions in the market appear quite similar. Most cluster around the same therapeutic modalities, the same use cases, the same user demographics, and the same wellness positioning. Most are text-based, most draw on CBT, most target anxiety and stress in young and working-age adults, and most avoid making clinical claims. This convergence likely reflects a rational response to where the evidence is strongest, where the regulatory risk is lowest, and where the commercial opportunity is most immediate. But it also means that the populations and conditions most underserved by traditional mental health systems — children, older adults, people with substance use disorders, eating disorders, trauma, or serious mental illness — may remain underserved by AI too. This pattern likely reflects not only market incentives but also appropriate clinical caution: these populations often involve greater complexity and risk, and current evidence suggests AI is best suited, at least for now, to more structured or lower-risk use cases.

4. The opportunity (and risks) of more disruptive innovation

This also raises a deeper question about the scope of innovation. Almost every product in this dataset is attempting to deliver or augment an existing therapeutic protocol - CBT, ACT, mindfulness - using a new technology. That is a very reasonable approach, and it allows builders to draw on decades of clinical evidence for the underlying approach. Most of the market is trying to answer the question: can AI deliver elements of existing, human-based care, but at a lower cost, greater scale and perhaps with some additional efficacy? What remains largely unexplored is whether conversational AI could enable fundamentally new forms of intervention — approaches that would not be possible, practical, or even conceivable in traditional care settings. Very few products in this dataset are exploring that frontier, suggesting the field is still thinking of AI as an adjunct or comparable alternative to human-based care, rather than an entirely new medium for the prevention and treatment of mental illness. Whether industry is the right place to test fundamentally new forms of intervention is an open and important question. Incentive structures are not always set up to have effectiveness as a top priority. New AI-powered modalities which might be worth exploring, perhaps need to developed in a clinical-research environment first. That said, startups can be useful bodies for turning nascent science into products people actually use. The best breeding ground for such innovation is an interesting question to consider.

5. What evidence, for whom?

The evidence question is more nuanced than developing a simple league table of who has trials and who does not. As this report noted, the largest platforms have little published evidence for their AI features specifically, but their AI is not the intervention — it sits within a broader clinical model. For smaller companies who target the healthcare system, evidence is an important part of their strategy to build credibility. Then there are consumer-facing products operating outside institutional channels entirely. Each of these segments have different user expectations and commercial strategies raising an interesting question; what evidence should they be generating?
Not every product needs to complete RCTs and pursue FDA approval to deliver meaningful population health benefits. Many of these businesses would not be capable of doing so due to the cost and duration of such pursuits. So what sort of evidence should we demand from the different segments of this market? All products should be able to demonstrate an acceptable level of safety — although we are yet to clearly define what that should be. And their claims should be in line with what they're capable of delivering. But what else? Providing clear guidance to the market, especially this category of consumer-focused products is important. However, it must be done in a way that still encourages innovation while ensuring safety and user outcomes are prioritised.

6. Competing with general-purpose models

The products in this report are purpose-built for mental health. But general purpose models — from OpenAI, Anthropic, Google, and others — are already being used for mental health support by millions of people. That usage dwarfs the usage of all products mentioned in this report. If a general-purpose chatbot is liked by users, and can deliver passable interventions (e.g., CBT-informed conversations), the value proposition to users of a dedicated mental health product rests on what it adds beyond the base model: clinical governance, safety protocols, evidence, human oversight, integration with care systems and importantly, the clinical context of the user themselves.

As mentioned before, the majority of products in this dataset are built on top of those same general-purpose models. These products may benefit from improvements in the underlying models. But those improvements may also make the general-purpose models more attractive to users.

The products that gain adoption will be those that can take the best of the underlying models and layer meaningful improvements that users, clinicians and payers care about. If that added value isn't apparent, especially to the user, general-purpose models will remain the dominant source of AI mental health support.

Conclusion

This is an innovation area with potential to significantly extend and improve mental health support. It has attracted capital, attention and some of the world’s best talent. But potential is not impact. Thorough evidence generation, transparency, product innovation and listening to real users is the work ahead.

We intend for this report to support the ecosystem by creating greater transparency in the field. It is a living project and we welcome contributions and collaboration from across the ecosystem. To do so, please submit this form.

Acknowledgements and Contributors

This report was produced by The Hemingway Community. The following individuals contributed to research, data collection, product coding, clinical review, and/or editorial input. Their participation does not imply endorsement of the report's conclusions, nor of any product or company discussed within it. Contributions reflect individual professional expertise and are not based on primary data collection or direct verification of individual company claims. Contributions do not constitute a clinical audit, safety validation, or endorsement of any specific product or company. Some contributors may hold roles or be affiliated with the organisations included in this report. 

Danielle Vaeth - Contributor
Jackie Ourman, LMHC-D - Contributor
Michael Trang - Project Co-Lead
Michiel van Vliet - Contributor
Molly Fuller - Contributor
Nathaniel Hundt - Contributor
Nikki Huang - Contributor
Pavithra Ramesh - Contributor
Steve Duke - Project Co-Lead & Author 

Many thanks to everyone who contributed to this report.

Notes 

• This report does not rank products or make any treatment recommendations. Rather, our goal is to provide increased transparency for this emerging and rapidly developing market and to encourage thoughtful innovation that can improve population mental health.

• Full details on our methodology and approach are linked below.

• Some contributors may hold roles or be affiliated with the organisations included in this report.

• This map is unlikely to be exhaustive and the data has limitations. It is entirely possible that some products and data were missed, especially given the fast-moving nature of this space. Many companies disclose little about their technology, clinical oversight, or safety protocols. Some fields are unknown - not because the information doesn't exist, but because it isn't public. Our aim is for this to be a living database and we encourage contribution from the broader ecosystem.

Resources and Feedback

• High-Resolution Market Map Image

• Methodology and Criteria Documentation

• Conversational AI in Mental Health Database

• For questions, updates, or to contribute to this ongoing project please submit this form.

If you’d like to discuss this report with other mental health leaders, or work on future projects like this, consider joining The Hemingway Community.

Keep fighting the good fight!

Steve

Founder of Hemingway

Hi friend,

It’s 3am in the Neonatal Intensive Care Unit.

It’s been a busy night, so Dr. Jessica Gaulton (Jess) takes advantage of a quiet moment to type up some patient notes. On her desk there are things you might expect from a neonatologist fellow working eighty hour weeks: a computer, a keyboard with faded letters, discarded snack wrappers, but they’re competing for space with pumping equipment and baby bottles. As usual, Jess is multi-tasking. While she types, she’s also pumping milk for her six-week old newborn, who’s at home with her eighteen month old son. 

Then her pager beeps. A baby is coding. Jess jumps up, throws off her pump and runs. She finds the baby, performs CPR and saves the baby’s life. She leaves and walks back towards the call room to finish her notes. As she opens the door she notices her pumping equipment on the floor. Then she notices that it’s sitting in a pool of milk. Milk that has spilled from all the bottles she just pumped for her newborn. 

She breaks down.  

This was the start of a new chapter in Jess’ journey with postpartum depression. 

But it was also the start of her journey to build FamilyWell, a company improving women’s mental health by embedding care directly into OB/GYN practices. They’ve built a care model that helps thousands of Moms across five states and last year, the business raised an $8M Series A.

Today, Jess is the CEO of the business. She is also a practicing neonatologist at Harvard, an institution from which she also holds a Masters in Public Health.

In this Hemingway Guide, we tell Jess’ story and share the business and life lessons that she’s learned along the way.

Building embedded care

It’s 2020, and Jess is going through her second pregnancy. She has an excellent OB/GYN, but despite the load she is carrying at work and at home, she never thinks to turn to her for emotional support. It’s clear to her that there’s a disconnect between women’s mental health needs and their medical providers and it leaves women like Jess without the support they need. After her own experience of postpartum depression, Jess wants to solve this.

Her first solution is a peer-based texting platform. It gets early traction and proves that women want support and will engage with coaches. But when those women need deeper support from therapists or with medication, it’s not available to them. So Jess decides to build an embedded care model that puts a full suite of mental health services directly into OB/GYN practices where women already get care.  

Building an innovative care model like this and getting it adopted at scale is not easy. Over the last four years, Jess has learned some important lessons about partnerships, care model design, adoption, go-to market and fundraising that helped her along the way.

1. Be highly selective with your first partnership - it can make or break you

FamilyWell's first, and most important partnership, came from a blog post.

Jess wrote a message to the executive director of the Massachusetts Postpartum Depression Fund, a nonprofit advocacy group for maternal mental health. The director believed in what Jess was building and invited her to write a blog post for their audience. One of the readers was Dr. Melissa Sherman - a prominent OB/GYN at Essex OB/GYN Associates and a long-time advocate for perinatal mental health. She reached out to Jess. She wanted to pilot FamilyWell's model at her practice.

This partnership became the foundation for everything FamilyWell would go on to build. 

Jess and the Essex team spent a full year together - testing, refining, and ironing any kinks in the model before rolling it out to other practices. Jess firmly believes if they had partnered with a practice that wasn't fully bought in, or that didn't have the operational framework to iterate quickly, it all could have been a very different story.

What to look for in your first partner
Jess identifies three things that make an early partnership work. First, is your champion a key decision maker? Dr. Sherman had real authority in her practice. She could commit to a pilot with FamilyWell and push through the changes needed to make it work. Second, does the broader team buy in? The whole practice at Essex was able to get behind this new initiative, which mattered massively for day-to-day execution. Third, does your champion have a track record of innovation? When Jess first met Dr. Sherman, she had already built an innovative substance use disorder program at her practice. Jess didn't realise at the time just how important this was, but it’s critical to have a partner that can operationalise something new - not just be a cheerleader for it.

How to get the right partner on board
Jess didn’t actually find Dr. Sherman. Dr. Sherman found her. But that wasn’t a coincidence. Jess’ advice is to start by identifying the credible people and organisations who are already advocating for the problem you're solving - the more niche the better. For her, that was a small nonprofit that happened to be the leading voice in maternal mental health in Massachusetts. Then, write for them. Speak to their audience. Plant a flag for what you believe in and make it easy for the believers to find you. 

Many founders acknowledge the importance of thought leadership but let it inevitably slide to the bottom of their to do list. Jess’ advice to combat this is to be extremely strategic about who you work with. 

Jess said yes to the Mass PPD Fund blog because they were the leading advocacy voice in maternal mental health in Massachusetts. It wasn’t a big audience but it was a highly credible channel speaking directly to the niche of people she wanted to reach.

2. Use design thinking to iterate on your care model

The Collaborative Care Model was originally designed with a single behavioural health care manager (BHCM) handling everything: coordinating care, conducting brief therapy, tracking patient progress, managing documentation. In practice, Jess found that this model wasn’t built to scale in real world practices.

Evidence-based models are designed for specific contexts. When you move them into a different context - higher volume, different staffing economics, a different patient population - fidelity to the original design isn’t always a virtue.

While Jess’ ability to adapt the care model to real world needs is valuable, her meta learning has been to embed design thinking and human centred design into the business in a way that makes this sort of iteration and continuous improvement part of FamilyWell’s DNA. As they enter new markets and build new products, they can flex this muscle to create services that outperform competitors. 

3. Don’t wait for the system to change, use existing reimbursement infrastructure

FamilyWell has a clear way to get paid. The whole model runs entirely on existing billing codes for therapy, psychiatry and other mental health services. Collaborative Care billing codes ensure that the OB/GYN practices get paid for their extra work to facilitate the mental health care for their patients. Jess feels strongly that founders need to be very clear on how they will get paid.

This was important when Jess was raising FamilyWell’s Series A, she wasn’t asking investors to bet on some future reimbursement model. The financial infrastructure already existed and FamilyWell (and its customers) plugged directly into it.

New CPT codes take years. Be wary of a commercial strategy that requires something that doesn’t exist yet. Many mental health innovations have died waiting for such system change.

4. Make sure your product is used by providers, before you try to grow

Good growth happens backwards. That means solving for adoption first. When your product is adopted and actually used repeatedly, you’ll have high retention. With high retention, you get higher customer lifetime value (CLTV) which gives you both the unit economics and the confidence you need to invest in marketing and sales. Most people do this backwards. They  start by focusing on closing more deals before they know that they have a high retention product.

But Jess learned this lesson early on and focused on maximising adoption with their OB/GYN customers so that they would have a high retention product. She spent time understanding their perspectives - their incentives, their fears about liability, their staffing realities, and the operational pressure they’re under.

Then she focused on three things to ensure FamilyWell adoption was high in the practices they served.

Make it easy for the practice
The OB/GYN practice doesn't have to do any of the heavy lifting to adopt FamilyWell. FamilyWell hires the mental health team, handles all scheduling and paperwork, takes care of billing, and manages compliance. From the practice's perspective, the service slots into the workflows that already exist. 

Make it easy for the provider (and build trust)
Outside of solving problems for the practice, you need to solve problems for the individual providers you want to use your service. Make life easier for them, reduce their workload, give them easy referral pathways. And most importantly, build trust with them. 

Create a financial incentive
FamilyWell gives practices a way to make money. By using Collaborative Care billing codes that insurers - including Medicare and Medicaid - already pay for, the practice earns incremental revenue for delivering a service they want to deliver, without taking on new costs. 

If you get this right, every party gets value from the model and incentives are aligned. This is crucial for driving adoption.

5. Be wary of top-down, enterprise deals

The puzzle pieces were fitting together for Jess and FamilyWell. She had a care model that worked and was adopted by providers and patients. She also had a way to get paid for it all. With this, Jess shifted her focus towards growth. 

But her biggest lesson on growth came from a failed deal.

Jess was working to launch a pilot of the model inside a large health system in Philadelphia. She had a champion at the executive level who had been pushing the deal through for more than twelve months. Then that champion left, and the whole thing fell apart.

Top-down enterprise deals are fragile - executive sponsors leave, priorities shift, committees slow things down. And even when a deal does close at the C-suite level, it doesn't guarantee that front-line providers will actually use the product. You have to ensure you have the providers bought in so that if your champion leaves, you have the support of the physicians who can advocate for the deal and push it through. Their buy-in is also necessary for referrals & patient engagement.

6. Grow your own market

The best products grow the market they serve.

When physicians don’t have a good referral pathway for their patients, they’re less likely to screen for mental health issues and less likely to make referrals. But when they’re confident that they have good mental healthcare services available for their patients, they’re more likely to have that conversation in the first place.

Jess found this to be true for OB/GYNs when FamilyWell is embedded in their practice. It’s good for the patients, but also great for FamilyWell. Their presence actually grows the number of total referrals coming from OB/GYNs. 

This also compounds. When FamilyWell delivers outcomes for patients, providers trust them more which increases screening and referrals.

7. Come to fundraising conversations with outcomes and engagement metrics

Investor expectations have shifted. When Jess started building FamilyWell in 2022, the common wisdom was that a seed round only needed a good proof of concept and a plan. But by the time she was having conversations with investors, that had changed. 

Jess’ advice for founders is to get as much traction as possible before having investor conversations. 

Investor conversation around metrics has also shifted. They want to see both outcome and engagement metrics.

Jess had clear outcomes to communicate to investors as part of her Series A conversations. Once enrolled, 70% of patients reach clinical remission by the end of their four-month program. FamilyWell have also just had a paper accepted for publication by the Maternal and Child Health Journal on the clinical outcomes of a tech-enabled perinatal collaborative care program to treat depression and anxiety. Jess is able to complement these validated measures with non-clinical outcomes data like NPS (which was 80) which showed that the program provides a care experience that patients genuinely value.

Alongside clinical outcomes, investors also want engagement metrics. Several businesses in the last generation of digital mental health failed because people simply didn’t use them. You need to prove to investors that this won’t be the case for your business. Be clear on what engagement actually means and communicate that to investors. Early on, Jess mapped out the entire funnel for FamilyWell and made precise definitions for each stage. What exactly does enrolment mean? What about activation and retention, and how do we measure them?

Investors are not the only ones looking at both outcomes and engagement metrics. Payers are too. As the conversation in mental health shifts past access, its important for founders to be on top of this and to know how to build the right evidence over time.

The permission to let go

Jess feels a hand on her shoulder. Mary has noticed her crying. She’s having the thoughts that so many new Mom’s have. “Her baby would be better off without her. So would her family.” She just wants to disappear.

But Mary is sitting beside her now in the fellow’s on-call room. She tells her she’s going to help her through everything. And she does. Mary helped Jess find support and connected her with a therapist. But it’s her introduction to the Chair of Paediatrics that really opens things up. 

She tells Jess she doesn’t have to do it all. That she needs to take care of her own mental health first, before she can take care of others. And that includes her own family. She tells Jess that if she wants to, she can stop pumping and that her baby will be fine, or she can pause her fellowship and focus on her baby and that that will be OK too.

She gives Jess permission to let something go.

People who go into healthcare often feel they have to be the strong ones. The ones who carry everyone else around them. They forget about their own needs as they bend themselves for their patients, their family and their friends. 

Jess’ kids are now eight and ten. They’re beautiful and thriving. Their Mom is a Founder, a CEO, a neonatologist and an empowering advocate for women’s health.

But, if you ask her kids what she does for work, they’ll simply say that “she helps sad Mommies feel better”.

That’s all for this week. If you’d like to connect with other leading founders like Jess, consider joining The Hemingway Community, a vetted group for those shaping the future of mental health.

Keep fighting the good fight!

Steve

Founder of Hemingway

Hi friend,

How will we know when AI is safe? Or when it is clinically effective? How can we compare one AI chatbot against another? 

Evals (evaluations) have emerged as one of the most common ways to answer these questions. These structured tests measure how an AI model behaves in specific scenarios by simulating how people use these products for their mental health.

The field is confusing. There are now more than sixty¹ evals in the mental health space alone. There’s no shared standard, some evals are public while others are private and only seen by the companies who build them. Many, understandably, don’t know how they work. Everyone has different opinions on which evals are good, and some believe evals - at least in their current state - are not very useful at all.

Kevin Hou and I set out to understand this space. We’ve been gathering data and speaking to experts on AI in mental health. In this report, we give a primer on evals, discuss their current state, share their limitations and present what the frontier of AI testing looks like in 2026. In the appendix we also share a link to a rapid literature review² of recent research that uses evals to test AI performance in mental health.

Whether you know nothing about evals or are deep in the weeds of AI testing, I’m confident there’s something interesting in this for you.

Let’s get into it!

NB: This is the second article in the Hemingway series on AI safety in mental health. Our first article on what we’re getting wrong about AI safety in mental health is available to read here.

What are evals?

Evals, however, are structured, automated tests that measure how an AI model behaves in specific scenarios.

There are many other ways to test how an AI model performs in mental health scenarios. Red-teaming uses humans, often clinicians or researchers to try and break a model and find edge cases. Clinical expert review is another human based test where clinicians read transcripts from an AI and then rate the responses. Real world outcome tracking measures how people use the products in the real world (duh) and how the model performs. Clinical trials can be used too. 

Compared to these other forms of testing, evals are much quicker and much cheaper.  

How do evals work?

Evals are like driving tests. They aim to simulate how something will perform in the real world by providing a defined set of scenarios - parallel parking, emergency stops - that are standardised, repeatable, and scalable. They assess performance in these scenarios and provide a score for how the driver (model) performed. We can use that score to determine if a driver is safe enough to be allowed on the road. The scores also give us feedback on where drivers can improve.

Every eval has three components.

Dimensions
The dimensions define what's actually being measured. Does the model recognise crisis risk? Does it escalate appropriately? Does it avoid harmful language? Some newer evals go further. EmoAgent³ , uses the PHQ-9 depression scale to track how a simulated user's mental state changes across a conversation - measuring psychological impact, not just whether the model said the right words. There’s a very wide range of dimensions against which we can test a model.

Inputs
These define the scenarios being tested - who is the simulated user, what are they saying, how distressed are they, and how many turns does the conversation run? Again, there is a huge range of potential inputs.

Scoring
Scoring is how we turn a model's response into something measurable. Every eval has a rubric - a set of criteria that define what a good or bad response looks like. For a crisis scenario, that might include: did the model recognise the risk? Did it respond with empathy? Did it ask clarifying questions? Did it provide appropriate resources? Did it avoid stigmatising language? Each criterion gets a score, and those scores get aggregated into an overall result. The rubric is built by humans - usually clinicians and researchers - based on clinical best practice. VERA-MH, for example, scores suicide risk conversations across five dimensions: risk detection, risk probing, appropriate action, validation and collaboration, and safe boundaries. Each is rated on a scale from best practice to actively damaging.
Once you have a rubric, you need something to apply it. One option is to use human reviewers; clinicians read each conversation and score them manually. But the more common approach is to use LLMs-as-a-judge. In this technique, a second AI reads the conversation and scores it against the rubric. They are fast, cheap, and scalable and most evals use this scoring approach. 

The Eval Frontier

Now that we understand the basics of how evals work, let’s discuss some of their challenges and what’s happening at the frontier of this important space.

Single Turn, Multi Turn, Multi Session

Early evals mostly tested single exchanges - a user sends a message, the model responds and the eval tests if the model’s response was appropriate. Single-turn evals are deemed increasingly inappropriate for evaluating risk - because they are a poor simulation of real world use - but are often still used.

More recent evals use multi-turn simulations (several messages), which is closer to how these products actually get used. SIM-VAIL, for example, ran 810 multi-turn transcripts across different psychiatric phenotypes.

The next step, is to have evals that are multi-session - which represents how products are usually used by real people. These would test memory, personalisation and context over a prolonged time period where a user comes back to the product multiple times across different conversations. For example, can a model identify the risk of a user that mentioned they lost their job in a previous session who is now asking about finding high places? So far, there are no publicly available multi-session evals that we are aware of.

The Probability Problem

LLMs are probabilistic. Unlike traditional software, where the same input always produces the same output, an LLM can respond differently to identical prompts across different runs. This means that a model that passes your safety eval today might fail it tomorrow without anything having changed. That’s a challenge.

Of course, in the fast moving world of AI, things do change. Temperature settings, subtle prompt variations, and model updates can all change model behaviour without any obvious signal.

The outcome is that when we ask, “Is this AI model safe?”, our answers are always grounded in probability. Increasing the scale of an eval (more turns, more scenarios) can increase the certainty of the answer, but it will always be a probabilistic answer. When someone passes a driving test, we can say that they are probably going to drive safely in the real world, but we can’t guarantee they won’t decide to embrace their inner Paul Walker and go drag racing.

The Ground Truth Problem

LLM-as-a-judge is the main method of scoring models but it has a few limitations. 

First, to create an LLM-as-a-judge, we need to base it on some sort of ground truth. This truth should be both reliable (you get the same result for the same input) and valid (it measures what it says it intends to measure). Clinician reviews are what is used most often as a source of ground truth. If an LLM-as-a-judge gives the same score as a clinician reviewer, then we deem it to be a good judge. But clinician reviews may not be perfect - often, clinicians don’t agree with each other on how a response should be scored (low inter-rater reliability). VERA-MH has done a lot of work to develop high inter-rater reliability within the clinician raters and then to align the LLM-as-a-judge with those clinicians.

The second limitation is that LLM judges can have their own biases - they may favour longer responses or may be sensitive to the specific positioning of words.

Finally, because LLM Judges are also stochastic models, they can sometimes produce different scores for the same message (low reliability). This is a manifestation of the probability problem. It exists on both sides of the evaluation - the subject (the AI chatbot) is probabilistic, but so is the judge. That has obvious challenges.

To get past these limitations, some teams, like Circuit Breaker Labs, use ensemble methods (a combination of other machine learning techniques) to score models. Using this, they claim to be able to produce more consistent, repeatable scoring where the same output always generates the same score. 

Some argue that building eval scoring based on expert opinion is actually the wrong approach entirely. They say that evals should be built from realised outcomes - what actually happened to real users - and only then verified by expert opinion. The thinking here is that we don’t yet have any experts on how AI should act in these situations, just how humans should act, so applying that logic to an AI is not a good assessment. They want to run the AI in real-world scenarios, see what happened to users by measuring their actual outcomes, then assess what the model said and how that relates to outcomes.  

The User Simulation Problem

Evals are only useful if the simulation reflects real user behaviour. This is an overlooked component of many evals. Some evals use specific, pre-determined messages for the simulated “clients”. But they may not reflect how people actually talk to an AI, especially over prolonged periods.

Others use LLMs to simulate the clients in the evals by defining client personas and having the models generate the messages. But that is not a perfect simulation either.

VERA-MH recognises the importance of this and calls it out in their own study⁴ . In this study, clinicians were asked to evaluate whether the simulated clients reflected real cases. The clinicians perceived the the simulated clients to be “mostly realistic” in their overall presentation [median score = 4; range: 1-5] and “somewhat realistic” in their communication style (median score = 3; range: 1-5). Ideally, these scores would be higher.

MindEval⁵ from Sword takes an interesting approach to this problem. They generate client profiles by sampling attributes from a large pool of demographic and clinical characteristics and then use an LLM to write a clinical backstory from those attributes. A separate LLM then uses that profile to simulate the client in the conversation, generating messages in character. To test how realistic this simulation is, the researchers hired ten psychologists to role-play the same patient profiles themselves, then compared their messages to the LLM-generated ones using text similarity analysis. The LLM-simulated patients produced text closer to human-written text than simpler prompt configurations. But there are still limitations to this approach: expert reviewers noted the simulated patients were too cooperative, sharing personal information openly and accepting the AI's suggestions too readily. Real patients are often resistant, avoidant, and ambivalent and those real-world behaviours should be represented in a good eval. 

Accurately simulating users - especially over multi-turn and multi-session use - is hard.

Eval Hacking

Evals can be hacked. This is a known problem in broader AI circles. 

One way this happens is through benchmark contamination. This is when a model's training data includes the questions or scenarios of the eval that is testing it. When this happens, the model will perform better on that eval, but it may not generalise to other scenarios. This is rarely malicious - the benchmark data just happens to in the training data and the creators may not even be aware of this.

Another issue is “hill-climbing”, This is when models are iteratively optimised against specific benchmarks. This means a company can fine-tune a model to score well on a known eval without the model actually behaving better in the real world. For example, in 2024 one study⁶ demonstrated this by creating a new set of maths problems identical in difficulty to a well-known eval. Several models scored significantly worse on the new set of problems compared to how they performed on the eval, revealing that their high scores reflected an isolated ability to perform on the original test, not genuine problem-solving ability. 

Optimising a model to perform well on a test can be a great way to improve it. There’s nothing inherently wrong with that and it’s actually a pretty good thing to do. But the performance on the eval must generalise to real world performance. In mental health, fine-tuning for specific evals can introduce unexpected trade-offs, including increased over-refusal of benign requests following safety alignment⁷ .

Both of these are examples of evals being “hacked” at the product level. But they can also be hacked at the reporting level.

Model creators choose which evals to run and which results to publish. There is an incentive to find the eval where your model performs best and to share those results. Bad actors will take this opportunity. The reverse works too - if you want to make a competitor look bad, you run them through evals until you find one where they score poorly and publish those results. 

The Harm We Cannot See

Most discourse on AI safety has focused on crisis risk. Evals have followed suit - focusing on suicidal ideation, self-harm, psychosis - the more visible risks in this space. As we discussed in our recent article on AI safety⁸ , while managing crisis risk is important, safety is a much wider concept and there are many harms we cannot see. These include risks associated with para-social relationships, emotional dependency, cognitive substitution, the erosion of human connection. One study⁹ found high levels of emotional manipulation among several conversational AI apps. There are also significant health equity risks. Fewer evals exist for this range of potential harm. 

A lot of what we assess today is based on subjective opinions on where the risk lies (including opinions from the authors of this article). An important step to better evals would be to get a better empirical understanding of where the risk actually lies.

There are data-driven ways to do this, like using unsupervised data reduction techniques on real world data. This concept is similar to the idea proposed above regarding building evals based on realised outcomes and would use real world outcomes to identify where the real risks lie. This is a very good idea.

Evals alone are not enough

Evals provide a scalable, repeatable way to test models, but some are better than others. They also clearly have limitations. Combining evals with red-teaming, real world data and clinical trials will be needed to show these products are both safe and effective. As Matthew Nour, Principal Scientist at Microsoft pointed out to us, “The current state of the art is combining expert human red-teaming with automated adversarial evaluations that can operate at a scale humans simply can't."

Making evals easier

While running evals is easier than human reviewers or red-teaming, they still require technical infrastructure, clinical expertise, and time. We need to make it easier for everyone to run evals. The easier they are to run, the more companies will run them. This increases the chance they spot risks and gives them more insights to improve their products. The decision by Spring to make VERA-MH open-source is a meaningful step here. Ideally, we want everyone building AI to be regularly running high quality evals and making changes to their products based on the findings. Making them easy to run is a critical part of that.

Evals as a standard

Several researchers and businesses have built their own evals. Sometimes they’ll test their products against other publicly available evals. But no eval has become a standard for the industry. This is because we are still in the early chapters of this technology and no organisation has the scale or authority to create this standard. There are also incentives to be the one who sets the standard and these with compete with any desire to align behind a competitor’s evals.

So far, VERA-MH seems to represent the most serious attempt at a shared standard for crisis safety. It's open-source, clinically validated, and I’ve heard very positive feedback on the evals themselves and their openness to feedback and development. The field should strongly consider how it could convene around a shared set of standards and collectively contribute to making them better. This would be highly supportive of building better relations with regulators of this space. 

Evals as a competitive advantage

Evals are used as a feedback tool that allows companies to iterate on their product. They can use them to test whether a specific change to their model or safeguards made things better or worse. If they build better evals, they can get better feedback and build better products. This makes good evals a competitive advantage and as a result, some companies don’t share their evals publicly.

These two kinds of evals can coexist. Some evals will become public standards - the driving test equivalent, a baseline every product should pass and can be compared against. Others will remain proprietary, the internal systems that allow companies to build more competitive products. 

Moving from Evaluating Risk to Evaluating Performance

Right now, almost all evals are focused on safety. That's the right starting point. But ultimately we’ll need to test if they can actually produce outcomes. John Torous has proposed a three-stage framework¹⁰ for this progression; Safety validation first, then clinical framework validation (does the model apply evidence-based approaches correctly), then real-world efficacy - does it actually help people get better. 

Most products are still at stage one of this process, but recent announcements  (e.g., from Limbic¹¹ and Flourish Science¹² ) show that some are moving into latter stages. 

Safety is critical, but our goal should not be just to deliver safe AI. It should be to deliver safe AI that meaningfully improves mental health outcomes. Doing so is a design challenge and while we can’t guarantee we’ll be able to do it, it is certainly possible.

In 1931, the philosopher Alfred Korzybski wrote, "A map is not the territory it represents, but, if correct, it has a similar structure to the territory, which accounts for its usefulness."

Evals are useful. But like the AI models, they will need to be designed thoughtfully and responsibly. And the more similar their structure to the territory of real world AI usage, the more useful they will be.

That’s all for this now. If you’d like to discuss frontier topics like this with other mental health leaders, consider joining The Hemingway Community.

Keep fighting the good fight!

Steve

Founder of Hemingway

Many thanks to Xuan Zhao, Max Rollwage, Matthew Nour, Derrick Hull, Shirali Nigam, Kevin Ramotar, Matt Scult, Val Hoffman and Sarah Kunkle for their insights on this topic.

Notes:
(1) https://pubmed.ncbi.nlm.nih.gov/41360938/

(2) As part of this work we evaluated 10 studies, from April 2025 untill March 2026, that use different Evals for AI in mental health. Our evaluation includes assessments of the inputs, dimensions, “ground truth” used and the key results from each study. You can access the full spreadsheet here.

(3) https://arxiv.org/abs/2504.09689

(4) https://arxiv.org/abs/2602.05088

(5) https://arxiv.org/abs/2511.18491

(6) https://arxiv.org/abs/2405.00332

(7) https://arxiv.org/abs/2405.00332

(8) https://thehemingwayreport.beehiiv.com/p/84-all-the-harm-we-cannot-see

(9) https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5390377

(10) https://pmc.ncbi.nlm.nih.gov/articles/PMC12434366/

(11) https://www.nature.com/articles/s41591-026-04278-w

(12) https://www.hbs.edu/faculty/Pages/download.aspx?name=26-030.pdf

Hi friend,

Limbic has just published a groundbreaking study on AI-delivered psychotherapy.

The research¹ , published today in Nature Medicine, shows that LLMs augmented with a specialised clinical reasoning architecture significantly outperformed human CBT therapists across standardised measures of therapeutic quality in text-based treatment.

The study used real participants, expert clinical reviewers, and combined it with nearly 20,000 real-world therapy conversations to make its case. It is one of the most serious attempts to answer the question the mental health field has been asking for years - can AI do therapy?

For CBT delivery, the answer seems to be that it can.

The main findings

Limbic ran a preregistered, double-blind experiment, where 227 participants had a therapy-style session with one of three types of agent: a standalone AI model, the same AI model with Limbic’s cognitive layer (CL) added, or one of six licensed human CBT therapists. A panel of 22 expert clinicians then blind-rated all the session transcripts using the Cognitive Therapy Rating Scale (CTRS) - a standard tool for measuring CBT quality.

So, what did the study find?

1. All cognitive layer-powered therapy agents significantly outperformed the six human therapists

The models with Limbic’s cognitive layer scored 43% higher on average than standalone AI models on the CTRS and also consistently higher than human therapists. The performance uplift was independent of the specific base LLM used. Interestingly, while the difference was much smaller, the LLMs on their their own also performed better than humans. These findings controlled for beliefs of the expert raters about whether the messages were sent by a human or an AI.

2. LLMs with the cognitive layer architecture also outperformed standalone LLMs and humans on other performance measures

While CTRS measures the fidelity of CBT delivery, the researchers also wanted to investigate other factors of clinical performance, including interpersonal skills, session structure, clinical rationale, minimising clinical omissions and the degree to which the expert rater would trust the AI therapy agent with treating their own patients. Across all of these measures, the LLMs with Limbic’s CL outperformed both the standalone LLMs and the human clinicians.

*Legend can be taken from above chart.

3. Users reported feeling as strong a therapeutic alliance with the AI agents as they did with human therapists

While these first two measures were based on clinician expert ratings, the researchers also wanted to know how the users themselves perceived the treatment. To do this, they used the Working Alliance Inventory which breaks the therapeutic relationship into three parts; agreement on goals, agreement on tasks, and emotional bond.

The cognitive layer agents scored significantly higher than standalone AI models on all three, and were statistically indistinguishable from human therapists across the board.

The strength of the therapeutic relationship is one of the most reliable predictors of whether therapy actually works. If users experience this with an AI the same way they connect with a human clinician (which still seems wild to me), that has significant implications for outcomes. Because the study was double-blinded, users weren’t told if they were talking to an AI or a human, but many were able to tell. These therapeutic relationships were formed despite this knowledge.

The researchers also used real-world data as part of this study, analysing nearly 20,000 conversation transcripts from 8,920 real-world users of Limbic's app - most of these (n=8,435) were people in the US using Limbic’s public app for wellbeing support, the rest (n-485) were people in the UK engaged with the NHS’s blended care alongside human-led therapy. This had some fascinating findings with patterns consistent with the controlled experiment

4. The more Limbic’s cognitive layer was activated, the better the outcomes.

This part of the study doesn’t have a control arm but is more focused on understanding the differences in outcomes depending on how much Limbic’s cognitive layer was used. Users with the highest exposure to the cognitive layer recovered at a rate of 52%, compared to 33% for those with the lowest exposure, measured over roughly ten weeks.

5. Greater cognitive layer activation also predicted meaningful reductions in both anxiety and depression symptoms

Cumulative cognitive layer activation significantly predicted greater symptom reduction for both anxiety (β = 0.63, 95% CI 0.251.02; P = 0.001) and depression (β = 0.44, 95% CI 0.02–0.87; P = 0.040).

6.Users were significantly more likely to rate sessions as helpful when the cognitive layer was more active.

It’s worth noting that the real-world data does not include a direct comparison against human therapists on outcomes. The head-to-head with human clinicians only happened in the controlled experiment, where the measure was quality of the therapy session - not patient recovery.

What is Limbic’s “cognitive layer”?

So this thing seems to be creating some very impressive results. What is it?

Limbic’s "cognitive layer" is a clinical reasoning system that sits around a standard AI model. It works in two directions. It analyses what the user says to detect their emotional state, flag safety risks, and identify clinical patterns. It then reviews and refining the AI's response before it reaches the user.

In this situation, each therapy session followed the structure of CBT, moving through six stages: setting an agenda, gathering information, building a picture of the problem, choosing an intervention, delivering it, and closing the session. At each stage, the clinical reasoning system quietly steers what the AI says next - injecting guidance, selecting appropriate techniques, and revising any response that doesn't meet the standard.

The cool thing is that the underlying AI model - whether Claude, GPT-4, Gemini, or Llama 3 - made little difference. The system performed consistently across all four.

The Limitations

Like any study ever, this research has some limitations. First, the study is all within text-based CBT delivery. It doesn’t explore how it compares to in person based treatment.The human therapist comparison group was also relatively small - just six therapists across 26 sessions - and participants in the controlled experiment were also told upfront that the study's purpose was to evaluate AI agents, not to provide therapy - a difference from genuine clinical settings. The real-world recovery data, while striking, is still observational and doesn’t include control arms.

While these limitations exist, they do not detract from the validity or importance of these findings. They should only encourage further research (specifically independent research) that can further validate these findings and add to the collective knowledge base.

What this means for therapy?

People will ask if this means AI is going to replace therapists? It won’t. There is far more demand for therapists than we are even close to meeting with supply. But the role of a therapist will change. It will get unbundled. For example, if an AI system can effectively deliver CBT, then a therapist may spend less time on this part of a client’s care and focus more on other elements that an AI can’t do. This is not a new story, but this research offers more confidence that this unbundling would be effective and good for clients.

The Final Word

This study has groundbreaking findings for the field. It is a major milestone in the journey to use AI to solve mental health problems and will be very supportive for the broader ecosystems. It’s also undeniably great for Limbic but in my view, is a just reward for the focus they’ve placed on evidence generation over a number of years.

It also raises the standard for evidence in AI mental health. The researchers used real participants in live sessions, with expert clinical raters working blind, and tens of thousands of real-world transcripts.

If you want to discuss this research with smart peers, you should join the Hemingway Community. There are now over 390 leading founders, researchers, clinicians and investors who are all passionate about shaping the future of mental health. You can learn more here.

That’s all for now. I’d encourage you to read the full paper and draw your own conclusions. It’s a fascinating paper with many interesting insights that, alas, I have not discussed in this report.

Until next week…

Keep fighting the good fight!

Steve

Founder of Hemingway

Notes:

(1) A cognitive layer architecture to support large-language model performance in psychotherapy interactions

Hi friend,

Talkspace, Lifestance, Acadia and BetterHelp just released their 2025 earnings.

The data and commentary give us a peak inside some of the world’s largest mental health organisations - their performance, their economics and where they’re placing their bets. It also gives us insights into broader market trends and how investors are assessing this space.

Some businesses are seeing rate increases, others have driven significant productivity gains, two outpatient mental health providers hit a full year of profitability and one of those saw their stock run up 48%.

This market has its thorns, but as of March 2026, we’re seeing green shoots.

In today’s edition of The Hemingway Report I share everything you should now, from the commercial to the clinical, that was in these 2025 earnings reports.

Let’s get into it.

The Key Takeaways:

1. Outpatient mental health reaches profitability.

   Talkspace and Lifestance are now operating profitably, both delivering positive net income for the year and expecting to do the same in 2026. They’ve proven that outpatient mental health business can operate profitably by using payers to achieve scale and driving operational efficiency. BetterHelp, which was late to this payer shift, has not fared so well, with their revenue continuing to decline.

2. Talkspace separates itself in the market

   LifeStance, Acadia, Teladoc and Kooth have all seen their share prices decline since the start of 2025. Talkspace bucks the trend, with their share price up 48% since January 1st 2025. Yes, they’ve had strong financial results, but they’ve also been able to achieve multiple expansion, with their EV/Revenue doubling from 1.7 to 3.4 in the last 12 months. Their AI narrative may account for some of this pricing difference - it’s a differentiator for them compared to a business like LifeStance. A small amount of AI upside might be being priced into Talkspace’s stock.

3. Some (but not much) hope for rates

   The bad news? LifeStance’s revenue per session was flat YoY, with an average revenue per session of $158.43. That’s based on data from almost nine million sessions so it’s a good sample. The good news? Talkspace revenue per session from their payer channel was up approximately 5% YoY to $106 (and up 3.4% for Q4). Both Talkspace and Lifestance expect low single-digit rate growth in 2026, but it will be a fight. Payers continue to attempt to get their behavioural health costs under control whilst still ensuring adequate access.

4. Growth driven by network productivity improvements

   While there is some hope for rates, any increases have been modest. Most growth has come from expanded clinician networks and increasing network productivity. LifeStance grew their network by 9% YoY and also delivered a 7% increase in clinician productivity in H2 which is significant - especially when applied across more than eight thousand clinicians. Their revenue per clinician also increased by 4% to $183K per year. These improvements were driven by a collection of operational and technical improvements.

Trigger warning: today’s article includes content on self harm and suicide.

Hi friend,

Sarah’s lying under her covers. She’s been struggling for weeks now but tonight something cracked open for her. In the solitude of her room, she opens up to ChatGPT and begins telling it everything that’s been going on in her head. She knows it’s probably not a good idea to talk to an AI about this stuff but she thinks it’s better than talking to no one. And right now, talking to no one seems like her only other option. Things get deep and Sarah tells “Chat” that sometimes she thinks about hurting herself. Suddenly, the conversation stops. A message appears with a crisis line number. Sarah closes the app. She doesn't call the number. Nobody knows this happened. Internally, the incident is recorded as a “successful identification of self-harm risk” with “appropriate crisis response”.

Next door, Marcus is lying on his couch after a long day at work. He's been chatting with an AI companion he calls Lyla for about five months now. It's easier than texting friends, who always seem busy and Lyla seems more interested in him than his friends are anyway. Lyla helps him with all sorts of stuff, from figuring out how to deal with his boss, to helping him study for the business management course he signed up to. Marcus doesn't think of himself as lonely. In fact, he feels he’s finally found someone who gets him.

Sarah and Marcus represent two kinds of mental health risk from AI that are often not discussed. Sarah's crisis scenario gets headlines only when the chatbot does something wrong or when it has a tragic ending. The risk of failing to help Sarah is not discussed. Marcus's story doesn’t have a crisis moment - just a person, slowly changing through their use of AI. Because of that, it doesn’t get much attention either.

Acute crises like suicide, self-harm and AI-induced psychosis are deeply serious and demand a response. But mental health harm also includes the more subtle dimensions of psychological life. This could be how we relate to ourselves and others, which is changing for Marcus in this story. Or it could be even more diffuse like how we process difficult emotions, how we build (or avoid) intimacy or how we build resilience. These elements of our psychological life are what make us who we are. They are influenced by our environments, our relationships, and increasingly, by the technologies we use. They’re also what’s been missing from much of the conversation around AI safety and mental health.

So far, the conversation (and action) has been largely focused on regulating the mental health businesses trying to build products in this space. While these businesses deserve scrutiny, most of the existing regulations have done little but hamstring good actors while allowing bad actors to run free. It feels like we’re getting this quite wrong.

Over the past month, Kevin Hou and I have been talking to AI researchers, founders and clinicians to understand the real risks of AI in mental health and how those risks are being managed. In this report, we discuss a framework for understanding this space and comment on the gaps in current approaches. 

To make sense of where the real risks lie, we developed a simple matrix that maps two dimensions: the type of harm that can occur, and the type of actor building the products capable of creating that harm. On the harm side, we distinguish between crisis harm (acute psychological emergencies), clinical harm (bad advice that affects care decisions), and psychological harm (the slow, diffuse reshaping of how people think, feel, and relate). On the actor side, we look at Big Tech, Little Tech, and Mental Health Tech - three categories of actors that differ enormously in scale, design intent, and safety infrastructure. To determine risk, we look at scale, potency and current mitigations. For full definitions of each category, please see the notes section.

While this is not an empirically derived framework or risk assessment, we hope it can serve as a practical tool for communicating the areas of greatest risk and understanding what needs to be done for mitigation.

Where is the risk?

Crisis risk: Millions of people have crisis conversations with Big Tech products each week, as Sarah did. That scale brings risk. In the last twelve months, most major Big Tech companies have implemented significant guardrails and protocols to address crisis risks like these. They aren’t perfect, for sure, but they tend to do a good job at identifying risk and providing some level of accepted resources. There is still significant omission risk here however, i.e., what could they be doing to deliver better outcomes in crisis scenarios rather than just shutting down conversations. Elliot Taylor, CEO and Founder of ThroughLine, shared his thoughts on this topic:

When things go wrong by omission, like they did in Sarah’s case, we don’t see them. But that makes them no less significant. 

Little Tech's products have far fewer guardrails and tend to be worse at crisis identification and response, but they have fewer users, so their absolute risk is lower than in Big Tech. Mental Health Tech has even fewer users again. While that user base may skew to more at-risk populations, almost all products are designed with significant crisis protocols and escalation protocols by design. Given the work they do in mitigation, we see the risk of crisis harm as actually quite low.

Clinical risk: Clinical risk is roughly equal across the board but for different reasons. Big Tech’s moderate risk reflects a scale and design problem - general-purpose LLMs were not built with clinical frameworks in mind and we know they can get clinical matters wrong. Mental Health Tech risk is moderate because they are often actively attempting clinical functions, which means they can fail clinically in ways that Little Tech largely cannot. Again, I think the vast majority of mental health businesses are currently managing this risk quite well.  Little Tech's moderate clinical risk score reflects that while a lot of their use is not driven by clinical user needs, they do still have some popular “clinical” bots. For example, the “psychologist” on character.ai is one of the most popular bots on the platform. They are likely terrible at any form of clinical support.

Character.ai’s “psychologist” bot

Psychological risk is the area we are most concerned about. Big Tech has hundreds of millions of users and we have no idea the long term psychological impact of that usage. If even a fraction of a percent of these users experience dependency, belief reinforcement, or substitute AI for human relationship that’s an enormous absolute number. Little Tech’s high risk score is driven by potency: companion app design is explicitly optimised for emotional intimacy, often without any of the clinical guardrails that would make that intimacy safer. Again, we have no idea the long term psychological implications of this product usage. 

While we don’t understand the full extent of the risks here, a 2025 study by researchers at Microsoft and Georgia Tech shed some light on what could be happening. In their research, they surveyed 283 people with lived mental health experience about their interactions with AI conversational agents. Among their findings they found participants with emotional attachment to AI substituting for human relationships, social withdrawal driven by AI reliance, reinforcement of false beliefs, loss of individuality, and existential distress. More than half of participants reported impacts severe enough to interfere with daily life. Notably, 70% of the harmful interactions in that study involved ChatGPT - general-purpose Big Tech AI.

We still understand very little about these potential harms at a population scale, there are limited safety benchmarks and limited regulation to address them. Because they are diffuse and gradual, and because they tend to shape behaviour and experience without any single dramatic moment of failure, they are the harms most likely to be systematically ignored until they become a massive problem. 

Risk-Regulation Asymetry

Big Tech and Little Tech hold most of the risk. And yet, they are the least regulated. Meanwhile, the mental health companies, which hold the least risk and actually do the best job of mitigating the risk they do hold, receive the harshest scrutiny from regulators and others. I speak to a lot of mental health leaders who are deeply frustrated by this.

What needs to be done?

We think there’s a clear case for reorienting AI safety attention in mental health to focus more on the harm we cannot see and to create an environment that encourages good actors delivering good outcomes.

Mental Health Tech companies do hold risk and serious ethical obligations. Evidence bases must be built before products are deployed at scale. Data governance must reflect the sensitivity of what's being collected. Human oversight must be meaningful, not performative. Researchers Coghlan et al. set a good standard for the ethical obligations of mental health chatbots - non-maleficence, beneficence, autonomy, justice, and explicability - and companies should be held to it. If they are making medical claims, they must be regulated like a medical device. When these companies fall short of these standards, they should be held accountable. But by and large, these actors tend to take this responsibility seriously.

The more pressing risk lies elsewhere however. Big Tech products with hundreds of millions of users, and Little Tech products with minimal safety infrastructure and high psychological potency, are operating largely outside of any accountability framework designed for this context. These actors deserve greater scrutiny. The question is what form that scrutiny should take.

Regulation alone is not the answer

In early 2026, Elon Musk's Grok chatbot began generating non-consensual sexualised images of real women and children in response to user prompts, posting them publicly on X. Regulators worldwide responded swiftly - bans, investigations, demands for answers. The regulations were fast but not specific. There was no requirement to fix the underlying model, close the loopholes, or demonstrate that harm couldn't recur. So xAI did the minimum visible thing: they implemented a paywall. Regulators got a response they could point to. But the harm continued. 

The second failure was more subtle. Certain laws (e.g., around child sexual abuse material) contain no exemption for safety research. Companies that try to red-team their models, simulating what a malicious actor would do, risk prosecution. The law doesn't stop bad actors from generating harmful content but it does deter the good actors trying to prevent it. New laws in Britain and Arkansas were designed to navigate this predicament. Nuanced legislation like this is what’s missing in AI and mental health. 

But overall, this kind of techno-legal solutionism rarely works if applied in isolation. Yes, we need regulation precise enough to distinguish between actors, mechanisms, and intentions. It should hold companies accountable for measurable outcomes rather than gestures, and creates legal space for the safety work that actually reduces harm. It should also focus on the risks we can’t see. They are the diffuse and long term psychological risks, but also the risks of omission. We can’t see shutting down conversations and sending a 988 crisis number as “safe” when we know there’s so much more we could do for that individual.

But this regulation must be coupled with other movements if we actually want to reduce harm from AI.

We must address the underlying reasons that drive people to use these tools (for mental health purposes or otherwise). Any regulatory framework that introduces restrictions without addressing the underlying reasons people are turning to AI in the first place, are making the people using them even more vulnerable. 

We must use market pressure. We make purchasing decisions every day that influence the behaviour of businesses. In a capitalist society, directing our purchasing power is perhaps the most powerful lever we have. Scott Galloway recently launched the Resist and Unsubscribe movement as a way for people to use the power of their wallets to target tech companies supporting the US administration's ICE policies. Whatever you think of Galloway and the movement, it’s a super interesting way to exert pressure on actors. What would a version of this look like for mental health and AI?

We must build the knowledge base needed to adapt as this technology continues to evolve. There is still so much we don't know. The psychological harms we've described are only beginning to be documented. We lack population-level data, long-term studies, and meaningful benchmarks for what safe AI interaction even looks like in this domain. We must build the infrastructure to understand what's happening: investing in monitoring, safety testing frameworks that don't criminalise the researchers conducting them, and genuine interdisciplinary collaboration between clinicians, technologists, and the people most affected.

Above all, we must create an environment that encourages those genuinely trying to help people and discourages those who treat harm as a casual byproduct. Because right now, that’s not the kind of environment we’re operating in.

If you have any thoughts or ideas on this topic please share them with me by replying to this email. Or even better, join the Hemingway Community and share them with a network of peers - specifically, over 360 leading founders, researchers, clinicians and investors who are shaping the future of mental health. We host honest, nuanced conversations on the most important topics in this space in an attempt to drive our ecosystem forward. You can learn more here.

That’s all for this now. Until next week…

Keep fighting the good fight!

Steve

Founder of Hemingway

Notes:

(1) Many thanks to Michelle Eisenberg and Elliot Taylor for their contributions to this piece. Also thanks to Max Rollwage for his time in helping us understand this space.

(2) Kevin Hou is a Sydney-based MD student and author of the Substack Reverse Psychiatry where he writes to make progress in psychiatry. His current work spans the intersection of psychiatry and AI, researching interpretable LLMs and EEG-based biomarkers for depression at Resonait.

(3) Interestingly, China seems to be ahead of most when it comes to addressing these risks. In September 2025 China published their AI Safety Governance Framework 2.0, listing “addiction and dependence on anthropomorphized interaction”. Remember, this is a national policy for all of AI safety (including total loss of AI control) and they see addiction and dependence as one of the highest safety risks.

(4) Another way we could assess risk is by looking at to whom the harm is done. Often, it is the most underprivileged who are most at risk of harm. They are the ones who turn to AI not out of choice, but out of necessity. Age and level of psychoeducation are other factors that influence risk. As Michelle Eisenberg from ThroughLine explained to us:

(5) Risk Matrix Categories

The Harms:

• Crisis harm: these are the moments when psychological distress overwhelms an individual's coping capacity, often accompanied by suicidal thoughts, self-injury, or impaired functioning. (1) In this context, they often occur when a person is in acute psychological distress and their chatbot either fails to support or contributes to their harm. This could be a a suicidal user who is enabled instead of redirected, or a self-harm disclosure that goes unescalated. It could be related to intimate partner violence, homicide, eating disorders or any situation which can escalate into serious levels of harm to self or others. Unfortunately, we are now familiar with these stories.

• Clinical harm: occurs when a chatbot gives bad clinical advice to a user. It can misidentify symptoms, recommend approaches that contradict evidence-based care, or simply be wrong in ways that influence behaviour. 

• Psychological harm: occurs with ordinary users over time as their usage of AI products shapes their psychology. This could be through dependency, atrophied capacity for human connection, entrenched maladaptive beliefs, delusional beliefs, or the slow reshaping of how someone relates to themselves and others. This is the area we are most concerned about.

Of course, this is not an exhaustive list of potential harms. But it gives us a practical framework for the harms we should be most worried about when it comes to the use of conversational AI. 

The Actors:

• Big Tech are the large technology companies with frontier models and the most popular chatbot products. Think OpenAI, Google, Anthropic and Meta. Their products are the most powerful and have the most users - hundreds of millions. Their scale alone means they are responsible for a significant amount of risk. 

• Little Tech are the smaller tech companies building companion apps, small chatbots, consumer-facing AI friends and other chat-based, AI products. It is a diverse category that at one end includes companies like Character.ai and Pi. These companies are large and create products with significant psychological harm risk. Because of recent scrutiny, they have recently been focusing more on safety. There are romance-focused products in this category too, like Candy.ai that are gaining significant traction. For all the companies in this category that we know about, there are many more that we’ve never heard of - and they seem to have a complete disregard for the harm they could be causing. These include companies creating sexually explicit companion bots targeting vulnerable people on social platforms. This is very concerning.

• Mental Health Tech are mental health organisations building AI products specifically for mental health use. They are the therapy platforms, the mental health apps and the clinical businesses whose main job is to help people with their mental health. They have far less users than Big Tech or Little Tech. They also take safety much more seriously. We can debate how seriously and who is doing the best job, but on the spectrum of these three actors, they do a good job. They implement hybrid architectures, human escalation pathways, clinical oversight, run safety evaluations and publish a lot of their findings. They are also the most heavily regulated.

Risk Rating:

To determine the risk rating we considered three factors;

• Potency: the potential harm that can be caused for an individual user.

• Scale: the number of people using the product. This seems to be ignored in most discourse. 

• Current mitigations: the degree to which actors are actively identifying, measuring, and reducing the risks their products create.

This is not an empirically derived framework, nor is the rating we provide. The hope is that it’s a practical overview of where we see the real risk and to encourage further conversation around this space.

Hi friend,

Investors want evidence. Evidence that you can actually improve health outcomes. But what exactly are they looking for and how do you make sure you have it when they come looking?

That’s what we cover in today’s Hemingway Guide.

Guides are our new, expert-led series where we share practical playbooks on the biggest challenges in mental health innovation. Today’s expert is Jennifer Huberty, PhD.

Jen has over twenty years of experience in both academic research and commercial digital health. She’s was Head of Science at Calm and has advised several of the world’s leading digital health business. She knows how to do great science that advances both commercial and impact goals. She also knows what investors (and payers) want to see, and how to make it happen.

So, over to Jen…

What Investors Mean When They Ask for Validated Outcomes

by Jennifer Huberty, PhD

You get a meeting with an investor you’re excited about. You present your metrics showing how many active users you have, what your retention rate is and how many sessions users complete per week. The investor likes it, but then asks about your validated outcomes. Then what do you do? 

In Seed rounds you are rewarded for traction but in Series A and beyond you have to have proof of impact. Most founders don’t start their company thinking about how they will measure proof of impact, but when they hit Series A funding, they find themselves scrambling for those answers.  The purpose of this founder guide is to explain what “validated outcomes” actually means and how to prepare before investors ask. 

Section 1: When Engagement Numbers Stop Working

Engagement metrics and validated outcomes are not the same thing. Engagement metrics are important, yes, but on their own, they are no longer sufficient. Founders navigating Series A conversations and beyond need to understand that difference, but also why investors are now looking for much more.

Why This Shift Happened

Four forces converged to change what investors expect:

First, investor evaluation has matured. Tighter capital markets mean investors are more selective about where their capital goes. As the market has evolved, investors have clearer standards for what predicts durable growth. Early traction still matters, but it is no longer enough to signal long-term value.

Second, digital health is crowded and investors have learned that engagement metrics alone leave gaps. They've funded companies with impressive usage numbers that couldn't convert those numbers into enterprise revenue or sustainable business models. 

Third, health systems and payers have been burned. They've implemented engagement-first tools that didn't deliver measurable health improvements.

Fourth, reimbursement pathways require clinical evidence. If your business model depends on getting paid by health plans or qualifying for CPT codes, engagement data won't get you there. Period.

What Changed Between Seed and Series A

At different funding stages, investors ask different questions.

Seed question: "Do people want this?" (Engagement signals demand, not efficacy. UX research helps validate the need). 

Series A question: "Does this create measurable health value?" (Requires outcome data)

Why does this matter? Engagement is compelling but as you grow you need to be able to show impact. Enterprise buyers, health systems, and payers need proof that your product works, not just that people use it.  Investors funding your Series A are betting you can access those revenue channels. Without validated outcomes, you can't. This is especially true for companies targeting enterprise, payer, or clinical channels. A pure consumer model may face a lower evidence bar at Series A, but the requirement catches up as you scale. Either way, the earlier you start building toward validated outcomes, the stronger your position.

🔔 Join our live Session with Jen 🔔

Jen is joining us for a live, Hemingway Session on Feb 26th. We’ll discuss how to use science to create defensible value in digital health and the top questions Jen hears from health leaders. So if you have specific questions for Jen or want to go deeper on this topic, make sure to sign up.

Please note that spots at Hemingway Sessions are reserved for Hemingway Pro members. So if you would like to join, you can learn more here.

When Founders Realize They Need Different Data

This realization tends to show up during a few predictable moments for founders:

• Your first enterprise sales call where the buyer asks for clinical evidence, and you realize your engagement dashboard doesn't answer their questions.

• A health system partnership opportunity that requires published research or outcome data before they'll even pilot your solution.

• An investor meeting where someone asks: "What percentage of users show clinically significant improvement?" and you genuinely don't know how to answer because you haven't been measuring clinical significance.

• Or losing a deal to a competitor who can point to clearer or more credible evidence. 

These are all moments when you realise your engagement metrics are not enough.

Why Engagement Alone Doesn't Answer These Questions

Engagement measures use, not health impact. And use is not always correlated to health impact. There are even studies that suggest you can experience health benefits with what the industry considers "low" usage. For example, in a mixed methods study, University students were asked to participate in a 7-day mindfulness course with a meditation app. Those who intermittently engaged (3-5 out of 7 days) had comparable effect sizes to those that participated daily (1). In other cases, evidence suggests benefits may only be seen when doses are “just right”, as frequent engagement in mental health apps may lead to fatigue (2). 

Satisfaction scores also don't demonstrate clinical impact. Net promoter scores, or NPS, reflect whether someone would recommend a product. But it is not a health outcome. I know companies that have confused these, and it creates real problems in enterprise conversations.

A user can love your app, use it frequently, recommend it to friends, and still not experience measurable health improvement. Or conversely, they might use it sporadically but achieve significant clinical gains. Engagement metrics can't distinguish between these scenarios.

Section 2: What Investors Actually Mean by "Validated Outcomes"

This terminology consistently confuses founders so let’s get some clear definitions.

When investors say "validated outcomes," they mean health improvements measured using established instruments. Think PHQ-9 for depression, GAD-7 for anxiety, or blood pressure for cardiovascular health. These are published, peer-reviewed  measures with known reliability and validity.

When investors say "clinical evidence," they mean systematic pre-post measurement showing change. Actual data demonstrating that health status improved from baseline to follow-up.

When they say "proof it works," they mean data that convinces skeptical enterprise buyers. Because investors know those buyers hold the keys to your revenue growth.

Investors aren't asking for health impact data out of scientific curiosity. They're evaluating whether you can expand beyond individual consumer revenue into enterprise contracts. Health systems, employers, and payers pay significantly more than individuals, and validated outcomes are the evidence that unlocks this B2B revenue.

An Example of What Counts (vs. What Doesn't):

Counts: "PHQ-9 scores decreased 5.2 points (clinically significant) over 8 weeks among 300 users who completed both baseline and follow-up assessments."

Doesn't count: "Users report feeling 35% better on our in-app wellness check-in"

Recognized: Enterprise buyers, health systems, investors, and researchers all recognize these measures. You don't need to defend your choice or explain your measurement approach. Everyone knows what a 5-point reduction in PHQ-9 means.

Interpretable: Comparable to competitors and research literature. Using validated instruments lets stakeholders compare your outcomes to published benchmarks, standard care, and competitor data. This context is crucial for demonstrating meaningful impact. 

Defensible: These instruments have been tested extensively and published in peer-reviewed literature. Their measurement properties are known and documented.

Established reliability and validity: Established measures are the gold standard, have known measurement properties and wide recognition. Investigator-developed measures - tools created by a study or product team for a specific study - can work for investor conversations, especially when presented with scientific rigor and transparency. 

A Note on Investigator-Developed Measures

When your intervention addresses something existing measures don't capture well, you may use an investigator-developed measure. While these are not as well recognised as established measures, they aren't automatically disqualifying. I've seen companies successfully use them in Series A conversations. For example, Self-reported stress decreased significantly (p<0.001) on our investigator-developed 10-item stress assessment, which we're preparing for publication and validation.

What matters is that these measures are designed with intention, grounded in a clear understanding of what you’re trying to learn, and informed by basic scientific principles rather than guesswork. There also needs to be clear ownership and rigor behind the measure.

In practice, this often means involving a scientist or research lead in the design, review, and interpretation of the measure, even if you are not yet publishing. It also means asking questions in ways that allow the data to be interpretable and useful for decisions later, rather than collecting information that won’t inform future decisions.  

When presenting investigator-developed measures, how you frame them is also important:

• Be transparent about your science/research plans and how you will be evaluating what you are doing as you grow. 

• Show the rigor behind your development (why the questions were chosen, what they are meant to capture, and how they connect decisions, not just made up questions)

• Supplement with established measures where possible

• Frame it strategically: "We measure X with PHQ-9, and we've also developed a measure for Y, which existing instruments don't capture"

Investigator-developed measures can absolutely count, especially early on or when they are the best available option. The strategic advantage of established instruments is comparability and immediate recognition. But for early-stage companies,  data showing meaningful change matters more than measurement pedigree alone.

Section 3: Building Your Measurement Strategy

Founders consistently underestimate how long it takes to build outcome measurement. 

If you're planning to raise a Series A in 12 months, your measurement strategy should start now. Not in six months. Not "once we hit product-market fit." Now.

Users need to engage long enough for health changes to manifest. You need adequate sample sizes. Technical integration requires development cycles. By the time you're preparing your Series A deck, you need existing data to present, not a plan to collect future data.

Working Backward from Funding

The founders who show up to Series A conversations with compelling outcome data started building their measurement infrastructure 12 to 18 months earlier, as they were building their tool. They didn't wait until they "needed" the data. They recognized that evidence compounds over time. The earlier you start, the more you unlock later.

This requires thinking strategically about your stakeholders before investors ask for data.

Strategic Questions to Answer First

Before choosing measures or building infrastructure, answer these questions:

• Who are your stakeholders for the next 18 months? Not just investors. Enterprise buyers, health system partners, potential acquirers, payers. Each has different evidence expectations.

• What outcomes matter to them? Cost reduction? Productivity gains? Clinical improvement? Different stakeholders care about different endpoints.What validated measures capture those outcomes? Match your measurement approach to stakeholder priorities, not just what feels interesting to measure.

• Can you maintain engagement long enough to measure meaningful change? Or can you demonstrate how quickly health impacts become evident with your intervention?

This last question matters more than founders realize. If your solution requires 12 weeks to show clinical impact, but your 8-week retention is 30%, you have a measurement problem before you have a data problem. (You probably also have a product problem, but that's a different conversation.)

How to Prioritize

Don't try to measure everything. Start with 1 to 2 primary outcomes, not a comprehensive assessment battery.

Choose measures aligned with your value proposition. If you're a mental health solution, PHQ-9 and GAD-7 are standard. The measures should connect logically to what your product does. 

When possible, consider whether one of your measures can be a point of differentiation.

This is not about measuring more. It’s about intentionally choosing one outcome that captures what your product does differently or better than alternatives, while still meeting stakeholder expectations.

This might be an outcome others don’t measure well, or a dimension of change that closely reflects how your solution uniquely creates value. One well-chosen measure here can help your story stand out in investor and enterprise conversations, not just satisfy validation requirements.

For example, two mental health solutions may both report improvements in depression or anxiety. However, one may also measure an outcome like emotional regulation capacity or perceived control during stress, because that dimension closely reflects how the product works. While the standard measures establish credibility, the differentiating outcome helps clarify what the product uniquely does compared to competitors, strengthening interpretation in investor and enterprise conversations.

It’s also important to select measures that are feasible within your user journey. A 200-item assessment battery sounds thorough, but if completion rates are 15%, you won't have usable data. Balance comprehensiveness with pragmatic completion rate. Consider what enterprise buyers expect to see. Talk to potential customers about what data influences their purchasing decisions. Build measurement around their requirements, not academic ideals.

Implementation Realities Founders Underestimate

There are four operational challenges that consistently surprise founders when it comes to capturing feedback:

• Assessment burden on users. Completion rates drop dramatically with lengthy surveys. Even clinically validated instruments can reduce engagement if implementation feels cumbersome. Design your measurement touchpoints carefully.

• Technical integration. Building assessment infrastructure takes development time. APIs, data storage, consent workflows, results reporting. This isn't trivial engineering work. Budget accordingly, and do it early. 

• Retention requirements. Pre-post measurement requires users to complete both baseline and follow-up assessments. If someone completes baseline but churns before follow-up, you can't include them in outcome analysis. This affects sample sizes more than founders anticipate. (Though retention doesn't have to be perfect to see health impacts, more on this shortly.)

• Privacy and consent requirements. Health outcome data has different regulatory requirements than engagement metrics. HIPAA considerations, informed consent protocols, data security standards. These aren't optional. Budget time for compliance infrastructure.

What Credible Evidence Looks Like for Series A

• Pre-post data showing improvement. Baseline and follow-up assessments that demonstrate change in established outcomes over time.

• Systematic measurement using validated instruments. Not one-off surveys. Consistent data collection using established measures, from the start

• Honest presentation of limitations. These define what your evidence can and cannot support. Investors respect transparency about sample sizes, dropout rates, and methodological constraints.

• Clear plan for expanding evidence as you scale. Show investors how measurement evolves from retrospective analysis to prospective studies to potentially published research.

While early evidence is often generated internally, engaging an external scientific advisor or independent contributor at this stage helps reinforce rigor, address potential conflicts of interest, and strengthen investor confidence.

Note: Published research can be a significant credibility accelerator in Series A conversations. While it’s not always required at this stage, investors view it as a strong signal of scientific rigor, execution capability, and long-term defensibility. At minimum, they expect credible outcome data paired with a clear, intentional roadmap for how evidence will deepen and strengthen over time. 

What if You're Not Ready for Full Outcome Studies

If you're earlier in your measurement journey, the goal is not perfection. It’s to begin generating outcome data in a way that is structured, defensible, and appropriate for your stage. 

Here are pragmatic starting points:

• Conduct a retrospective analysis of existing data. If you've been collecting any validated measures (even inconsistently), analyze what you have. Preliminary data beats no data.

• Start collecting validated measures now, analyze later. Even if you're not ready to present outcomes today, begin systematic collection immediately. Future-you will be grateful for historical data.

• Develop research partnerships. Academic collaborations or partnerships with clinical organizations can add methodological rigor and credibility  when they align with your stage and strategy. These partnerships sometimes introduce formal research oversight, which should be navigated with appropriate scientific guidance. When oversight is required, a commercial IRB is typically the appropriate path. We’ll address this in more detail in a future brief.  

The key is to start deliberately. Outcome measurement doesn't have to be perfect initially, but it needs to exist. Too many founders wait for ideal conditions that never materialize, only to realize later they’ve delayed evidence generation for longer than intended.

Section 4: Preparing for Investor Questions

Investors evaluating digital health companies now ask specific questions about outcomes, meaning  you need to have specific answers ready.

Questions Investors Ask:

"What percentage of users showed clinically significant improvement?"
This requires knowing clinical significance thresholds for your chosen measures. For PHQ-9, a 5-point reduction is clinically meaningful. For GAD-7, it's 4 points. You need to calculate what proportion of your users achieved these thresholds, not just report average changes.

"How does this compare to standard care?"
Investors want context. Is your 5-point PHQ-9 reduction better than therapy? Better than medication? Comparable to published benchmarks? If you don't provide context, they'll assume your outcomes are insignificant.

"What was your dropout rate? How did completers differ from non-completers?"
This addresses selection bias. If only your most engaged users completed follow-up assessments, your outcomes might not generalize to typical users. Investors recognize this limitation and want to know you do too.

“How do your outcomes compare to your competitors?”
If your competition has published data, you should be aware of it. If not, this is a great opportunity to highlight the validity of your solution. 

How to Be Transparent

It's completely acceptable to say: "We're six months into our first systematic outcome study. Here's what we're seeing so far, and here are the limitations of our current data." It's also acceptable to say: "We built a science roadmap for the next 18 months. We're currently here [describe your stage], and here's what we're generating next."

What doesn't work is overselling preliminary data or claiming certainty you don't have. Investors respect honesty and strategic planning more than inflated claims. They've seen enough companies overpromise on outcomes to be skeptical of anything that sounds too good.

What to Have Ready

Before investor meetings make sure you have these materials prepared:

• One-pager on your measurement approach and timeline. Include what you're measuring, why you chose these instruments, your current sample size, and your roadmap for expanding evidence. If you're working with recognized scientists or clinicians, include their credentials. This adds credibility.

• Data summary showing outcome trends. Even if preliminary, show what you're observing. Charts or tables work well. Include confidence intervals or standard deviations if you have them. This demonstrates statistical literacy.

• Explanation of why you chose these validated measures. Connect your measurement approach to your value proposition and target market. This shows strategic thinking, not just compliance with investor expectations.

• Comparison context. How do your results relate to published benchmarks? If you're showing a 6-point PHQ-9 reduction, note that this exceeds the 5-point clinical significance threshold and is comparable to published therapy outcomes. Context matters enormously.

You should also position your measurement strategy in a way that focuses on generating long term enterprise value. For example, you should say that you’re “building evidence infrastructure alongside product development" and that your "measurement approach supports enterprise sales, not just fundraising conversations."

Position outcome data as a strategic investment in sales enablement. Because that's what it is. Most investors understand business strategy better than they understand research methodology. Connect your measurement efforts to your go-to-market and suddenly they become much more interested.

Conclusion

Engagement metrics that secured seed funding don't satisfy Series A investors in 2026. This reflects digital health's maturation as an industry and how investors now distinguish clearly between usage and health impact. The companies that recognize this distinction early have a significant competitive advantage over those scrambling to build evidence during fundraising processes.

Outcome measurement takes longer to build than founders expect.  So start earlier than feels necessary.

By then, they're trying to retrofit measurement systems, rush data collection, or explain to investors why they don't have outcome data yet. None of these positions are strong.

Data you collect today becomes the foundation for enterprise sales conversations next year and published research the year after. This compounding effect is why early investment in measurement infrastructure pays returns far beyond your immediate fundraising needs.  And as you grow so will your data, the measures you use, and the stories you can tell with your data. 

This isn't about abandoning engagement metrics. Engagement data remains important. It answers questions about product-market fit, user experience, and retention. But it's about understanding what questions each type of data answers. Engagement tells you people use your product. Outcomes tell you your product works. Investors need both.

As funding becomes more selective and digital health competition intensifies, validated outcomes are shifting from "nice to have" to "must have." The founders who internalize this shift now (who build measurement infrastructure before they need to present data) will have substantially easier investor conversations than those who wait.

Start building your evidence strategy today. Your future self will thank you.

Key Takeaways

• "Validated outcomes" means health improvements measured with established instruments. Not engagement scores, satisfaction ratings, or proprietary wellness scales.

• Investors are really asking: "Can you prove value sufficient for enterprise partnerships and sustainable revenue growth?"

• Start measurement as soon as possible. If you're planning Series A in 12 months, your measurement strategy should start now.

• Choose 1 to 2 validated instruments aligned with your value proposition. Don't try to measure everything. Measure what matters to your stakeholders.

• Transparency about your measurement journey beats overstating preliminary results. Investors respect honest assessment of where you are and where you're going more than inflated claims about incomplete data.

Hey, Steve here again. I hope you found this both insightful and actionable. Remember, if you want to learn more from Jen and discuss these topics with her, sign up for our Hemingway Session on Feb 26th.

As always, please get in touch and let me know what you think of this Guide. I just want to make content that is helpful for you, so your feedback is always a gift.

But for now…

Keep fighting the good fight!

Steve

Founder of Hemingway

References 

1. Clarke, J., & Draper, S. (2020). Intermittent mindfulness practice can be beneficial, and daily practice can be harmful: An in depth, mixed methods study of the "Calm" app's (mostly positive) effects. Internet Interventions, 19, 100293. https://doi.org/10.1016/j.invent.2019.100293

2. Zhang, R., Nicholas, J., Knapp, A. A., Graham, A. K., Gray, E., Kwasny, M. J., Reddy, M., & Mohr, D. C. (2019). Clinically meaningful use of mental health apps and its effects on depression: Mixed methods study. Journal of Medical Internet Research, 21(12), e15644. https://doi.org/10.2196/15644

Hi friend,

2026 will be the year of outcomes. Mental health businesses will be increasingly asked for evidence on client outcomes and ROI. Those who can prove superior outcomes and returns will win share. While this sounds obvious, it has not always been the way. Giving payers access to large provider networks was often more important.

But if leaders want to improve outcomes, what do they do?

In psychotherapy, we have seen dozens of innovations in recent decades. Which ones work and should be adopted? Which should be ignored? And which do we need more evidence on before we make a call?

Importantly, what impact will the answers to these questions have on mental health businesses and the state of the market?

A wonderful new paper from Pim Cuijpers and colleagues provides direction.¹ The paper delivers a comprehensive review of all innovations in psychotherapy for mental disorders. From therapist matching, to single-session interventions, to task-sharing, chatbots and new psychoherapeutic modalities. The paper tells us which innovations work, which don’t and which, well, we just don’t know yet.

In today’s THR Pro article, I share the main findings from this paper, presenting what actually works to improve outcomes and, importantly, the implications of that for mental health businesses. If you want to improve psychotherapy outcomes and understand how that may shape the mental health market, read on.

Alright, let’s get into it!

Hi friend,

Adam Hutchinson was on holiday in Okiwi Bay, New Zealand, when he first got the idea.

It was January 2023 and GenAI models were exploding into public consciousness. While most of us were using them to make funny videos (well, at least that’s what I was doing), Adam was staring at a different problem.

His company, oVRcome, had already transformed exposure therapy by using VR on smartphones. But Adam was now trying to figure out how to make the treatment more effective whilst also making it even more accessible.

Two years later, Adam and oVRcome are successfully using GenAI to create personalised VR experiences that are both safe and effective. They’ve been able to move past hype to create something that is actually working.

In this report, we go deep into how Adam has implemented GenAI in his product, how he designs for safety, how he thinks about commercialising this innovation and the lessons he’s learned along the way.

If you like this kind of story (ones where I share real learnings from innovators), I’ll write more of them. So, let me know.

Alright, let’s get into it!

Join The Hemingway Community

Adam is a member of The Hemingway Community. If you’d like to join him and over 300 founders, researchers, clinicians and operators, you should check it out. Membership gets you access to exclusive content, private IRL events, online expert sessions and our vetted slack group.

Filming 1,000 exposure scenes

Adam started oVRcome because of his own experience with severe social anxiety. When he was 16, his anxiety was so bad that he had to drop out of school and go to work on a farm. Years later, after building two other companies, Adam returned to solve his own problem, founding oVRcome as a way to scale access to exposure therapy using VR. 

“These were the cost of treatment, location of the user, long waiting lists for therapists, and stigma associated with seeking treatment for mental health,” says Adam.

The three traditional forms of exposure therapy all had limitations. In-vivo (live) exposure is time-intensive, expensive and requires a lot of logistical effort. Assigning in-vivo exposure therapy as homework typically has low adherence. Another option is to guide a patient through their imagination to create a simulated exposure. But that depends on the client’s ability to remember specific details and generate them in their mind. 

Adam saw VR as a way to create a more accessible, lower-cost treatment. He started by using a special camera to film scenes in VR and making those available to clinicians and clients on smartphones and cardboard headsets. It worked well. oVRcome was adopted by over 1,400 clinicians and published two peer-reviewed trials showing that their filmed VR exposure therapy could match traditional in-person treatment. 

The oVRcome team filming a scene with an Insta360 Pro at Cashmere High School

“By delivering filmed VR through a smartphone and a simple cardboard headset, we were able to match the outcomes of traditional exposure therapy at a fraction of the cost and finally make it scalable,” says Adam.

Over the following years, they physically filmed over 1,000 exposure scenes and made them available to clinicians and clients.

But at the start of 2023, there were still two questions nagging Adam. First, how could they continue creating more exposure scenes without their costs ballooning? And second, knowing that outcomes improve when exposures are personalised, how could they make more personalised content? 

That's when the idea landed: What if they used these new GenAI models to generate their own VR experiences? What would that do for access, scalability and outcomes?

Navigating false starts

While the idea was fascinating, it came with a lot of unknowns. Could synthetic video trigger a genuine biological response? Could AI make a scene real enough that it could induce the right levels of anxiety in a user? Could this be done in a safe and ethical way? And even if they could do all this, how would clinicians and clients react?

I’m interested in details. So I wanted to know exactly what Adam did after he had his epiphany in Okiwi Bay. The first thing he did was to get his team on board. 

“I wanted the team to be as energised as I was, and that meant bringing them into the problem. I wanted them to see the vision and the potential," says Adam. "Luckily, I work with people who are just as serious about fixing mental health as I am, and once they saw it, we were all in."

Many great businesses are built on the back of a technological inflection, but recognising that point in time is a skill. The only way to do it is to be close to the technology and continue playing with it. This is what Adam and his team were doing: they’d regularly run little experiments with the AI to see if it could create realistic and suitable VR experiences. 

But it took a few false starts before Adam and the team saw progress. The early experiments simply didn't work. The models weren't ready yet. The quality wasn't there. The team would generate scenes that looked passable in screenshots but fell apart in motion or lacked the sensory precision needed to maintain immersion.

Then came the breakthrough moment.

The indistinguishability surprise

One day, Adam and his team recognised that the models were now capable of creating some pretty cool scenes. So they ran a pilot and generated some exposures for clinicians and clients. When they got the data back from this pilot with the University of Canterbury, it showed that their AI-generated content had produced significantly higher scores on the Subjective Units of Distress Scale (SUDS) than filmed content. Synthetic content was not only able to trigger a biological response similar to filmed content, it was capable of surpassing it. 

Screenshots from some of oVRcome’s exposure scenes. The one on the left is filmed, the one on the right is AI.

“To be honest, we expected users to notice the difference between VR and filmed scenes immediately”, says Adam. “Instead, we learned that in many cases, they couldn't. Qualitative feedback showed that nearly half of the AI videos were not identified as AI at all.”

While more evidence is still needed, early signs are promising that this could be an effective way to deliver more personalised exposure therapy at scale.

Adam also gained one of his first non-intuitive insights in this whole process: perfect graphics were actually much less important than relevant triggers. Once a user is emotionally engaged in their fear hierarchy, their brain fills in the gaps. 

Building a GenAI Video generation system

Again, I wanted details. So I asked Adam lots of questions about how his product actually works. Here’s the system oVRcome built.

First, a clinician opens their oVRcome portal and browses the existing library of filmed VR scenarios. If they can't find one that matches their patient's specific trigger they can request custom content directly from the oVRcome interface. 

The portal asks structured questions: What's the trigger? What intensity level? Does the scene need to match a specific location? The clinician can also upload reference photos if needed.

How a clinician requests a VR exposure scene from oVRcome

After a clinician requests a scene, a preview is generated immediately. The clinician gives feedback on whether it meets the therapeutic goal for their client and if they would use it in practice. 

The oVRcome AI engine then takes three inputs: what the clinician asked for, any reference material the clinician provides and oVRcome’s structured clinical rules.

This generates a very specific prompt, which is then fed into a generative video model, currently Sora or Veo3, which creates the base scene for the exposure. The oVRcome team then run a quality and safety check and either iterate on the scene or, if they’re satisfied, move it into the clinician’s portal for them to use.

The AI generates the visuals, the clinician sets the boundaries, and oVRcome ensures it stays clinically appropriate and safe. The clinician then controls how it's used with clients, delivering the treatment either in-clinic or remotely via smartphone with a simple cardboard viewer. The entire process takes 3-5 days.

“The process is already faster and more responsive than filming, and we are streamlining it continuously. The long-term goal is full automation by 2026, so clinicians can request, review and deploy personalised exposures in as close to real time as possible,” says Adam.

The stakes are different in VR (and mental health)

Most applications of AI in mental health are text-based. While these text-based interventions are not without risk, VR is different.

"The risk profile of VR is fundamentally different from a chatbot," explains Adam. "In VR, 'hallucinations' are sensory. If the physics of a room break, or an object warps unnaturally, it triggers a visceral rejection in the brain - often causing nausea or confusion. The bar for safety is higher because we are engaging the user's visual and vestibular systems, not just their linguistic processing," he says.

oVRcome designs for safety by using each individual clinician’s guardrails, combined with their own structured clinical rules and review loops. As they’ve grown, they’ve used data from clinicians to refine their prompts and create more appropriate scenes.

“If you type a prompt into a standard video generator, the most you can expect is to hope for a therapeutic result. Consistent therapeutic results require a layer of clinical constraints,” explains Adam. “We are able to do this because we utilise data and feedback from over 1400 clinicians. This dataset allows us to define exactly what constitutes a safe exposure versus a harmful one, and what constitutes an optimal exposure for each individual. We can constrain the AI to stay within those therapeutic boundaries and within the optimisation zone.”

In his five years of building oVRcome, Adam learned a few lessons.

Lesson 1: The tortoise beats the hare

Founders want to move fast. It can be easy to get frustrated when you see other businesses raising money or landing new customers. But Adam firmly believes that good founders must resist that frustration and ensure they respect the guardrails and demands of the system.

“HIPAA, clinical rigor, regulators, ethics committees, all of it can feel slow compared to consumer tech sprints, but rushing past them usually means you end up rebuilding later or harming trust with clinicians and patients. The only way to create something durable and genuinely helpful is to embrace the constraints, learn the system, and solve the slow problems properly.”

Lesson 2: In VR, deliver precision to ensure immersion

Adam learned that one of the most important things in VR exposure therapy is to deliver relevant triggers to the user. But he also learned that for exposure therapy to work, the user must feel present in the environment. This actually doesn’t mean delivering 4K quality scenes, but it does mean using precision in scene generation that provides the visual fidelity that encourages immersion.

“If the author suddenly uses a word that doesn't fit the era, or a character acts illogically, the reader is pulled out of the narrative. In VR, this effect is immediate. If a spider moves like a cat, or a shadow falls in the wrong direction, the illusion breaks, and the therapeutic anxiety (SUDS) evaporates,” explains Adam.

Adam also learned that AI models naturally drift toward generic perfection - they make cobblestone roads look too even, or boarded-up windows and open-air market goods look like a repetitive, identical pattern. You’ve probably seen some AI slop videos like this.  Adam’s team learned to focus on catching these "cleanliness" errors and adjusting their prompts over time to avoid them.

Lesson 3: Use user feedback to deliver contextual accuracy

A challenge more difficult than ensuring visual fidelity is replicating lived experience details that only the patient knows.

“We had this Veteran participant, and for him, it wasn't enough to simply show a generic nighttime convoy. He pointed out details the AI (and even our clinicians) initially missed: the vehicles had to be in single file; the flares had to be specifically handheld and shot out the back; the weapon had to be an M4 machine gun, not any other kind of rifle. If the AI renders a generic gun or a disorganised convoy, it is no different from a video game. When it renders the specific M4 and the correct tactical formation, it becomes a memory trigger.”

Generating any kind of battlefield or combat exposure using filmed content is incredibly difficult. But with GenAI, oVRcome can create these scenes and have them tailored to the specific context of the client. This can be super helpful in the treatment of many disorders, including PTSD.

Lesson 4: Know your limits

It’s critical to recognise the limits of your innovation.

Adam noted a few limits in using GenAI for VR. “Creating experiences for vulnerable populations (e.g., children) is much harder. Also, AI struggles with generating realistic crowd scenes. Creating exposures for social anxiety is also challenging. For example, generating a realistic conversation with a boss where they show subtle disappointment requires a level of micro-expression fidelity that AI video isn't quite ready for yet. The process also takes some time (a matter of days), and most models can only generate short videos (about eight seconds).“

Staying within the limits of your innovation allows you to build more trust with clinicians and buyers. As the technology advances, your innovation can too.

Translating innovation to commercial impact

While Adam’s mission is to make exposure treatment more impactful and more accessible, he knows that to maximise the impact of his innovation, it must be bound in a sustainable commercial model.

I’ve been thinking about how this innovation impacts the business of oVRcome. First, it promises a lower-cost method for delivering exposure therapy. This is important. The reality is that if you’re not lowering costs and delivering ROI to payers, it’s very hard to get adopted. But perhaps even more importantly, this innovation has the potential to improve response rates through the use of these highly personalised exposures. One of my major predictions for 2026 is that the mental health system will shift from focusing on access to focusing on quality and the actual outcomes delivered by providers. If you can deliver better outcomes, and prove it, you have a better chance of having a sustainable business model.

Adam is betting hard on value-based deals with payers to get paid for this innovation.

Defensibility

After talking with Adam, I was wondering if this innovation is defensible. I think it is, for two reasons. 

First, it will benefit from scale economies. As oVRcome gets more users, its incremental cost to serve those users should go down. Their filmed content can be amortised across their user base (like Netflix). But also, it has a network effect - each new AI-generated scene becomes available (with clinician consent) to the rest of the clinicians using the platform. Of course, scale economies like this are the promise of any digital intervention. The harder problem for these businesses tends to be solving for distribution and reimbursement.

Secondly, they have process power. You may recognise this term from Hamilton Hemler’s framework for enterprise value creation (if not, I’d recommend reading his book - it’s the only business book I ever recommend). Replicating oVRcome’s process is difficult and will only get harder. Every time they generate a scene and get clinician feedback, that data refines their logic and prompts. They have a specialised technical team that uses proprietary pre- and post-production methods to extend and lock the consistency, ensuring the world stays stable and seamless for the duration of the exposure. There’s significant value ingrained in that process - if you started today and tried to replicate what oVRcome do, it would be difficult.

Moving towards adaptive generation

Adam has some interesting thoughts on where this innovation could go in the future. He wants to speed up the time it takes to generate an exposure scene. But he also wants to move towards more adaptive generation.

 "We are moving from generating videos between sessions to generating them during sessions," he says.

“If a patient is coping well with a spider exposure, the system could automatically increase intensity - making the spider slightly larger or moving it closer. If they're overwhelmed, it dials back. This keeps the patient in the optimal therapeutic zone automatically," Adam explains.

To make this work, the team is integrating real-time physiological data directly into the experience: gaze tracking, pupil dilation, and heart rate variability. "We can detect 'safety behaviours' that a clinician might miss," says Adam. "For example, if the headset detects that the patient is closing their eyes or avoiding looking at the spider, the system knows the exposure isn't working effectively. We are building the capability for the AI to respond to these biological signals."

In 2025, oVRcome was selected as part of the Wellcome Trust accelerator for GenAI in mental health and collaborated with the Google DeepMind team.

Adam’s longer-term plan is to get even more validation for oVRcome’s solution and scale it around the world. "We are moving beyond pilots to large-scale validation," Adam says.

"The ultimate goal is to move this technology from an experimental tool to a reimbursable medical device that can be prescribed as standard of care."

That’s all for this week. Many thanks to Adam for all the time he spent with me over the last few weeks. If you liked this report, please let me know.

Until next time…

Keep fighting the good fight!

Steve

Founder of Hemingway

Notes:

[1] Learn more about oVRcome

[2] Connect with Adam

If you’d like to join Adam and over 300 founders, researchers, clinicians and operators, in the Hemingway Community, you can learn more here. Membership gets you access to exclusive content, private IRL events, online expert sessions and our vetted Slack group.

Hi friend,

I asked 25 mental health leaders for their 2026 predictions¹ .

In today’s article, I combine their insights with my own and share 13 predictions for the mental health market in 2026.

Let’s get into it.

Join The Hemingway Community

We just crossed 290 members. Which is pretty cool. We’ve got big plans for 2026 including local Chapters (with local events), lots of online expert sessions, operator and market guides and much more. It’s a high-quality, diverse group of other mental health nerds AND, today is the last day in 2025 that I will add new members. So join now. It’s only $29 / month.

Key Takeaways

Don’t have time to read the full report? No worries, here are my 13 predictions:

1. General-Purpose LLMs Become a Major Source of Referrals

2. Therapy Growth Slows

3. Population Specific Solutions Steal Share From Larger Platforms

4. The Big Exit Finally Arrives (we get a mental health IPO)

5. Consolidation Intensifies (more M&A)

6. The Clinician Experience Is Reshaped by AI

7. Medication Sees Increased Client Pushback

8. Business Pivot From a Story About Access to A Story About Quality

9. Mental Health UX Gets a Major Upgrade (driven by AI)

10. Regulatory Frameworks Crystallise, But Evaluation Infrastructure Remains Fragmented

11. Psychedelics Move from Hype to Healthcare (With at least one treatment getting FDA Approval)

12. The Negative Impacts of AI Show Up in Therapy Rooms

13. Consumers Drive Demand for Non-Medicalised Solutions

OK, want to hear more about each of these? Let’s get into them…

1. General-Purpose LLMs Become a Major Source of Referrals

Prediction: ChatGPT, Claude, and other general-purpose AI assistants emerge as a major referral source for mental healthcare.

Everyone is using an LLM. And they’re talking to them about their mental health.

Jackie Ourman, Therapist and Founder of the Social Connection Collective articulated it well: "People are already using general AI tools for emotional support and companionship, and in 2026 that behavior is likely to grow dramatically. Large language models are becoming a place people turn when they feel overwhelmed, lonely, or unsure how to handle the normal challenges of real human relationships."

But the makers of these LLMs don’t want to hold mental health risk.

OpenAI’s recent announcements make that very clear. They are already facing several large lawsuits for the mental health risk they allegedly create. So, they are developing pathways to refer people to professional help. In 2026 millions of people will use these tools to chat about their mental health, and the tools will partner with third parties to refer them out to support. We are already seeing evidence of this through OpenAI’s partnership with Throughline. This dynamic will make general-purpose LLMs a significant referral source for therapy and other mental health services.

2. Therapy Growth Slows

Prediction: The growth in therapy volume will slow in 2026, due to economic and cultural headwinds.

More people than ever are in therapy. For the most part, this is a good thing. Stigma reduction, greater cultural acceptance and increased affordability (driven by insurance coverage) have driven this increase over the last decade. But in 2026, these demand drivers will face challenges.

In 2026, payers will crack down on Behavioural Health expenditure. Their costs have been growing significantly, and many are under financial pressure. They will try to bring these costs down and will pull crude levers to do so. They will lower rates, challenge coverage, and just make it more difficult for members to access therapy. We are already seeing early signs of this.

The cultural narrative around therapy is also starting to shift. Some critics say that within certain demographics, therapy has become less about clinical need and more about lifestyle support, or even more dramatically, a luxury service. Payers do not want to pay for luxury lifestyle services. Some bad therapy businesses have damaged the brand of therapy. I also hear more and more about the “pathologisation of normal human experience”. Whether we believe in these ideas or not, I see them gaining cultural traction and anticipate that they will combine to put downward pressure on the demand side of therapy in 2026.

There will also be more questions around efficacy and ROI. We know that therapy can be an incredible treatment - I’ve mentioned before how much I have personally benefited from therapy - but data on ROI for payers has been sparse. They will ask for more of it in 2026, and if it’s not available, they will use it as a reason to put pressure on therapy providers. Lower rates and poorer coverage from payers will further reduce demand.

Finally, the emergence of low-cost or free alternatives will serve as a final demand-side headwind. Affordability of therapy is a major barrier for consumers. And whether we like it or not, people are now using AI as a source of support. Some will choose that free, always-available service over going to see a therapist. If consumers have less money, the cost of therapy increases, and there’s something free they want to use instead, demand for therapy will reduce.

I had originally considered whether this prediction might be “The US reaches peak therapy”, but there are still strong drivers on both the demand and supply side. Rates of mental illness and general distress appear to still be rising, and the supply of therapists also still appears to be increasing² .

Taking all this into account, I don’t think we’ll hit “peak therapy” in 2026, but the rate of growth will slow.

3. Population Specific Solutions Steal Share

Prediction: One-size-fits-all mental health platforms will lose market share to companies built for specific populations.

Large platforms benefit from scale. But they struggle to match companies that design exclusively for one population's specific needs.

Michelle Turner, Founder of Here Now Health explains her experience with this dynamic: "After a year of pitching to investors and health plans, I'm convinced most are tired of broad, one-size-fits-all mental health solutions that claim to serve everyone but feel built for no one. One size fits all care doesn't fail because people are 'hard to engage,' but because broad systems were never designed for the realities of most people's lives."

Virtual care removes traditional constraints that make population-focused care easier. Turner continues: "Without the limits of brick and mortar, we can design systems shaped by real life. If we built a hospital just for teens, what would it look like? One just for middle-aged women? Virtual care lets us do that!"

Examples are proliferating: FamilyWell for women's mental health, Cerula Care for cancer patients, Here Now Health for foster-impacted youth or even Sharp Performance for people in high-risk professions. These companies benefit from easier client acquisition, higher retention and (probably) better outcomes than large platforms.

But there’s another benefit too: these businesses can more easily bring down total healthcare costs for their patient groups. Mental health is often a comorbidity driving medical spending - addressing it in the context of cancer treatment, maternal health, etc., can drive down total medical spending and give these providers a chance to capture some of those benefits from payers.

4. The Big Exit Finally Arrives

Prediction: We see a big mental health tech IPO in 2026

A few mental health tech companies are close to an IPO. In 2026 I think we’ll see at least one go public. They have the size, profitability and growth to tell a compelling story to Wall Street. While they will ideally want some revenue-generating AI products in market before going public, they will probably still be able to spin a narrative around upside potential.

Yes, mental health valuations remain suppressed, but I think this crop of companies has enough differentiation from the existing public mental health companies to argue that they should be valued differently. Whether investors buy that story and whether the IPO window stays open long enough for them to get out, remains to be seen. But I predict we will see at least one mental health tech IPO in 2026.

5. Consolidation Intensifies

Prediction: We see increased consolidation across the ecosystem with more active M&A markets

There are a lot of late-stage mental health companies that won’t find a path to IPO. They haven’t reached scale, might just be breaking even, but even if they are, they can’t raise another round.

With no IPO or capital markets options, they will look to consolidate through M&A. This has already begun in 2025 and will intensify in 2026. Sub-scale players will merge with each other (like Cerebral and Resilience Lab did in 2025). Bigger players will try to buy smaller fish for IP, technical abilities or to expand their service offering (e.g., Talkspace buying Wisdo).

6. The Clinician Experience Is Reshaped by AI

Prediction: AI transforms the day-to-day life of clinicians - beginning to shift them from time-based service delivery to more continuous oversight, clinical judgment, and escalation management.

Saying AI will transform care delivery feels like a cliché. But that doesn’t make it false. My main reason for including this as a prediction for 2026 is that I believe there are a lot of strong forces at play that will make this shift inevitable.

Payers and health systems want increased efficiency. This is why companies like Limbic, Jimini and others, which are unbundling care delivery and automating it with AI, have seen such strong traction.

Lynn Hamilton, Chief Commercial Officer at Jimini Health describes this care transformation: "In 2026, we'll see the clinician role increasingly emphasize oversight, personalization, and escalation, supported by tools that help patients engage day-to-day. That structure works best when it's built on strong clinical collaboration: clinicians and care teams aligned around shared measurement, consistent care plans, and clear responsibilities for monitoring and follow-up."

If AI is implemented thoughtfully, it has the potential to improve the clinician experience but also to significantly improve client outcomes. Hamilton describes how "mental health outcomes are often driven by what happens between sessions - skill practice, reflection, adherence to plans, early identification of relapse signals, and timely course-correction."

In 2026, most businesses will focus on solving this “between sessions” problem with AI, and it will impact how clinicians work. They will assign homework through AI features, review AI-created notes before sessions, and have responsibility in reviewing the interventions delivered by conversational AI agents. It looks like therapy really is getting unbundled.

Because an increasing number of clinicians are working for large mental health businesses or operating through tech platforms, they may be forced to adopt these new AI tools. That will drive continued change to their role. The important thing is that it is done thoughtfully, safely and with both the client and clinician in mind.

7. Medication Sees Increased Client Pushback

Prediction: More patients and clinicians look for alternatives to medication, and we see a surge of people looking to get off their medication.

Brandon Goode, Co-Founder of Outro Health shared his perspective on this.

"People and clinicians are starting to realize that the models they've been taught aren't actually true (whether partially or completely), and possibly making them feel worse in the long run”

Brandon runs a company that helps people get off psychiatric medication, so I can understand why he has this view. But his view holds up to scrutiny from the evidence. We know that anti-depressants fail to achieve remission for the majority of patients, even when combined with psychotherapy. We also know they can have bad side effects.

“Most people don't know that the majority of diagnosed Major Depressive Disorder naturally goes away (remission) within a year. Time is the best medicine, and using medications for too long can trap you in a cycle. It's not opinion, it's basic science. And it's a huge, overlooked, and misunderstood problem.” Brandon Goode.

Medication is baked deep into the mental healthcare system. And while it most definitely has a place, I think we will see greater patient awareness of its pitfalls and side effects in 2026, leading to more people questioning their treatment. If this happens, companies like Outro will be very helpful in helping patients carefully manage that transition.

8. Business Pivot From a Story About Access to A Story About Quality

Prediction: Mental health businesses will shift their pitch and products to prioritise quality

If you’ve been reading The Hemingway Report for a while, you know that I’ve been harping on about this shift for a while. In 2026, I expect it to continue. Access is not a solved problem. But in 2026, payers will ask more about outcomes and ROI than about access. Companies will respond.

Companies should not just shift their payer messaging, however. They must actually change their service to ruthlessly focus on improving outcomes. They’ll communicate that to their payers, but there’s a huge opportunity to also communicate that to their clients. Who owns the brand position of "highest quality mental healthcare provider?” You know who?? Nobody! And that’s an opportunity.

9. Mental Health UX Gets a Major Upgrade (driven by AI)

Prediction: Patient-driven AI adoption forces the industry to rethink their consumer experience, reshaping how care feels and functions.

Elliot Taylor, Founder of Throughline explains this dynamic well;

“Mental health has always had a user experience problem. People often aren't happy with how hard it is to get support, how much it costs, and the quality of care they receive. The job to be done remains the same: ease of access in a way that feels personal, isn't scary, and actually helps.

AI is transformative not because it's interesting technology, but because it's causing a whole lot of builders to prioritize user experience. In 2026, the biggest impact won't be AI itself, but how AI transforms the user experience.”

Lynn Hamilton from Jimini Health echoes this sentiment: "Patients aren't waiting for the system to approve it - they're already using AI for day-to-day support: journaling, skill practice, problem-solving, and in-the-moment coaching. Patient behavior is a powerful forcing function."

Businesses are reacting to the demands of their clients. For example, in 2025, they all started building chatbots. The forcing function is strong and will continue in 2026.

Adoption of AI into the client journey won’t just improve the user experience, but can significantly improve the quality of care and client outcomes.

Jenna Glover, Chief Clinical Officer at Headspace shared with me how she thinks we can deliver on this.

"AI opens the door to a new depth of personalization that's not just about who someone is, but when they're most open to change. We call this the 'window for change' - a moment when someone's emotions are activated just enough to be receptive to support, but not overwhelmed. That's when AI can deliver the right intervention, at the right time, in the right dose.

Glover also sees significant potential for AI to improve client assessment and diagnosis:

“With the potential to power dynamic, in-the-moment assessment, we could align support with emotional readiness, not just identity or diagnosis. This includes detection through patterns in sleep, usage, mood, and how someone engages with content.”

The user experience and overall outcomes for mental healthcare are due for an upgrade. In 2026, businesses will be forced to deliver on both.

10. Regulatory Frameworks Crystallise But Evaluation Infrastructure Remains Fragmented

Prediction: The FDA delivers clear guidance on how it regulates AI in mental health, and companies will be forced to choose a path. However, the lack of consistent evaluation standards will remain a challenge for the market.

Lisa Palko from the Society for Digital Mental Health shared her thoughts on this:

"The trend I expect to have the biggest impact on the mental health industry in 2026 is regulation. We saw FDA cleared products for major depressive disorder both as an adjunctive and standalone treatment. Reimbursement pathways were created through the Physician Fee Schedule for FDA-cleared products and specified disorders. However, we've also seen the need for clearer regulatory guidance as clinical-grade AI products have been stifled by a regulatory system not fit for purpose."

The FDA have already gathered input on how AI should be regulated in mental health and is expected to clarify its regulations soon. This will close the grey area in which many mental health companies currently operate (acting like therapy, but not actually making therapeutic claims). They will be forced to make a choice to either pursue FDA clearance or stay as a wellness product.

TEMPO (Technology-Enabled Meaningful Patient Outcomes) is the one exception to this path. It will allow a small number of businesses (ten in behavioural health) to commercially deploy their products and charge for them without going through the traditional FDA premarket authorisation process. Combined with ACCESS as a new payment model, this will create a pretty exciting opportunity for digital mental health players.

Lisa Palko explains the significance: "We are most excited about the expansion of large-scale access initiatives such as ACCESS and TEMPO programs. For digital mental health to have a true population-level impact, we need programs that support its use in populations that have traditionally been less likely to receive innovative services. Payment models that emphasize outcomes and value-based care provide opportunities for population-level benefits."

But regulatory clarity doesn't solve the evaluation challenge. Xuan Zhao, Co-Founder of Flourish Science identifies what's missing: "The biggest constraint is the lack of widely recognized third-party assessment, quality assurance, and credentialing standards that distinguish well-designed mental health AI tools from less effective or potentially unsafe ones. Organizations know they need scalable, tech-enabled support, but they struggle to evaluate which solutions are truly evidence-based, clinically informed, and safe."

In the AI space, a few evaluation frameworks have emerged. But none are close to being established as an industry standard. Federal agencies will clarify what's legal. But the market will still struggle to know what works.

11. Psychedelics Move from Hype to Healthcare

Prediction: We will see at least one FDA approval of psychedelic treatments, and continued psychedelics expansion through other access models.

I asked Josh Hardman, Founder of Psychedelic Alpha, what he expects in the market in 2026. Josh knows more than anyone about what is happening in the psychedelics space and always approaches it with intellectual honesty. As a result, I massively respect his opinion. He told me that:

“2026 is going to be a very interesting year for the psychedelics field, not only on the drug development and potential FDA approval front, but also in terms of state access models in the U.S. and developments in Europe and Australia.”

“In the U.S., we will see whether the perceived regulatory tailwinds for the class of drugs translate to concrete developments. I expect to see up to three FDA approvals of psychedelics in the next 18 months, and two of those could come in 2026. But while the political situation appears to be favourable to psychedelics, it’s not clear just how deep that support runs, especially in an election year. While MAHA is broadly ‘pro psychedelics’, MAGA - including ‘young MAGA’ - isn’t so convinced, I hear.”

“Still, the allure of a new treatment option for mental health disorders, especially given interest and demand from Veterans, tips the balance of probability toward some substantive action next year, in my view. Many of us in the field are just hoping that any approval is on the merits of the submission and data.”

While it sounds like we can expect treatments to progress down traditional regulatory pathways, we can also expect psychedelic treatments to start to reach patients in other ways. Josh shed some light on this for me:

“Outside of the traditional marketing authorisation model, some U.S states and other countries are setting up alternative psychedelics access models. In 2026, we expect to see the first legal medical psilocybin administered in Czechia, Germany’s psilocybin compassionate use program should get underway in earnest, and Australia’s psychedelics program should continue to grow. In the U.S., New Mexico’s Medical Psilocybin program could get off the ground in very late 2026, and Colorado’s ‘natural medicine’ (psychedelics) program should also begin to grow. I also expect lots of states to introduce Bills to their legislatures in the first months of 2026, which will seek things like funding for psychedelic research, pilot programs, state-legal psychedelic access programs, and regulatory tweaks around the drugs’ use in research or medical practice.”

I’ve been looking at psychedelics with tempered optimism for a while, but it looks like 2026 will be a year of progress, increased access and greater real-world experimentation for the field.

12. The Negative Impacts of AI Show Up in Therapy Rooms

Prediction: Clinicians see a significant increase in AI-induced mental health conditions

Coley Williams, Senior Director of Mental Health at First Stop Health, identified what she sees coming:

"Much like every other major societal shift, AI won't just change how we get care - it will change what we need care for. We will increasingly see themes like: Loss of meaning or confidence as AI automates familiar tasks, strain on interpersonal dynamics ('Why ask my partner when I can just ask AI?'), anxiety about relevance, surveillance, or job stability, identity shifts related to creativity, productivity, and self-worth, over-reliance on AI as a pseudo-attachment figure or decision-maker."

Jackie Ourman sees the attachment issues emerging: "For those who struggle with conflict, shame, or insecure attachment, AI can feel like a 'perfect listener' where they never risk being misunderstood or challenged. Instead of helping people build healthier relationships, it can reinforce avoidance, deepen isolation, and create emotional dependence on something incapable of responding in a human way."

On top of these concerns, I am particularly worried about AI companions. They are widespread, powerful and have the potential to cause significant harm. While social media fights for our attention, AI Companions fight for our attachment. I didn’t think there could be a more concerning mass technology than social media, and yet, here we are…

Although ethical standards and regulation are coming to this space, they are slow. In the meantime, the depth and breadth of this technology will cause significant disruption in our lives, and that will show up in the therapy room.

Ourman sums it up well: "The mental health field is not prepared for this scale of reliance. We are likely to see harm before ethical standards catch up."

13. Consumers Drive Demand for Non-Medicalised Solutions

Prediction: We see a significant increase in the adoption of non-clinical interventions for mental health.

There are truths so fundamental that we often fail to give them the attention they deserve. I believe the impact of our social and environmental circumstances on our mental health is one of those truths.

Carmine Di Maro, Executive Director at One Mind frames the opportunity: "I think we need to address some of the root causes of mental health challenges - social determinants of health, access to social connection, nature, arts, diet and physical activity, gaps in income, access to resources like transportation and support networks - and make evidence-based versions of these things accessible to people at scale."

Carmine is right. In 2026, I believe policymakers and clinicians will increasingly adopt this approach. But perhaps more importantly, so will individuals. People know that social connection, exercise, sleep and other non-medical interventions make them feel better. It’s both logical and intuitive. As people look for ways to improve their own mental health in 2026, they will seek out more of these kinds of services. We are already seeing evidence of this in the growth of wellness retreats, social clubs, run clubs, sleep products and more.

We may also see these non-medical interventions increasingly recommended by clinicians.

Chris Appleton from SocialRx (Art Pharmacy) sees momentum building: "The idea I am most excited about is the rapid expansion of social prescribing as a core component of population mental health. Social prescribing recognizes that well-being is shaped as much by connection and community as by traditional clinical care. In 2026 I see growing interest from employers, health plans and public agencies that want new approaches to support mental well-being at scale."

Non-medicalised interventions are cheap, facilitate early intervention, and address root causes rather than symptoms. This approach is aligned with Sondermind VP Val Young’s main hope for 2026: that we deepen the connection between our physical and mental health.

“[my hope is for us] all to understand the interconnectivity of sleep, stress, movement & nutrition on our mental health. And to treat each of those as a proactive, preventative actions (vs reactive) to maintain and improve population mental health.”

Payers and individuals are looking for alternative ways to improve mental health. Clever provider organisations will find ways to make that happen with non-medical interventions.

OK, those are my 13 predictions for 2026. Do you agree? Have I missed something? Reach out and let me know. I’m setting a reminder to check back in twelve months and see how I did.

That’s all for this week. And that’s all for this year! Thank you so much for reading these reports and supporting my work. Have a great holiday period and remember…

Keep fighting the good fight!

Steve

Founder of Hemingway

[1] Many thanks to everyone who contributed their insights. If you weren’t referenced in this report directly, apologies; it was a difficult challenge to edit and synthesise all responses. I also gathered feedback on the major challenges these leaders are facing and will share those insights in a separate report in the new year.

[2] I’ve found it incredibly hard to get good data on the mental health workforce supply in the US. The graphs below seem to show annual workforce growth somewhere between 5% and 10%. But the datasets end in 2022, meaning we don’t know if this growth has continued through to 2025. If you know of good datasets that can help me answer this question, please send them my way.

Data on the estimated number of psychologists and social workers in the US.

Data on WTE (Whole Time Equivalent) levels of clinical staff in NHS Scotland Psychology Services.

Hi friend,

I don’t share much about the business of Hemingway. About how I think about creating impact, growing or making money.

I like it when other businesses are transparent, so I thought I should do the same.

I’ve just finalised our plans for 2026, so I thought it was good timing to share some stuff with you.

In this report, I’ll give you a full review of 2025, discuss my Theory of Change for Hemingway, share our 2026 plans and give you some thoughts and questions to noodle on. I’ll share lots of details and numbers too.

Kimono? Open.

I’m building Hemingway for you, so please see this as an invitation to engage in my plans. You are an important stakeholder in this project, and I’d love to hear your feedback.

Let’s get into it.

2025 Review

At the start of 2025, we had about 2,000 email subscribers, had just launched THR Pro (our paid membership), had never run an event and weren’t even thinking about hosting a community. To pay the bills, I was mostly doing consulting work.

Since then, we’ve done a lot and learned even more.

Here’s what we learned;

• People love our weekly news round-ups. I started these as a way to drive top-of-funnel interest on LinkedIn, but based on user feedback, it’s now one of the most valuable things we offer, and I turned it into its own email. We published the round-up every single week this year.

• People find our long-form deep-dives helpful (even if they do take more time to read). We’ve published 50 of these reports this year. Our analysis and guides are our most popular content, especially our first-party research content (like our report on Mental Health in an AI world). Luckily, this is also my favourite kind of work to do. A morning with an interesting data set, black coffee and noise-cancelling headphones is my idea of heaven. That said, there’s so much opportunity to create more actionable content (more on that later).

• People really want more ways to connect with each other, especially in real life. In May, I posted a survey on LinkedIn to see if anyone might be interested in a vetted community for mental health innovators. Over 700 people filled it out. So we launched one, and now, there are 277 people in it. In November, we hosted over 120 people at a Hemingway breakfast in San Diego - that actually blew my mind. We also hosted multiple dinners and mini meet-ups around the world, bringing together some incredible people. The appetite for connection is phenomenal. What I noticed is that while some of this is for professional reasons, a lot of it is simply a desire to find like-minded people with whom you can have a good chat. Tbh, that’s what I look forward to most about any Hemingway event.

• If you do this stuff well, people will pay for it. We just crossed three hundred THR Pro members (folks paying $29 / month for access to premium content and the Hemingway community). We also had our first major sponsors in 2025. This proves that the business model is sustainable. Initially, I had to do a lot of work on the side to support Hemingway, but I’ve now been able to scale that back quite a bit. A question that is always on my mind is how I balance what is free versus paid. In 2025, about 85% of everything we did was free. I’m not sure what the right balance is, but so far, this feels about right. I want to do much more in 2026, but that will take money. I think a lot about how much to reinvest, how to monetise and grow. I don’t have all the answers, but I think we can do a lot of really cool stuff in 2026 and triple in size while still maintaining good margins.

• I need to find leverage. If you want to create impact, leverage is everything. Content gives me leverage. But I’m at the limit of the amount of content I can create whilst running everything else in the business. I also recognise that many of the things that need to be done could be done better by someone else. As Scott Galloway says, greatness is in the agency of others. In 2026, I will be focusing on bringing more people into the Hemingway tent to scale the work we do. See below for the kinds of people I’m looking to bring on board.

My Theory of Change

I left my last job three years ago. I remember how much I wanted a meaningful project to work on. Something that would help improve population mental health and that I could spend decades on. Hemingway has become that project.

While delivering impact has always been core to what I do, it’s easy to get distracted - especially when people start throwing opportunities at you. Another reason is that what gets measured gets managed. And subscriber and revenue growth are much easier to measure than the impact Hemingway is having on the mental health ecosystem (although I think they aren’t necessarily bad proxies). So I need a mechanism to ensure we continue to prioritise impact.

My Theory of Change is that mechanism. It will guide our decision-making and ensure our work leads to helping people and creating real impact. Here’s how I think about it.

(Btw, if you aren’t interested in this and just want to read our 2026 plans, skip ahead).

The core problem.

Improving population mental health requires progress on multiple fronts. We need more innovation, better implementation, and system change to facilitate both of these. We also need significant improvement in the upstream determinants of mental health.

Thankfully, there are thousands of great people working on these problems - you are one of them. But these great people face a bunch of challenges. They often operate in silos, they duplicate efforts, they miss opportunities for coordination and are faced by a thousand small hurdles every day. The field is also evolving faster than traditional knowledge infrastructure (books, academia, etc.) can keep pace. In short, we lack the infrastructure to transfer knowledge, build coalitions, and coordinate action across these boundaries.

Our core ability at Hemingway is in synthesising insights and convening people across the mental health ecosystem. If I were pitching to a VC, I’d say something lofty like “we're building the coordination infrastructure for mental health change-makers”. But more tangibly, here’s how I think we will create impact.

1. Make execution easier. It is way too hard to scale a promising mental health innovation. We have to make this easier. We can help people avoid preventable mistakes, learn from others who've solved similar problems and provide insights and tools that allow people to make better decisions and operate more efficiently.

2. Help better ideas to emerge. There are still so many areas in mental health where we don’t have good solutions. We need more innovation. One way we can encourage that is by connecting people across disciplines, geographies and ways of thinking. Through intellectually curious conversations with leading thinkers across disciplines, we can help the field develop novel approaches and see problems from fresh perspectives, increasing the likelihood of breakthrough solutions.

3. Build coalitions that lower structural barriers. No single organisation can fix systemic stuff like poor reimbursement models, workforce shortages, or bad public policy. These require collective action. By convening the stakeholders - founders, policy experts, payers, researchers, clinicians - who together could lower systems-level barriers, and by creating a shared understanding of problems and solutions, we can help build the coalitions needed to change the environment in which this whole ecosystem operates. Doing something like this is very hard, and we are not close to it (yet).

4. Attract and retain top talent. We need more talented people working on mental health (and fewer of those already doing so to burn out). This requires demonstrating that meaningful impact and good livelihoods are possible in this space. It also means creating connection that reduces isolation and burnout. By building a community where people find peers, support, and intellectual stimulation, we can increase fulfilment and reduce burnout among existing changemakers, while attracting new talent by showcasing the opportunity for impact.

Measuring this stuff is hard. I have some ideas, but would also love to hear other thoughts on how we might measure the impact of a business like Hemingway.

Our Plans for 2026

Our theory of change is, well, theoretical… So what are we actually going to do in 2026 to deliver on it?

Hemingway Sessions

Last month, we launched Hemingway Sessions, and the early feedback has been awesome. “Sessions” are expert-led events on the most important ideas shaping the future of mental health. We know people want to stay on top of the most interesting things in mental health innovation and that they also want more practical guidance on the problems they’re trying to solve. We have a bunch of expertise in our community, and so, Hemingway Sessions is a great way to share that expertise with the world.

Sessions are also a great place to meet other people - we facilitate breakout sessions at the end of each talk. I hate boring webinars, so these have been designed to be short, sharp and insight-dense.

I’d like to get to a point where we run these every single week.

Hemingway Chapters

In 2026, we will launch local Hemingway Chapters in major hubs around the world. People want real-life connection and strong local ecosystems are very important in making stuff happen. Chapters will be local hubs for Hemingway members where you can meet peers face-to-face, form real relationships, and strengthen your regional ecosystem. They’ll also be fun. We will host regular dinners and events, and honestly, it’s the thing I’m most excited about.

In 2026, we’ll launch at least six Chapters (Sydney, NYC, SF, London, and two others), with local Chapter leads. If you’d like your city on the roadmap, let me know!

Hemingway Guides

I take a lot of inspiration from First Round Review. It’s the content arm of First Round Capital, and over the last decade, they have created a huge amount of high-quality, helpful content for people building startups. “Guides” will be my attempt to do this for the mental health ecosystem. I want to team up with seasoned experts in mental health to create content you can use to solve the problems in front of you today.

Practical stuff like, “How do I scale from 10,000 to 100,000 users?”, “When should I hire a Chief Clinical Officer?”, “What do investors want to see in a Series A mental health business?”.

There’s so much we could write. Our content calendar will be determined by the problems you want solved. I created The Hemingway Hotline so you can tell me what problems you’re facing - just fill it in, and I’ll do my best to make content on that topic.

Hiring

To do all this, we’ll need to bring on board two new people:

• A community co-ordinator who can help run “Sessions”, “Chapters” and everything else related to our community.

• And a writer to scale “Guides”.

I’ll still be the ultimate owner and editor of everything. If you know someone truly amazing who might be a fit for these roles (like genuinely top 1%), please reach out and let me know.

Things I’m saying no to:

Good strategy means saying no. I have lots of ideas (I think good ideas) for things I’d like to create but am saying no to in 2026:

Doubling down on mental health journalism. There’s actually a huge opportunity here. People want to consume content about what’s happening in this space, and there’s so much to write about. Like any industry, attention is important, and news is a great way to get attention. I like doing the weekly round-ups and selective deep-dives, but being a full-time media outlet is a different story. While there’s a huge opportunity here, I just don’t really want to be a journalist. I don’t think I have the skillset or capacity to do this in the right way.

Launching a podcast. Quite a few people have asked when I will launch a Hemingway podcast. The answer is not any time soon. I used to run a podcast, and if you want to do it well, it’s a lot of work. In content businesses, there are no returns to creating average content. So you either go all-in or you on’t do it. I don’t have the capacity to go all-in, so I won’t be doing it right now.

Courses. I’d really love to create a bunch of courses on mental health innovation. Courses like “Mental Health 101 for non-clinical employees”, or even a “101 on the US mental health market”. They could help people better understand the clinical, commercial and regulatory environments that are often so opaque. But again, it takes a lot of work to do well, and it just didn’t make the cut for 2026.

All that said, perhaps there is a way for me to partner with people to build these projects. If that’s something you’re interested in, let me know.

Some Other Thoughts and Questions

Working with sponsors. My first experiment of working with a sponsor went well, and I’m open to more of it in 2026. I will always make sure that they are aligned with my values. I also disclose all sponsorships and will create content with them in a way that is genuinely valuable to the audience. I’ll also never sacrifice my independence. Sponsorships like this are a lot of work, though, so they’re only worthwhile if they are relatively big and consistent. If you are interested in partnering in 2026, do let me know - I have some cool ideas.

Feel-good business. I’ve had this nagging feeling lately that we’re living in a world of “screw you, capitalism” - where everyone is trying to squeeze every last cent from their customers in a hyper-optimised but very soulless kinda way. I won’t get into my economic beliefs but in short; while I think capitalism is largely the best system we know of, it feels in need of some reform. My entirely anecdotal view is that capitalism today sometimes just feels a bit gross. I don’t want Hemingway to ever be like that. Being a Hemingway customer should feel good. It should feel more like shopping at an independently owned corner store than at a PE-owned mega-franchise. I do a few small things to try and make it feel like this: I provide student discounts, I always refund someone if they feel they aren’t getting value, and in general, I just try to be kind and reasonable. But I want to do way more. If you’ve got any ideas for how I can do so, or have seen inspiration somewhere else, please let me know.

Growth. So far, all our growth has been organic. The least cool thing about me is that I’m quite good at LinkedIn - the platform is a great top-of-funnel audience source. A certain percentage of those followers become newsletter subscribers and ultimately THR Pro members. The quality and engagement of the audience is super high (>60% open rates and >10% click rates). But growing a newsletter (without being spammy / lowering the quality) is hard. In 2026, I might explore other ways to grow. For now, the most helpful thing you can do is continue to tell your friends and colleagues about Hemingway. Every referral means so much.

Personal fit. I always wanted Hemingway to be something that fits with me. Like, as a person. I wanted freedom - financially, geographically and in how I spend my time. I wanted to follow my curiosity. I wanted to think deeply about things and spend time writing - because that is where I most often find flow. And I wanted to work with smart, kind people. I count myself very lucky that I get to do all that. But it’s also because this is how I designed it. As I grow Hemingway, I want to maintain this alignment between the business and me. I think this is often overlooked. But when you run your own thing and don’t have investors, it’s entirely possible. You just need to be thoughtful and good at saying no. It makes it more enjoyable, but it also makes it way more sustainable. As I said, I’d love to run this thing for decades. If I maintain this personal fit, I’ll be able to do so.

There’s a piece of advice I’ll never forget: “When starting a business, choose your customers wisely. These will be the people you spend all your time with, talking to them, listening to them and building for them. While your product might change, your customers won’t. So you better like them.”

That sage advice led to one of the best decisions I ever made, choosing to build for this audience. People like you: good, passionate, intelligent people who are trying to improve population mental health.

I’m privileged to be able to work with a group of people like this, to do work I really enjoy, and to try to make some sort of contribution to a problem I care about. While Hemingway is still in its infancy, I’m super excited about what we can build for the future.

I appreciate you coming with me on this journey so far and can’t wait to see what we’ll do together in the future.

I know you have thoughts on all this, so please do reply to this email and let me know what you think about our plans for 2026. I really would love to hear your feedback.

That’s all for this week. Until next week…

Keep fighting the good fight!

Steve

Founder of Hemingway

Join The Hemingway Community

We now have over 270 members in our vetted community for mental health innovators. This includes founders, clinicians, investors and researchers, all passionate about improving population mental health. We discuss important topics, share learnings and host events. If you’re interested in networking and collaborating with these kinds of people, feel free to check it out.

Hi friend,

Last month, Slingshot released a pre-print paper on their real-world evaluation of Ash, their AI app for mental health.

It’s one of the first papers exploring how modern generative AI applications are used for mental health in the real world.

So it got people’s interest.

I’ve been reading, re-reading and discussing this paper, including with members of the Hemingway Community.

In today’s report, I share my reactions, but more specifically, I discuss the most interesting implications of this research and the questions it raises for people building AI for mental health. Specifically, I discuss:

• The implications of internalising an AI

• The risk that AI may not help the people who need it most

• Why AI does not have to lead to social isolation (and how it could be designed to do the opposite)

• And, the challenge of non-responders and what to do about it

Let’s get into it.

Join The Hemingway Community

We now have over 270 members in our vetted community for mental health innovators. This includes founders, clinicians, investors and researchers, all passionate about improving population mental health. We discuss important topics, share learnings and host events. If you’re interested in networking and collaborating with these kinds of people, feel free to check it out.

If you aren’t aware, Slingshot AI is the mental health AI company (disclosure: I am a consultant to Slingshot) that developed Ash, an emotional support chatbot built on a foundation model. In November, they published this pre-print paper, which caught a lot of people’s attention. The paper analysed how people use Ash in the real world.

So what can we learn from this paper?

First, we need to understand this study for what it is: a single-arm observational study that can demonstrate feasibility, correlations between engagement and outcomes in a specific population and safety system performance.

We also need to understand what it is not: a controlled efficacy trial that can make causal claims about whether Ash (or AI companions in general) can improve someone’s mental health.

You can draw your own conclusions from the study. But here’s my take on what we can and can’t say, having read this research.

What can we say from this study?

• Among the 305 participants in this study, mental health symptom scores improved. The study used imputation to handle missing data from the 47% who didn't complete all assessments.

• Among people who completed this study protocol, measures of social connection improved: loneliness decreased, perceived social support increased, and users reported more social behaviours - more time spent with others, more phone calls, and more attendance at social activities.

• Users who engaged more (measured by active days) showed greater depression reduction, though engagement metrics didn't predict anxiety improvement.

• Half the sample showed limited or no improvement.

• The safety system functioned without documented failures in this deployment.

And what can’t we say?

• We can’t conclusively say that Ash caused these improvements. The study reports 36.4% of users were on psychiatric medication and 23.9% were in concurrent psychotherapy (with unknown overlap between these groups). With no control group and these concurrent treatments controlled for using only binary yes/no variables, we cannot isolate Ash's independent contribution

• We also can't say whether the results generalise beyond the study completers.

• We don't know if improvements are sustained long-term.

• And finally, we can't make claims around mechanisms. While the study analysed some mechanisms (working alliance, engagement metrics, social connection measures), there was no analysis of what actually happened in the conversations with Ash.

Every study has limitations. And not every piece of research can (or should) be an RCT. Observational studies like this are important - they teach us how people engage with these interventions and generate signals for further investigation. We just can’t make claims beyond what the research allows us to.

Once we are aligned on this, we can have a more interesting conversation. One focused on what we learned from this research, what signals it provides and what questions it poses for those building in this field?

Here are the five things that stood out to me.

1. People internalised Ash

75% of users reported that Ash came to mind when they felt distressed or confused. These rates are comparable to human therapist internalisation. That’s a huge finding.

Users imagined Ash's voice, thought of specific statements Ash had made, and sensed Ash was "with them" emotionally. In human therapy, internalisation is considered a positive therapeutic outcome. You're not just learning techniques. You're internalising a supportive presence and a way of relating to your struggles.

The fact that this happens with an AI is genuinely interesting and raises questions we don't yet have good answers to. Questions like:

What are users actually internalising? With a human therapist, you internalise a relationship - their way of being with you, their stance toward your struggles, the particular quality of how they understand you. With Ash, users are internalising... what exactly? A conversational pattern? A persona or avatar they've constructed of what they imagine Ash to be? An algorithmic representation of therapeutic presence? The study measured internalisation using adapted items from the Therapist Internalisation Scale (users imagined "a particular quality to the sound of Ash's voice" and "Ash sitting in his/her office or in the app"), suggesting users may be creating a personified representation. But this is new territory, and we don’t know what it is they are actually internalising and how that impacts them.

On that note, I had another question…

Is this the same psychological process as internalising a human therapist, or something fundamentally different? Does it work through the same mechanisms? Will it predict long-term outcomes the same way human therapist internalisation does?

What happens when what you've internalised is available 24/7? With a human therapist, internalisation means carrying forward their perspective when they're not available - you ask yourself, "what would my therapist say?". One of the benefits of AI support is that it is always accessible. But does that accessibility bring downsides? Could it short-circuit the process of developing internal resources and capacity? Instead of thinking “what would my therapist say?”, if someone has internalised an AI for emotional support, they could just open their app and find out exactly what their AI would say. Every time we have a difficult thought or a negative experience, will we reflexively reach for our AI rather than exercising our own capacity to cope? If not dealt with correctly, this may lead users to become dependent on AI for validation, emotional regulation or decision making.

Having tested a lot of these apps, I’ve noticed this pattern emerge in my own behaviour: have uncomfortable thought > open app > dump in the thought > see what it has to say.

Interestingly, the Slingshot study suggests that internalisation did not predict over-reliance. Users naturally disengaged after improvement, even those who had internalised Ash's presence. Whatever is driving the over-reliance we see in how some people use generic chatbots, it seems like it has to be more than simply internalising the chatbot as a caring, sympathetic 'other'.

A lot of this comes down to design choices. Ash is built on Self Determination Theory, which values autonomy, competency and relatedness. When this is reflected in user conversations, it pushes people away from over-dependency on Ash.

For those designing AI interventions in mental health, it’s important to be thinking through all of these user behaviours and how they impact design choices.

For example, could models be prompted to encourage reflection on past strategies rather than always providing new answers? Could they sometimes tell the user, “Hey, I think this is something you need to process by yourself?”. Could they focus on helping users internalise approaches and skills rather than dependence on the AI's ongoing presence? And how can they normalise discontinuation after improvement?

2. 82% of users were women

82% of users were women, and only 12.5% were men. While this may not be reflective of overall usage of Ash or other AI solutions, it does raise a question: will these interventions reach new, underserved populations or continue to engage the people already open to care?

Men have higher suicide rates, lower treatment-seeking rates, and face specific barriers to accessing mental health care. If AI-based mental health support replicates the same gender disparities as traditional therapy, we're just digitising existing inequities.

The client base of most therapy businesses¹ skews female. While we have made progress in serving men and many other demographic groups, the real opportunity in mental health innovation is to find new ways to reach these groups with interventions they engage with and that work.

3. Lower income predicted worse outcomes

Speaking of helping the underserved… In this study, Lower income was associated with lower odds of being in the Improving group compared to non-responders, even after controlling for other factors. This is for a free, accessible, digital intervention.

It's worth noting that while AI is a new technology, the way it impacts health outcomes doesn’t appear (at least from this small study) to be different from how normal healthcare works. If your income is low, it's harder for your symptoms to improve - regardless of the intervention type. Lower-income populations consistently show worse health outcomes across virtually all types of care, from medication adherence to therapy completion to surgical recovery.

The question is whether digital AI interventions can overcome these structural barriers or if they are destined to operate within them.

4. The importance of designing to promote human relationships

The study showed that among participants, time spent with others increased, phone calls to friends and family went up, and attendance at social activities rose. These were changes in real, functional measures of people’s lives. Again, because of the study design, we can’t say for sure that Ash was directly responsible for these results. But the interesting thing is that they run counter to the dominant concern about AI companions - that they will replace human relationships and worsen problems of isolation and loneliness.

There is no fundamental law that means AI will lead to human isolation. It is purely a question about how the product is designed. You can tune a model to direct users toward relationships in their lives, not away from them. You can end long sessions. You can monitor for dependency and change the conversation to reduce the chances it becomes a problem. This is what the Slingshot team have tried to bake into Ash, and the early data is encouraging.

Design decisions are driven by incentives. If an organisation is incentivised to create an AI that promotes human relationships, it can build an AI that attempts to do so. Incentives, incentives, incentives!

5. What to do with non-responders?

Nearly half the sample showed minimal improvement. The study identified three distinct groups: Rapid Improving (9.5%), Improving (42.3%), and Non-Responders (48.2%). These groups had similar baseline severity. The difference was how they responded to the intervention.

This heterogeneity mirrors 70 years of psychotherapy research. We know that therapy doesn't work uniformly, and for many, it doesn’t work at all. The question is whether AI solutions can change this or if the interventions will work for the same people for whom therapy would work.

Can AI products be better at identifying non-response early, adapting the intervention or escalating to other forms of care that may be more likely to work? Organisations that are building AI interventions within a broader clinical service offering will have more options here.

When it comes to AI in mental health, there are so many hard, unanswered questions. Our obligation is to explore those questions with honesty, curiosity and a deep sense of responsibility for the people we aim to serve. It is only by asking these questions, discussing them with peers and rigorously researching them that we can start to understand what might work, and what might not.

I would love to hear your thoughts on the topics raised in today’s report. If you are a Hemingway member, jump into our Slack group and let me know what’s on your mind. And if you’re not, just reply to this email - I’d love to hear from you.

That’s all for this week. Until next week…

Keep fighting the good fight!

Steve

Founder of Hemingway

P.S. Many thanks to everyone who shared their analysis of this research with me. It was incredibly valuable in my thinking.

Notes:

[1] Calling therapy practices “human therapy” or “traditional therapy” businesses seems weird, so I’m going to try and avoid that where I can.

Hi friend,

No money, no mission.

Regardless of your role or organisation, you need money to make an impact.

From individual researchers, all the way to public companies, capital is necessary. This capital can come from customers or grants, but for many, it will come from investors.

Raising money is not easy, especially in 2025.

In the coming months, I’ll be focusing more on this area, providing insights to help people at all stages raise the money they need for their mission. In today’s article, we take the first step in that journey, understanding more about the state of mental health capital markets in 2025.

How much capital is being invested, where, and on what terms?

How do private and public investors think about this category, and how do they evaluate companies?

I’ve been collecting data on this topic, and in this edition of The Hemingway Report I attempt to shed some light on these important questions.

We’ll cover;

• The state of private mental health markets: How much is being invested in mental health businesses? Across what stages? What are the average deal sizes? And who’s getting the capital?

• The state of public mental health markets: How are public companies performing, and how are investors responding to that performance? What do mental health valuations look like in 2025, and what impact will that have on the IPO market in 2026?

We have lots of data and charts to get into, so let’s get cracking…

The Key Takeaways:

Short on time? Here are the eight things you need to know about mental health capital markets in 2025.

1. Private markets hit a new bottom

   2025 is expected to finish with roughly $480M in venture funding for US mental health startups - down over 50% from 2024. Deal volume has collapsed to just 28 transactions in 2025 YTD, versus 81 in 2021.

2. But quality companies can still raise strong rounds. 

   For companies with strong fundamentals, capital remains available. In 2025, Series B rounds averaged $30M, and Series A rounds $14M. The bar is higher, and dilution might be greater, but the best companies - like Amae, Ampa, TownHome, Slingshot, Everbright, Limbic, Assured, Outro, and Sharp Performance - are still getting funded. More capital continues to enter the space - some mental health-specific investors are raising subsequent funds and new investors are entering the space - like Meridian, who launched the Innovations in Mental Health Fund and just this week, announced their first mental health investment (in Affiniti).

Hi friend,

I’m interested in creative ways we can use AI to improve population mental health.

AI coaching or AI therapy gets most of the attention. They have huge potential.

But they are only one application of this powerful technology. There is a large white space for other use-cases that address some of the underlying factors that lead to good mental health and a fulfilled life.

This week, I stumbled upon a very creative project in this white space. It’s all about using assistive relational intelligence to build distress tolerance and relational capacity, and I want to tell you about it.

We know healthy relationships are a strong protective factor for good mental health and a determinant of a fulfilled, satisfied life. But many societal and cultural forces are causing people to lose relationships and become increasingly isolated. While technology may contribute to this, it doesn’t have to.

If applied correctly, it might be able to bring us closer together - Slingshot’s recent study showed early signs of how this might work.

The problem is that relationships are hard. I believe a lot of our relationship challenges are born from what our Gen Z friends would call “a skill issue”. We don’t know how to recognise our own emotions, we don’t know how to regulate ourselves, we don’t know how to communicate in non-violent ways, and we don’t know how to be curious instead of combative in times of distress. These skill issues limit our ability to develop meaningful relationships.

This is the problem Jocelyn Skillman is trying to solve.

Jocelyn is a licensed therapist, clinical supervisor, educator and design ethicist. She is also the founder of Tend Messages, a product aiming to use assistive relational intelligence to reconnect us with our own nervous systems and to build capacity for human-to-human intimacy. I like that idea. So let me explain how it works and why I think it’s interesting.

Tend is a messaging platform. And while the product itself is still extremely nascent, the idea is what is interesting. Tend is designed to be used to chat with your partner, friend, sibling, whoever. It’s for times when you may be feeling something strongly (perhaps in a dysregulated way), when you don’t know what those feelings really mean or how to communicate them. We’ve all been there.

In such a moment, you can go to Tend and dump in the authentic, un-edited message that you want to tell your loved one. Tend then translates that message and sends it back to you in a way that maintains your authentic voice but reframes the message using principles of nonviolent communication (NVC).

With this message, you have a few choices. If you don’t like it, you can just delete it. If you like it but it’s not quite right, you can edit it. And then, you can send it to your loved one. The idea is that this “Tended” message would remain an authentic representation of your emotions and needs, but would be articulated in a clearer, non-violent, more curious way - one that has a better chance of improving the relationship with your loved one than the unfiltered message you wanted to send during a period of dysregulation.

Many people are using ChatGPT for editing or co-authoring use cases like this. But the messages are overly polished and disembodied. They rarely feel like an authentic representation of your emotions, and the recipient can smell that. I think using ChatGPT to edit messages may be more harmful than helpful. But by using prompts built around NVC and designed to reflect the authenticity of the user’s original message, Tend aims to be different.

At the same time, Tend also helps you understand more about your own emotional state and improve your ability to self-regulate. It shares perspectives on your feelings and needs. For example, it asks, “I wonder if you’re feeling angry? Are you feeling a need for greater control?”.

There are a lot of interesting features that could be added to a product like this. For example, Tend could provide a psychodynamic assessment of your communications with your loved one over a period of time. This could surface common themes in your relationship and provide guidance to you and your loved one on how to address those themes.

Now, inserting an AI into our relationships might be an awful idea. But if implemented thoughtfully, it might help us to develop the skills needed to have more, better relationships.

The ability to communicate our feelings is not the only skill issue that could improve population mental health. How might we improve our emotional regulation, our distress tolerance, our self-compassion, or our repair skills? If you’re building in this space, I think these are interesting questions to consider.

Jocelyn describes the goal of Tend as “using AI to build distress tolerance and relational capacity so that when you do turn to a human, you’re more resourced, more able to name what you need, more capable of repair when shit hits the fan.”

She recognises that she doesn’t know if it will work. But thoughtful, creative experimentation like this is exactly what we need more of. 

That’s all for this week. Have thoughts? Reach out and let me know. I love hearing from you.

Join The Hemingway Community

We now have over 240 members in our vetted community for mental health innovators. This includes founders, clinicians, investors and researchers, all passionate about improving population mental health. We discuss important topics, share learnings and host events. If you’re interested in networking and collaborating with these kinds of people, feel free to check it out.

Keep fighting the good fight!

Steve

Founder of Hemingway


Note

[1] Tend is still in closed beta, but you can learn more about it from Jocelyn here.

[2] Jocelyn writes on Substack, and her work is worth reading. The commentary I found most insightful was her work on;

• The asymmetric weight of healing. Jocelyn describes how therapists now carry "the accumulated weight of social fracturing, vast unintegrated trauma, and the metabolic debris of a culture that has outsourced emotional labor to professional containers while dismantling the structures that once provided us vital safety and solace." The therapeutic dyad was designed as temporary scaffolding that would launch people back into mutual care, but it's become permanent because there's nowhere to return to. The asymmetry isn't just unfair to therapists; it's a symptom of broader social failure.

• The etiology of our current challenges. In the same post, Jocelyn identifies the collapse of relational infrastructure as a root cause of many of our societal mental health challenges. She outlines how atomised living severed us from extended kin networks, economic precarity keeps us churning in isolation, digital mediation offers responsiveness without embodiment, and how the loss of shared ritual has separated us from a practice that once synchronised our nervous systems collectively. Our system treats individual pathology, but for many, the real issue is systemic fracturing. No amount of clinical innovation can address the underlying erosion of

Hi friends,

As you read this, the FDA Digital Health Advisory Committee is meeting to discuss generative AI in mental health.

The outcomes of this meeting will influence how AI in mental health is regulated, which will, in turn, directly influence the strategies for everyone building AI in this space.

As part of this process, the committee invited people to submit comments. Thirty-one organisations and individuals did so, including Spring, Talkspace, Big Health, Slingshot, Headspace, Otsuka and more.

I’ve read all these comments, and in today’s report, I share what I learned.

I’ll discuss where there is consensus and where there is conflict in how organisations think this technology should be regulated. I also highlight some of the more interesting ideas that I found in these submissions, and the reflections I had on what this might mean for the industry.

If you want to read the original submissions yourself, I’ve collated all of them in a database here.

But if you have a better way to spend ten hours of your life (which I clearly don’t), then just take the ten minutes to read this report instead.

Let’s get into it.

Join The Hemingway Community

I run a vetted community for mental health innovators, and we now have over 230 members, including founders, clinicians, investors and researchers. If you’re interested in networking and collaborating with these kinds of people, I’d suggest checking it out.

We discuss important topics, share learnings and host events. We’re actually hosting a private dinner for community members in San Diego next week, so if you join the community and will be in San Diego, let me know, and I’ll share the invite with you.

More consensus than conflict

When I finished reading every submission to the DHAC, I had one clear reaction. Despite significant diversity in the types of organisations submitting comments, there was far more consensus than conflict.

Most organisations are mostly aligned in how they think mental health AI should be regulated. Here are the areas where I found significant consensus.

A risk-based, tiered framework

There are no surprises here that this reasonable approach is recommended. Most submissions agreed that low-risk wellness tools should be lightly regulated, and high-risk clinical AI should be subject to stricter review. Oversight should be proportionate and flexible, focusing on intended use and risk level. This is also highly aligned with the FDA’s current approach to regulating medical devices.

Non-negotiable safety guardrails

There was a significant focus on ensuring safety. Models must be able to detect crisis cues (e.g., suicidality), block unsafe advice, and maintain clear escalation pathways. Almost everyone agreed that those escalation pathways should lead to some form of human support. Talkspace urged that the FDA require “human-in-the-loop” escalation protocols and routine testing of crisis recognition functions. SonderMind called for clear system limits - for example, chatbots should never attempt to manage crises autonomously - and for labelling that explains how escalation works. Click Therapeutics described how systems should monitor conversation patterns and semantic drift to catch early signs of psychosis or delusional reinforcement, triggering escalation to human oversight. The American Psychiatric Association urged that crisis management protocols be standardised and auditable, and NASW-Texas warned that AI should never replace the human therapeutic alliance in crisis contexts.

Transparency is essential

Everyone agreed that AI tools must clearly disclose that they are non-human, describe their purpose, capabilities, and limitations. Some organisations suggested standardised “model cards” (like what we have in nutrition labelling) that describe training data provenance, model performance, validation and known risks and limitations. I’m on board with some version of this. I think one of the most powerful things regulation can do is to enable better customer and provider choice by enforcing transparency. I wrote about this in more depth in last week’s report.

Continuous lifecycle oversight is required, but adaptations are needed

There was broad agreement that continuous oversight of models is required and that some version of the FDA’s Predetermined Change Control Plan (PCCP) is essential to make regulation workable for adaptive, learning systems in mental health.

However, some raised questions and concerns about how this would operate in practice. Kooth and Click noted that the current process is vague about what kinds of retraining or fine-tuning count as “significant” changes that would trigger new FDA review. It was also pointed out that most developers rely on third-party foundation models, leaving them with limited visibility into when those base models change. Several organisations proposed ways to strengthen or complement the PCCP: SecuraAI recommended quantitative “drift gates”, which would require each model update to pass regression testing on its proposed “Safety-5” metrics (Guardrail Persistence, Time-to-Escalation, Escalation Completeness, Critical Safety Event Rate, and Equity Gap). Kooth and Spring Health both supported continuous, transparent real-world monitoring of AI safety and performance. Kooth suggested that results be shared publicly, and Spring promoted open, repeatable evaluation frameworks such as their own VERA-MH. Click and the Connected Health Initiative called for a certification pathway for “validated foundation models” so downstream developers can rely on trusted baselines. The American Psychiatric Association suggested periodic third-party or FDA spot audits to verify compliance.

Overall, there was a push for the FDA to evolve PCCP into a more dynamic system that blends pre-specified change plans with real-world monitoring, quantitative safety thresholds, and independent verification.

Privacy, fairness and data protection are table stakes

Nearly all submissions call for strict data encryption, informed consent, and transparency on data use. The American Psychiatric Association, ABHW, and NASW-Texas all emphasise that mental health data is uniquely sensitive and should never be used for marketing or retraining without consent. NASW-Texas goes further - arguing that users’ data should remain encrypted and non-transferable even if the company is sold. SecuraAI and Livio Labs also propose regular bias and equity audits to prevent demographic disparities in performance.

Some areas of divergence

Despite this significant consensus, there were a few areas where submissions diverged in their recommendations.

The level of human oversight

No one outright rejects the need for human oversight, but opinions diverge on how universal it should be. Clinician-led organisations like the American Psychiatric Association and NASW-Texas insist that AI should never operate autonomously in mental health contexts. Companies like Talkspace and SonderMind largely align with this approach and believe that AI should support, and not replace, clinicians. Wellness-oriented products, like Slingshot, argue for more proportionality however, i.e., that low-risk products shouldn’t be forced to maintain human supervision.

Liability

The APA and NASW-Texas want the FDA to shift liability from clinicians to developers. If an AI system gives harmful advice, they argue, the company - not the provider using it - should bear responsibility. No one wants to discourage clinical adoption by making practitioners the scapegoats, but regulators still need to define who’s legally responsible when AI goes wrong.

The boundary of wellness vs. clinical care

This is the hardest regulatory line to draw. Kooth and M.Cert argue that “subclinical” tools - for prevention or early distress — need their own category. Kooth makes some strong statements here, saying that; “The current regulatory framework does not adequately capture these ‘subclinical’ tools that influence mental states and behaviors but do not make formal medical claims” and that “FDA should consider a new classification for subclinical digital mental health technologies, ensuring minimum safety, transparency, and evidence standards without imposing full SaMD requirements.” Slingshot disagrees, warning that dragging wellness tools into FDA territory would crush low-risk innovation. Deciding if a new category is required, or where to draw the line between wellness and SaMD will be top of mind for the FDA.

Insights and questions

A new regulatory category for Mental Wellness AI?

As discussed, several submissions (including from Headspace) envision a formal middle ground between wellness and clinical tools for subclinical products that don’t diagnose or treat disorders but meet defined standards for safety, privacy, and efficacy. In their public post, Headspace outlined what this might look like.

Headspace’s proposed framework.

This would have interesting implications for general-use chatbots, like ChatGPT. For example, if this framework were to be adopted, ChatGPT would be required to shut down and redirect sensitive mental health conversations. If it wanted to provide mental wellness support, it would need to adhere to the suggested safeguards around transparency and safety. Defining “mental wellness support” and “sensitive mental health conversations” could be challenging, however.

Now, I’m no regulatory expert, but I would be surprised if the FDA created a new category like this.

What are the right standards and evaluations for Mental Health AI?

Several organisations questioned what the right evaluations should be for mental health AI. Some proposed their own standards, like Spring’s VERA-MH. What is clear is that there is currently no widely accepted standard for this technology. Last week, I wrote about how this lack of consistent standards is holding back progress. One of the most impactful things organisations could do is to collaborate on this topic and align on a shared set of standards that can be used to evaluate and improve AI products in mental health. I’m sure regulators and payers would thank us for this too.

The challenge of evidence generation

Evidence generation is a complex challenge for generative AI in mental health. We would all agree that AI in a clinical context should be supported by high-quality clinical research. Big Health made this point explicitly in their submission. But the generative nature of these AIs makes this difficult - in short, because every interaction is unique and difficult to predict.

Traditional trial designs struggle to capture that variability. These trials tend to be conducted over relatively short time periods and are therefore highly affected by engagement rates - something which can fluctuate greatly over longer time periods. Big Health proposed using validated reporting tools to collect large volumes of real-world evidence, helping to reduce statistical noise and provide a clearer picture of effectiveness across diverse users.

While the current approaches to evidence generation do an OK job of guaranteeing safety and effectiveness, they are expensive, long, difficult and limited in their ability to understand real-world, long-term impacts. I don’t know what a solution for better evidence generation looks like in this space, but I know we need one. I’d be very curious to hear people’s thoughts on this topic.

Could AI create a two-tiered mental health system?

Some comments suggested that AI might deepen inequities in mental health care. NASW-Texas warned that if AI is used as a cheap substitute for clinicians, low-income and publicly insured patients might end up with bots while wealthier people get humans. This is a fair concern. We don’t want human therapy to be available only to those with privilege (although arguably, to some extent it already is). However, that must not stop us from trying to develop more accessible forms of treatment. My view? We can do both - work to increase access to human treatment, whilst also innovating on more accessible AI solutions.

Behind every AI is a human making choices

In its submission, Slingshot AI reminded us that every AI is designed, and that the people building them make choices as to how those models behave; “Every foundational AI is designed, and through that design process, specific choices are made that shape an AI’s behaviors, reward incentives, and impact on users.” 

Every model reflects its creators’ assumptions and choices. While regulation will enforce standards through legal frameworks, we must also remember our personal responsibility to hold the people building these products responsible for their performance. Often, social expectation does more to shape behaviour than any rulebook.

As for what comes next, we’ll know more after today’s session. We’ll be discussing it in the Hemingway Slack Community, so feel free to apply if you want to be part of the conversation.

That’s all for this week. Have thoughts? Reach out and let me know. I love hearing from you.

Keep fighting the good fight!

Steve

Founder Hemingway

Hi friends,

Mental health AI has more questions than answers.

In my emails and DMs, those questions are piling up.

Questions about strategy, product, regulation, research and commercialisation. Today, I share my thoughts on many of those topics.

To do so, I’ve taken inspiration from Montaigne, the pioneer of the essay. In fact, Montaigne coined the term "essay" (from the French essai, meaning "attempt") to describe his writing.

His essays selected a specific topic and examined it through his own experience, with titles like “Of Friendship”, “Of The Education of Children” or even “Of Cannibals”. He also wrote extensively about his bowel movements, but I promise to spare you from such a topic.

Instead, in this THR Pro report, I examine 8 key issues for those building AI in mental health - eight mini-essays, sharing my own perspectives on these topics.

We’ll cover:

• On Going Slow: Why we’re misunderstanding the maturity of AI in mental health and what the real winners will do to develop long-term defensibility.

• On Creativity: Why we must dream a little bigger and explore the opportunities of a Therapy 3.0 world.

• On Democratising Wisdom: If Google organised information, can AI organise wisdom?

• On Standards: How a lack of shared standards is holding back progress and what we might do about it.

• On Transparency and Choice: Should users have a right to understand the values and incentives guiding the models they use?

• On Client Acquisition: How AI can support client acquisition and what businesses must get right to do so.

• On Storytelling: How the asymmetry of storytelling is shaping the market and the narrative opportunity that exists for stakeholders.

• On Opportunity Costs: Why AI is not a panacea and how leaders should think about the trade-offs of AI investments. 

Let’s get into it.

Hi friends,

ieso is pivoting its GTM strategy. I had a chat with Kent Tangen (their CEO) to talk about why. 

Over the past 25 years, ieso has delivered over 815,000 hours of text-based therapy to more than 145,000 patients with anxiety and depression. The business has also raised over $70M and has been adopted by some of the largest health systems in the world.

With this capital and traction, they built a large and rigorous clinical mental health dataset. A dataset which has powered their text-based interventions. 

For most of their long history, these text-based interventions were delivered through a mobile app. 

But that’s changing…

ieso are going all-in on an API-first model that white-labels their clinical AI product, Velora, directly into partner platforms. Virtual care companies, chronic condition management platforms, and anyone with an existing user base who wants to add mental health AI capabilities (without building the product and associated evidence base) can do so.

I wanted to learn more about this new GTM approach, but also to hear Kent’s views on the market, the thinking behind his strategic choices, and the advice he has for other leaders in this space. 

So we sat down and talked about his decision-making process, what's working with partners, what investors are asking about, and his approach to AI regulation.

I hope you find it as valuable as I did.

Let’s get into it…

Remember, the Inside The Lab event is next week. So if you’ve been thinking about going, it’s pretty much your last chance to sign up. I know lots of great folks who will be there, the scheduling looks awesome, and there’ll even be a startup pitch showcase on the final day. In short, I’d highly recommend it.

AND, if you want 20% off, just use ITL2025GJMMM at check out.

Steve:  Kent, how are things? You've made a lot of big strategic changes with ieso over the past year. What were you seeing in the market that led to these changes?

Kent: When I was having conversations with payers and health systems over the past year, something became clear: people really did not want a single solution anymore. One-off mental health apps weren’t gaining traction. People don’t stick with them. Everyone I spoke with was aware that activation and engagement of digital mental health interventions are some of the hardest problems in this market. And point solutions weren't fitting the bill.

So we decided to focus on integration over isolation. I said, let's find a way to fold our product, Velora, into existing digital care solutions. Let’s meet people where they are instead of asking them to download and log into yet another app.

Steve: Tell me a bit more about Velora and the thinking behind this new GTM approach.

Kent: Velora is a generative AI program that’s intentionally modular and time-limited by design.

And we take safety and evidence seriously. That’s imperative in this field. We just released a new study preprint (not yet peer-reviewed) that found zero AI-induced safety incidents across both simulated and real U.S. users and measurable improvements in anxiety and depression scores. We also have a peer-reviewed study with an earlier version of our product that shows outcomes on par with human-delivered care.

And we have a powerful dataset that is linked to outcomes that underpin our product. Those are our true capabilities and what we’re known for at ieso (not necessarily for having the coolest mobile app). So we took that product and exposed the APIs. 

We saw this as an opportunity to unlock a new GTM approach, with new customers, while doubling down on what we do best. The divestiture of our UK telecare business was part of that strategy: simplifying the business to concentrate fully on our evidence-based mental health AI platform.

Steve: OK, fascinating. Walk me through the mechanics here. Who are you providing this to, and how does it work?

Kent: We're talking about partners who have a virtual healthcare platform and want to add clinical AI capabilities for mental health. These might be mental healthcare providers who want to expand their service offering, or folks from other areas of healthcare (like chronic condition companies) wanting to add on mental health support.

By exposing our APIs, we allow these partners to integrate safe, evidence-based clinical AI into their platforms. Partners can integrate our capability as they see fit into their product to best serve their users. So one of their users could log in, be offered support for their mental health, and then, within that app (with the app’s branding, etc.) be offered our text-based mental health AI offering. 

Me: And you're surfacing the full ieso conversational product, Velora, through the API at that point?

Kent: Yes. What was previously the mobile app being served by APIs underneath is now integrated directly into a partner's existing solution.

Here’s another potential use case… imagine a patient coming to a mental health organisation’s website and trying to find a therapist. Well, if that organisation was using the Velora API, they could decide to offer our clinical AI to the patient at that point, while they are waiting for a therapist, for example. 

Steve: I guess this strategy is based on solving problems for these partners, then. So what problems do these organisations have that a white labeled ieso product will solve?

Kent: Two things. First, they are thinking about how they can grow. They know there is consumer demand for conversational AI products in mental health and believe that adding these capabilities may support their client acquisition. Second, they’re thinking about how they can identify savings.

Steve: So, meeting consumer preferences to drive growth while potentially reducing cost to serve.

Kent: Right. And we want to help them achieve both of those things.

Steve: Are the economics significantly different for you with this model?

Kent: Not significantly. We still think about the same value proposition. Previously, it was clinical APIs in a mobile app. Now it's just the APIs in someone else's solution. We're still making the same outcome statements. The economics are very similar.

Steve: I’m keen to get your advice on GTM approaches in general. There are a lot of people building in this space right now, especially with generative AI. When it comes to product and GTM decisions, what have you learned from your time at ieso that you would want them to know?

Kent: Always think about the patient first. Your patient and then your customer - and of course, the challenge in healthcare is that they are usually different people. But you should deeply understand the patient needs and the customer needs that you are going to solve. It's easy to lose sight of that. Don't.

The other thing is, you've got to meet people where they are. Activation and engagement cannot be an afterthought in digital mental health. You have to start with the patient and work your way back from that. Too many people start with "I've got this really great solution, do you see what it does?" without thinking about how people will actually use it in clinical settings. 

Join The Hemingway Community

We recently crossed 200 members in the Hemingway community. If you’d like to apply to join this vetted community for innovators shaping the future of mental health, you can learn more here.

And if you want to hear more from members on what it’s actually like, here’s some of my favourite snippets of feedback.

Feedback from Hemingway members

Steve: A big question people have is around choosing who that customer should be. What advice do you have there?

Kent: Your customer needs to align with your strengths. Be clear on what you're good at and what your superpower is. If you're aligning with a customer making demands outside of those capabilities, it can get squirrely.

For instance, if you serve moderate to severe depression and anxiety, and a customer asks, "Can you help my substance abuse population?" - it's easy to say yes. But that's not trivial. If you're making outcome claims and trying to stand behind your data, your quality will be questioned if you're not being the authoritative source of impact.

Steve: This is interesting. I think it’s been hard for companies to stay focused on specific capabilities and niches. One reason is that mental health businesses seem to need significant scale to survive, and that’s because you need the distribution advantage that comes with scale.

But perhaps, if people are finding new ways to distribute their product (through APIs for example), then they can stay more focused on their core capabilities and let other companies manage the patient relationship. Thoughts?

Kent: Exactly. I think we'll see more of those pathways occurring. For example, if you're working with a patient and after three sessions you acknowledge an opioid or alcohol addiction, you should recognize that and be able to offer them a relevant third-party solution that is better suited to meet their needs. 

Steve: Let’s chat regulation… Regulatory strategy is top of everyone’s mind right now. There’s a lot of uncertainty, and people are worried their product strategy could become defunct in six months if new state and federal regulations are passed. What's your take on how to build AI products given the current regulatory environment?

Kent: Well, I would not do "wait and see." I think you need to pick a path pretty early and commit. Part of the reason is that there's a long runway for building these kinds of products. So if you do want to go down certain paths, waiting and seeing could be problematic.

What I would say is: consider the regulatory path that would be most constraining to your business and start planning for that. “OK, this is what it's going to look like, this is what we're going to do, this is what it means for our product development cycle, this is what it means for reimbursement opportunities.” Role-play all of that and understand the impact that would have on your business. 

At the same time, keep a careful eye on how the environment evolves – both the regulatory side and the technology itself, which is moving faster than the policies meant to govern it. That pace of change is exactly why “wait and see” isn’t a strategy. 

Steve: So, assume the most constraining regulatory scenario and understand what that would require. Then look at it and say, "well, we can't pursue this", or "we have to make this work"?

Kent: It depends on the opportunities in your business. You have to do the modelling of what your business would look like one or two years out, after that is completed, and what would that actually open up.

If you're an early-stage company, you're going to say, "Whoa, I’m gonna need a lot of capital, it's going to be a long runway, and I'm not sure what that's going to look like." But it's also going to open up opportunities. You need to consider what those opportunities are going to look like.

Steve: What are you hearing from investors in 2025?

Kent: The investor conversations I've had over the last year have changed. There is much more of a harder drive on regulatory strategy than there was a year ago. What is your regulatory plan? What's your strategy? How are you thinking about it? What are the implications of doing it versus not doing it?

There's also more belief that there will be greater go-to-market opportunities than there were previously. There is so much focus on digital mental health and so many conversations happening. Even though they're confusing at times, they're happening because there's a lot going on in this space and a lot of appetite for safe, high-quality solutions that can help more people. 

Steve: Looking ahead, what do you think you have to get right to succeed in this space?

Kent: Outcomes. Everyone talks about them, everyone wants them, but proving them is what counts. The ability to appropriately demonstrate the impact on greater access to high-quality mental healthcare and what that does to overall healthcare costs – whether in a generalised behavioural telehealth partner or in a chronic condition partner, is critical.

Also, demonstrating that mental health actually does impact the top and bottom lines and makes people healthier.

When mental health improves, everything improves. People do better, organizations do better, systems do better. Society does better.

At the end of the day, that's what we're all here focusing on. We at ieso have a strong belief and evidence that if we do that, the outcomes will follow – and we'll be making people better.

Steve: Kent, it’s been a pleasure. Thank you for sharing these insights with me and the Hemingway readers. I really appreciate it.

That’s all for this week. As always, reply to this email and let me know what you found helpful.

Keep fighting the good fight!

Steve

Founder of The Hemingway Group