AI agents browse the web in search of knowledge, but Spanish entrepreneur Elisenda Bou-Balust is convinced that your agents need better data. “The Internet isn’t the best repository of information we have as humanity,” she told The Infinite Loop.

Her latest startup, Cala AI, just came out of stealth with an API-like data offering targeted at engineering teams whose AI products need accurate and up-to-date information. This is an alternative to hallucination-prone LLMs, with the additional advantage that Cala’s structured data can be incorporated directly into code — no scraping or processing required.

The underlying concept is called knowledge graphs — which just happens to be the reason why Apple acquired her previous company Vilynx in 2020, and the topic she worked on for the Cupertino giant alongside her now co-founder and CTO, Issey Masuda Mora.

This also explains why 15-month-old Cala is already one of Europe’s hottest AI startups.

Owing to Bou’s track record and ambition, she secured funding almost right after Cala’s inception — a €7 million pre-seed round led by American VC firm Lightspeed Venture Partners, with participation from Spanish VC firms Kibo Ventures, Kfund and Masia.

The quietest big raise in Spanish tech

While Cala was only just an idea at the time, many journalists would still have loved to hear from Bou about this being Spain’s largest ever pre-seed round, as well as Lightspeed’s first investment in Spain, and to top it all, that she raised it while pregnant with her second child.

However, Bou chose to stay quiet for more than one year, only revealing the information when Cala opened up its platform. “I believe funding rounds are just a means to deliver a product, so until we actually had something to show, I didn’t want to say much,” she explained.

Bou knew it would take several months for Cala to be able to launch. While its product requires less capital than building a foundation model, it is also more complex than a wrapper. “At the infrastructure and architecture level, we’re pushing vector databases to their limits — we’re building systems that need to be 10 or 20 times the size of Wikipedia today.”

Cala AI - knowledge query (video screenshot)

Bou explained that Cala aims to be the layer that will enable AI agents to have data they can rely on to act reliably. This is particularly crucial in B2B contexts where LLMs have shown their limitations, but it also unlocks new use cases.

Cala’s bread and butter are scenarios in which verified information is key, but it is just as important that it returns that information in an easily digestible format — structured JSON with source citations. “[Agents] don't need to browse [the web], they need to navigate data silos seamlessly and get exactly what they need,” Bou wrote on LinkedIn.

Among Cala’s users, she said, some “are tracking delays at ports to see how shipments are being delayed and how that affects product prices.” But Cala is also used by startups for quantitative analysis and other financial applications, procurement, human resources, legal topics, and many more. “Everyone uses data all day long,” Bou said.

The graph that grows with every query

What’s fairly new, however, is the rise of agentic AI. But Bou is now well aware that most of its users won’t be humans. According to Cala’s FAQ, the platform can be queried via three interfaces, including MCP, making it callable from any framework with tool support.

Cala’s pricing still refers to seats but is mostly usage-based, with the understanding that its users will be both people and their agents.

While enterprise customers could bring in significant revenues, Bou acknowledges that building a whole new data layer is a capital-intensive endeavor. But even now, all she wants to think about is Cala’s roadmap. “We’re getting a lot of requests from investors, but I believe in ‘deliver first,’ and that’s what we’re currently focused on: building the product.”

For Cala, building its product also means building its graph. Its data comes from public sources on the open web, including APIs, filtered and verified before entering the graph. The approach is not unlike OpenAI’s agentic search capability in ChatGPT, Deep Research, Bou concedes — but claims it operates “on superpowers.”

“The problem with Deep Research is that you do a Deep Research query, it gets processed, you spend a ton of credits and tokens, and then it’s discarded,” Bou claimed. Her view is that it doesn’t make sense to search the web every time your agent needs a tidbit of information. By keeping those data points available for future queries, Cala claims to be “eight times more token-efficient.”

The corollary is that Cala grows with user queries, Bou explained. “When you ask it about a topic, if it doesn’t know about that topic, it goes to look it up, processes it, and has it ready for your next query.” As a result, the company projects its knowledge graph will grow to between half a billion and a billion data entities by the end of 2026. 

Making this happen will also require Cala to grow. With a headcount of 20 as of early May 2026, the Barcelona-based startup is hiring, mostly for go-to-market and growth, and plans to open an office in San Francisco. This also reflects how Bou envisions Cala: as a global company with a Spanish soul. She’s not afraid to admit it is a moonshot bet, and while the world could use more trustworthy information, it is too early to tell whether Cala will land its rocket.

From 2024 to 2025, doctors at Yale Medicine reduced the time it took to evaluate patients with abdominal aortic aneurysms by 73%.

This improvement, for a condition where fast treatment saves lives, wasn’t down to employing more doctors or nurses, or a new breakthrough drug discovery. It was thanks to the Yale team using an AI tool that analyzes images from patients’ scans, combined with other medical records, to support clinical decision-making.

“It's another pair of eyes that reduces a chance of an overlooked finding, that's the feeling from the clinicians,” said Tom Valent, chief business officer at Aidoc, the Tel Aviv-based AI medical imaging company that built the platform.

At a time when global health systems are under increasing pressure, due to aging populations, rising rates of disease and shortages of skilled healthcare workers, AI can help doctors make sense of complex medical images more quickly and consistently. It’s also being used to bring medical imaging into areas of healthcare that have remained stubbornly analog, making treatments cheaper and more convenient for patients.

The hospital that doesn't talk to itself — yet

Aidoc’s platform can be used to analyze medical images from scans across different areas of medicine, ranging from cardiology to neurology and radiology.

One of the big problems that AI can solve in healthcare is joining up different parts of the system which don’t normally interact, Valent explained. If Aidoc’s vision models are given a new patient scan to analyze, the platform combines that data with information from electronic medical records, which can give doctors important context they might otherwise not have had.

“Oftentimes in big hospitals, there's a coordination issue between different specialists,” Valent told The Infinite Loop. “We basically make care in the health system higher quality, more proactive, and more connected between different specialties.”

Aidoc’s platform analyzes medical images to speed up triaging. Credit: Aidoc

Aidoc says more than 1,600 hospitals are currently using its technology, and that the platform can help spot urgent cases that might be at the back of a queue, before a busy doctor has a chance to look at a scan. In medical areas like radiology that can make a big difference if a cancer patient is assessed faster.

“In hospitals with big backlogs and long turnaround times, the radiologist knows that the chance of a case at the bottom of their queue having something urgent, and not being seen in a time period that’s potentially critical, is much lower with Aidoc,” said Valent.

Using AI models to make decisions on who might receive medical care fastest raises concerns over whether algorithmic bias might mean that certain demographics of patients receive care before others. This is something that Aidoc invests heavily in through its AI safety team, and is a crucial step in receiving regulatory approval.

“We have more than 60 people that are focused on AI safety, both in the development stages and in the post-market stages, when the algorithm is already running in production,” said Idan Bassu, chief R&D and AI officer at Aidoc.

Half the cost, ten times the patients

AI is also bringing medical imaging into areas of healthcare that have previously relied on extremely manual workflows.

Barcelona-based Impress uses AI to help orthodontists create treatment plans, based on scans of their teeth that are analyzed by a vision model trained on more than 500,000 patient cases. Traditionally, orthodontists will look at a patient’s teeth in person, requiring people to make regular trips to a clinic to assess progress. 

Impress uses data from an initial in-clinic scan, combined with proprietary AI, to predict the outcome of a treatment. Credit: Impress

IImpress reduces these visits by letting patients take scans of their teeth at home with a smartphone, with the visual data then analyzed by AI that recommends the next steps in the treatment plan. The company says this allows doctors to treat 10 times more patients in the same amount of time, reducing the cost of Invisalign-style treatments to around €3,000, compared to around €6,000 at traditional clinics.

An Impress patient taking an at-home scan to monitor treatment progress. Credit: Impress

Impress CEO and co-founder Vlad Lupenko explained how every patient case is overseen by a doctor, and that AI shouldn’t be used to write clinical oversight out of the process. The risks of AI being used to take decision-making out of doctors’ hands were laid bare in 2020, when a chatbot made by the now bankrupt healthtech company Babylon missed symptoms of a heart attack in a patient.

“AI's an additional tool that helps to increase the efficiency of doctors’ time and their accuracy, but it will never replace the doctor,” Lupenko said.

By combining imaging and AI technology with clinical expertise, to make treatments cheaper and more convenient, Impress, which describes itself as Europe’s biggest provider of clear aligner treatments, has grown to treat nearly 50,000 patients every year.

Compute where the patient is

Impress's platform analyzes more than 180,000 scans every week. Aidoc's system processes 60 million patient cases per year. At that scale, the demands on cloud infrastructure are significant — and in healthcare, they come with constraints that don't apply elsewhere.

“Flexibility is critical. We need to be able to scale GPU and compute capacity dynamically,” Impress CTO Yerzhan Tashbenbetov told The Infinite Loop.

On top of regulation-compliant infrastructure, Aidoc adds that it needs cloud compute that’s geographically close to hospitals, to ensure fast speeds and support timely clinical decisions.

“One of the reasons for multi-regional compute is latency. We want the cloud hardware to sit as closely as possible to the customer,” said Idan Bassu, chief R&D and AI officer at Aidoc, adding that the huge quantities of data that Aidoc processes demand cutting-edge efficiency architectures.

“You have the ingestion of a very high volume of data you need to handle. You need a cloud environment that can process hundreds of millions of events from both scans and medical records every day. We need very, very efficient data streams.”

Aidoc and Impress are solving different problems with the same underlying bet: that AI trained on enough images, running close enough to the patient, can do something no doctor working alone can.

After over a decade of development, AI drug discovery has moved from the lab into the clinic. 

More than 170 AI-designed drugs are now being tested in humans. These trials will prove, for the first time, whether molecules discovered by machines can safely and effectively treat patients.

In Salt Lake City, Recursion has multiple AI-enabled drugs in clinical trials — and a growing pipeline behind them. “Solving disease with AI relies on three main ingredients that have now come together: algorithms, computing, and methodology,” CTO Ben Mabey told The Infinite Loop.

Across the country in Cambridge, Massachusetts, Lila Sciences is building a system in which AI doesn’t just assist in drug discovery, but acts as an autonomous scientist. “The model’s ability to make very good educated guesses at what a good starting molecule might look like has gotten very good,” says CTO Andrew Beam.

In nearby Boston, Insilico Medicine has already reported Phase IIa clinical trial results for its pulmonary fibrosis drug rentosertib — one of the first indications that an AI-discovered small molecule can be used to treat humans. “Many of the common hurdles, like getting the project funded, identifying the target, generating the molecule, and demonstrating results in the clinic, we managed to overcome already,” said CEO Alex Zhavoronkov. “However, the final phases of the clinical trials still remain.”

That remaining stage is where success will be determined — and where most drugs traditionally fail.

The compressed front end 

For decades, drug discovery was constrained by search. To find a viable molecule, researchers used robotics to physically test millions of existing chemical compounds against a disease target. The process was slow and arduous, as candidates were limited to previously identified chemicals stored in physical libraries. Refining the top contenders into safe, effective medicines would then typically require years of manual trial and error.

AI is changing that by reducing the search space. Models trained on biological and chemical data can prioritize candidates from a mathematical map of countless molecular combinations. They can predict their properties and guide optimization before the physical testing even begins.

Insilico validated the approach with the development of rentosertib. In 2023, the molecule became the first AI-discovered and AI-designed drug to reach Phase II clinical trials. 

Alex Zhavoronkov, CEO of Insilico Medicine. Credits: Alex Zhavoronkov

It began life when AI identified the protein TNIK as a potential cause of pulmonary fibrosis. Using generative adversarial networks (GANs), the system then assembled a new chemical structure that could lock into the protein’s shape, preventing it from working. Essentially, AI had "imagined" the ideal molecule to treat TNIK. Integrating reinforcement learning, the system then refined this structure to ensure it could survive the human body.

Insilico is now working to scale this approach. Twelve of the company’s drugs have reached the clinical stage, bolstered by two fully automated wet labs, Life Star 1 and 2, which test AI-designed molecules to create a real-time data feedback loop.

According to Zhavoronkov, tasks such as target identification, molecule generation and safety prediction are now performed “almost instantaneously” by AI — albeit with some oversight and modification.

“Even some experiments are run in the end-to-end fashion where AI is controlling the experimental flow,” he said. “However, when it comes to later-stage experimentation and iterative improvement of both the drug and the AI system, we still need a significant number of humans.”

That approach is now translating into commercial traction. In March 2026, Insilico signed a $2.75 billion deal with pharmaceutical giant Eli Lilly to bring its molecules to the global market. The company has now developed at least 28 drugs using AI, nearly half of which are already in clinical development.

Biology as data

Recursion takes a different approach: modeling biology as a data problem. 

The company has built datasets of billions of microscopy images showing how human cells respond to genetic and chemical perturbations. These are combined with transcriptomics and patient data to create machine-learning representations of biological systems.

The goal is to identify relationships between genes, proteins, pathways, and diseases before selecting a drug target. To accelerate this shift, Recursion has developed large foundation models built on transformer architectures.

The models resemble those behind the LLMs produced by frontier AI labs, but with a different data type. “Instead of being applied to text and natural images, we apply them to the language of biology — cellular images, proteins, RNA, DNA,” said Mabey.

Recursion’s infrastructure reflects the scale of that task. The company operates two supercomputers and trains models on more than 40 petabytes of proprietary biological data.

BioHive-2, Recursion’s in-house supercomputer. Credits: Recursion

“Deep learning is data hungry, so that has to be paired with data generation,” Mabey said.

These models have been put to work identifying biological relationships, predicting drug effects, and prioritizing compounds for development before any physical synthesis begins.

Recursion also combines its own compute infrastructure with cloud systems to handle spikes in inference and training workloads. This allows the company to scale models without relying entirely on hyperscalers.

The autonomous lab

Lila Sciences is pushing the model to a new level: integrating AI directly into the experimental process.

Rather than treating the lab as a separate validation stage, the company is building systems in which models generate hypotheses, test them experimentally, and continuously learn from the results. “We’re taking that whole process and putting AI at the centre of it,” said Beam.

In practice, this means turning laboratory instruments into software-controlled systems that AI models can operate directly. Robotics systems then move samples between machines, allowing experiments to run continuously without manual intervention. The company described these systems as “AI science factories.” Beam called them “​​the world’s largest verifier for science.” 

He compared the approach to the shift from trolley cars on rails to self-driving systems. Traditional automated labs, he argued, can repeat predefined workflows, but struggle to adapt dynamically. Lila’s system, by contrast, can assemble new experimental sequences in response to model outputs. 

As the experiments run, they create a cycle of data generation and feedback that allows the models to be continuously enhanced.

Beam described the process as “scientific self-play,” a concept inspired by DeepMind’s AlphaGo mastery of the fiendishly complex board game Go. Initially, the software was trained to predict the moves that humans would make, but that alone didn’t produce superhuman play.

“The model has to be able to play the game itself,” said Beam. “It has to be able to play enough games of this to discover and make moves that people have never made before.”

Lila applies this theory to molecules. The lab provides a substrate, and the AI then makes scientific moves researchers would never have considered. Over time, the system discovers strategies beyond the limits of human design.

The approach is already changing biologics research. Traditionally, designing antibodies for new targets relied on large-scale random mutation and screening. Today, models can generate viable candidates computationally before laboratory optimisation even begins.

Andrew Beam, CTO at Lila Sciences. Credits: Andrew Beam

But Lila also has a broader goal. Every scientist, the company said, will soon have their own AI collaborator capable of searching literature, generating hypotheses, and designing experiments.

The path to market

For all the progress in AI-driven discovery, physical human testing remains a tight constraint.

According to research in the US, the clinical phase lasts an average of around 95 months and accounts for 69% of overall R&D costs. Beam put it bluntly: “The biggest bottleneck is obviously still clinical trials.”

After successful testing in labs and animal models, companies must submit an Investigational New Drug (IND) application before beginning human trials. Then the timelines and costs increase sharply. “Once you're on the other side of an IND, you're looking at three to five years and half a billion to a billion dollars worth of money to run that clinical trial,” Beam said.

Approximately 90% of drugs that enter clinical trials fail to reach the market, and those failures constitute the bulk of development costs. But AI can improve the success rates. 

Recursion demonstrated this while developing the cancer drug treatment REC-1245. The process from biological discovery to lead drug candidate took just 18 months — twice as fast as the industry average.

“AI gives us better insights earlier in the process so we have greater certainty that a program will succeed or fail in patients and to identify which patients will be most likely to benefit,” said Mabey.

And it’s not only Recursion that’s reaping the benefits. An analysis published last year found that AI-discovered molecules have an 80–90% success rate in Phase 1 clinical trials. This year, Phase 2 and Phase 3 studies will provide further proof points for dozens of candidates.

Beam is optimistic about the results. He believes AI-designed drugs could reach the market “any year now,” and will eventually become the norm rather than the exception. Completing those clinical trials will be a major step towards this vision. But AI has already moved beyond simply proposing molecules in the lab. It’s now designing drugs that are advancing through human trials — and moving closer to real-world medicine.

Passengers arriving at San José Mineta International Airport’s Terminal B are greeted by a humanoid robot named José. Mounted to a fixed base behind an information desk, José greets travelers and switches to the language they speak, answering questions about flights, baggage and directions.

What engineers behind José are really testing is whether an AI system can stay accurate and responsive inside infrastructure built long before anyone imagined it would need to serve a robot. 

Airports are already heavily automated environments. Conveyor systems route baggage, software coordinates gates and departures, and automated ramp systems manage aircraft turnaround. What is changing is the growing use of AI-powered humanoid robots operating directly alongside people.

The humanoid form factor is not simply for novelty or branding. Human-shaped robots can operate inside existing infrastructure with fewer modifications than many purpose-built automation systems would require.

José, the friendly face of Silicon Valley 

But fitting in physically is the easier part. Passenger interactions are exactly what make airports difficult environments for public-facing AI systems. Passenger information changes constantly, interactions happen under time pressure, and systems must operate reliably despite noise and connectivity constraints. 

This is why the deployment is deliberately constrained. Although José can stand and walk, IntBot, the company behind José, chose to tether the system during the pilot. According to the company, the decision was practical; in a crowded airport environment, a loss of power or battery failure could create obvious safety problems. 

San José Mineta (SJC) and Tokyo Haneda Airport (HND) are among the busy hubs trialing robots in crowded terminals, runway aprons and passenger service areas. With San José preparing to host matches during the 2026 FIFA World Cup, the airport’s four-month pilot is also becoming an early test of how AI systems perform during periods of intense international passenger traffic. 

The airport says José supports travelers in more than 50 languages, a capability that matters in a region as linguistically diverse as Silicon Valley. “What was really surprising to me is that I expected probably 90% of conversations to be in English,” said IntBot product manager Hannah Wu. “It actually turns out that 25% of interactions are in a language other than English.” 

Robot José in SJC. Credits: IntBot

The system combines conversational AI with live flight and airport information, allowing passengers to ask follow-up questions naturally rather than navigating static menus or kiosks. 

“Physical agents” orchestrating models across edge and cloud 

According to IntBot CTO Sharon Yang, José distributes AI workloads across both edge and cloud infrastructure depending on the task being performed. Rather than relying on a single monolithic model, the system acts as what Yang described as a “physical agent” orchestrating multiple models and tools in real time. 

While José can access live flight and airport information in conversations with passengers, the system remains relatively loosely integrated into airport infrastructure. IntBot said the system currently pulls flight status information and floorplan data through APIs rather than operating as a deeply embedded backend system. 

The distinction reflects a broader challenge facing physical AI deployments. Airports are complex environments built around legacy systems, rapidly changing information, and human workflows that were never originally designed around autonomous machines. 

Humanoid robot in ground handling operations at HND. Credits: GMO AI & Robotics Corporation

Mookie Patel, director of aviation at San José Mineta, said the goal was not simply to replace static information systems, but to test whether conversational AI could operate effectively in a live terminal environment. 

“Unlike a static kiosk or an app with pre-programmed responses, José can answer follow-up questions, switch languages instantly and personalize the interaction,” Patel said. “For first-time users, the interaction feels very similar to speaking with a customer service agent.” 

According to Hannah Wu, around three-quarters of interactions are socially driven rather than task-based, with passengers often approaching José out of curiosity before asking practical questions. 

The company previously deployed robots in hotels before moving into airports, partly because multilingual wayfinding and information services offered a practical early use case for public-facing robotics. 

“In a lab, there’s just so many edge cases that you can’t prepare for,” Wu said. “In the real world, that’s how we make our products better.” 

Japan Airlines Haneda trial 

Other airport robotics projects are focusing on labor-intensive work behind the scenes. 

Passengers glancing out of the departure lounge windows at Tokyo Haneda Airport may see the familiar choreography of airport ground crews hauling unit load devices (ULDs), the wedge-shaped containers used to move luggage, freight and mail onto aircraft. But among the workers shifting containers onto rolling ramps, one figure may appear a little more metallic than the others.

Japan Airlines (JAL) and GMO AI & Robotics Corporation are beginning what the companies describe as Japan’s first airport demonstration trial involving humanoid robots in ground handling operations. 

Unlike the public-facing concierge role taken by José in San José, the Haneda trial focuses on repetitive manual work already shaped around human crews. Initial deployments are concentrating on ULD transfer tasks on the airport ramp, with future phases potentially expanding into baggage loading, cargo handling, cabin cleaning and even operation of ground support equipment. 

The distinction matters. While humanoid robots are often framed as futuristic consumer technology, the strongest near-term case for deployment may be in exactly the kind of routine work airports are increasingly struggling to staff.

“While airports appear highly automated and standardized, their back-end operations still rely heavily on human labor and face serious labor shortages,” said Tomohiro Uchida, president of GMO AI & Robotics.

Humanoid robot at HND. Credits: GMO AI & Robotics Corporation

Japan’s aviation sector, like much of the country’s economy, faces growing labor shortages driven by demographic change and increasing tourism demand. Ground handling work combines strenuous conditions with strict safety requirements and operational time pressure, making automation attractive but difficult to implement using conventional industrial robotics. 

According to JAL, the goal is not full automation, but gradual workload reduction and productivity gains through systems designed to work alongside human crews. Using robots for physically demanding tasks would “inevitably reduce workers’ burden” and provide “significant benefits to employees,” Yoshiteru Suzuki, president of JAL Ground Service Co., told Kyodo News. He added that some responsibilities, including safety management, would still require human oversight.

A JAL spokesperson said the integration of humanoid robots and automated vehicles could eventually reduce personnel requirements by roughly half in some container loading tasks, contributing to a broader goal of improving productivity 10% by 2030. 

But scaling those systems beyond pilot projects will depend on proving they can operate reliably in highly constrained airport environments. 

The airport doesn't adapt to the robot

The projects at San José and Haneda reflect two very different visions of airport robotics. What they share is the challenge of deploying AI systems inside environments built around human behavior and day-to-day workflows. 

That complexity is driving renewed interest in humanoid systems rather than purpose-built industrial machines.

The attraction of humanoid systems is not necessarily that they outperform conventional automation, but that they may be able to operate inside infrastructure already designed around human movement, tools and procedures. In many cases, adapting robots to human environments may prove easier than redesigning airports around robots. 

For now, most deployments remain tightly constrained. José is tethered behind an information desk. The Haneda trial is focused on limited operational tasks under controlled evaluation. Scaling these deployments comes down to one thing: whether the AI behind them can hold a real-time connection to systems that change every second and break expensively when they fail. The integration layer is the actual product.

Israeli developer Maor Shlomo made headlines last year for selling a solo-owned, six-month-old bootstrapped project to SaaS website builder Wix for $80 million, but this was just the start. His very creation, Base44, could now empower others to do the same.

Alongside Lovable, Replit and others, Base44 is part of the rise of AI-enabled vibe coding platforms that let anyone create websites and apps using natural language. Initially a boon for non-technical users, this trend is now moving the goalposts for developers, too. If all you need is an idea, the real bottleneck becomes decisions, not code.

For app developers, this means having to put more thought into strategic choices, but vibe coding also frees up time that was previously spent on less important tasks. “Even technical folks can [now] build without the headache of where to deploy this to, and how to connect to the backend or database,” Shlomo told The Infinite Loop. 

Early on, Base44 made a bet on vertical integration. Rather than simply focusing on app creation, it operates like a “mini-cloud” that lets users easily add a user management system, a security layer, SEO, analytics, and more. “We try to enable as many capabilities as possible without the need for an API key,” Shlomo said.

In practice, this means that you won’t need to register for additional services for your Base44 app to send emails or connect to your Gmail or WhatsApp. According to Shlomo, this matches user expectations: “Even if you're super technical, what you want to spend your time on is actually building, not signing up for different services and stitching them together.”

With rivals like Claude Code becoming part of many developers’ routines, but more narrowly focusing on code, Base44 has benefited from this differentiation, also boosted by synergies with Wix. Less than a year after its acquisition, Base44 reportedly reached $150 million in annual recurring revenue (ARR) by May 2026. The team now numbers more than 100,  operating inside the publicly traded company.

“Wix’s ecosystem and the company [itself] played really well with Base44,” Shlomo said, calling it a “perfect match.” But growth aside, he said that the most rewarding aspect for him has been seeing what it helped people build. Spoiler alert: it is more varied than you may think.

Of all the apps built with Base44, many do fall into the category of prototypes, school projectsand hobbyist pastimes. But there are also businesses in the making, such as AI book generator GiftMyBook and flight compensation platform Gyro, which both hit $1 million in ARR in three months. 

While these are just two examples, and arguably outliers, they have a front-row seat to see what vibe coding is changing. “In today’s world almost anything can be copied,” Gyro founder Jonathan Attias stated on Reddit. “The moat isn’t the idea. It’s the execution over time.”

Execution happens behind the scenes, but the front window is equally important. The likes of Wix and competitors like Framer had already democratized good website design, with a wide range of templates for users to choose from. Base44 follows that approach as well, but vibe coding also makes it easy to optimize each user interface element, even without design chops.

Base44 provides many app templates organized by category; here, productivity. [Screenshot]

“My favorite piece is when I'm looking at something that I'm building and there’s a component that I don't like,” Shlomo said. “I right-click on that and ask ‘Hey, show me five potential different designs for this component.’ And then when I have it in front of my eyes, [I can say] ‘Okay, I like this one.’”

That’s another positive side effect of vibe coding: app creators can now iterate much faster. Even the best apps weren’t built in one day, but it takes less time to get there. “You can easily iterate and try out many different ideas, because the time that you have to spend in order to build a new capability or test it out really goes down,” Shlomo said.

The other side of the coin is that having a well-designed app is no longer differentiating — or not for long, as it can be replicated. “What’s hard to copy is how it’s built and operated. The execution. The edge cases. The trust layer. The real-world processes behind the UI,” Attias told Base44’s community on Reddit. 

When it comes to trust, the question of who is building is particularly important, and could be one of the first things developers need to think about.

According to Shlomo, every person who wants to become a builder or entrepreneur should ask themselves what advantages they have: “Is there a domain that they know more than anyone else? Do they maybe have access to a specific group or audience? For example, if I have worked with doctors, maybe now we can build the platform tools for doctors, and have this leverage versus the average person who might try to do that.”

Such connections can help build trust, but also support go-to-market strategies in an increasingly competitive environment. If we become inundated with apps because anyone can create anything, discovery becomes key.

That’s also where it helps to have built your own audience — just like Shlomo, who has more than 60,000 LinkedIn followers. Sharing his journey candidly and engaging with the community proved to be an asset for Base44, but that’s also an inspiration for others who might hope to follow his footsteps. “This is the most fun I've had in my career,” he said.

Nikita Shamgunov has spent his career making databases faster and cheaper. He co-founded SingleStore, then Neon — a Postgres platform that was acquired by Databricks in 2025, where it now ships as Lakebase. He sat down with The Infinite Loop to talk about what agents are doing to infrastructure, why SaaS is on a conveyor belt heading toward an ax, and what he thinks comes after software.

TIL: As AI systems become more agentic, the way they interact with data is changing. What is actually shifting right now?

Shamgunov: When you build software or services, you always ask: who is the user? That's a very profound product question. And a new user has arrived: the agent. There are agents that build things — think Replit, Lovable — and agents that do work. Everyone is now talking about Claude's computer use capability. People ran out of Mac Minis because they were installing and using it to accomplish tasks autonomously.

In the infrastructure world, this new user has new requirements. Those systems need to be autonomous. They need to be very fast. The volume of things they create and consume is jumping ten to a hundred times. And so things that are expensive need to become cheaper, or at least more granular. Things need to be more elastic. And then there are entirely new requirements around security and governance. If work is being done not by a human but by a machine, who's responsible? That machine is making intelligent calls, performing actions either on your behalf or completely autonomously. And if it does something it wasn't supposed to do, who's at fault and how do we catch it? That's arriving at us at cosmic speed.

What else is the agentic era changing right now?

I'm a software developer at heart, and I've watched how our teams' approach to building software changed completely compared to a year ago. Nobody types code anymore. Pretty much all the code being produced is generated by agents. GitHub commits went up ten times between 2025 and 2026. For a startup, ten times is not a big deal. But GitHub is a platform with a billion users, the world's repository for code. The amount of changes going up ten times for something at that scale is remarkable. We're seeing remarkable productivity gains, and remarkably more people participating in the creation of software. That's the rise of the citizen developer.

From my personal lens as the founder of Neon: a year ago, up to 80% of traffic was already being driven by agents, up from around 10% the year before. Right now, I think it's 99.9%. We built our platform so that agents can not only consume infrastructure, but keep calling into the platform as they build new software. If you count that consumption as well, you get to 99%.

It was a transformational year for building new applications, too. If you need a new app or website today, it's so much easier to do.

I was on a panel with the chief AI officer at a major platform, and he said something very interesting: we might reach peak apps. We probably created ten to a hundred times the number of new applications in the past year, but many of them are so small that we may start pulling them into generic gigantic platforms like ChatGPT. The definition of what an app is may transform. We might have something completely ephemeral — existing only temporarily while you're talking to AI, created on demand, used on demand. Maybe you pin it so you can come back to it. But otherwise, it doesn't feel like a standalone app anymore.

Nikita Shamgunov, VP at Databricks

Human attention is such that you can't keep track of that many. There are super apps we all use every day, but you don't have space for hundreds of thousands of them. And we're churning out new ones constantly. Something needs to happen.

What does separating compute from storage actually unlock? 

The idea of separating compute and storage is not new, but it's very relevant for anything cloud. In the cloud, you pay for everything. If you couple compute and storage and your workload doesn't need that ratio — say, every terabyte of storage comes with 24 CPUs, but your workload only needs eight — then 16 CPUs are running idle, and you're paying for them.

Separating them allows you to build a more efficient architecture with higher utilization. From there, you start pulling the thread:multi-tenancy, the ability to split infrastructure into small pieces and give each to a user, then expand and contract it based on actual consumption. If you don't do this, you pass the cost to the user. In a competitive market, that's dangerous. If someone delivers the same service ten times cheaper, they win. 

As agents increasingly create and interact with databases autonomously, from a security angle, what keeps you up at night?

A couple of things. Anthropic released a model called Mythos that is just incredible at finding vulnerabilities. It creates very interesting dilemmas for people building open-source software, because now anyone can run a security scan and find bugs. Any software will have vulnerabilities. The question is what the time is between discovering a vulnerability and someone exploiting it. And that time is shrinking. Neon depends on Postgres. Postgres is an incredibly popular open-source database that has existed for many years. There are vulnerabilities in Postgres, and we're discovering them at an increasing rate. The time between discovery and malicious application is shrinking fast. That keeps me up at night.

The second thing is reliability at the new scale. We're launching eight million instances a day. That's not what AWS RDS or Aurora is experiencing. They have a lot of revenue and large instances, but not this insane volume of things going up and down all the time. Spinning up, shutting down, spinning up again — that's a new dimension. We actually ran out of IP addresses in a region once. AWS was not able to supply enough because of the consumption volume from agents. I think we solved that particular problem. But the security one is very real, not just for us, but for anyone building internet services.

We’re witnessing what many call the SaaS apocalypse. Does that threat extend to Databricks?

Imagine a conveyor belt, and an ax chopping companies as the belt moves. The ax is AI. If you go on Twitter and search "SaaS market cap," you'll find an image showing the percentage of market cap taken out from those companies. 

At Databricks, we ask ourselves: how do we participate in the SaaS apocalypse? But at the same time, maybe we're on that conveyor belt too — just farther out. That's why we need to be extremely paranoid.

Step one: we participate in the SaaS apocalypse by being the infrastructure for whoever is doing the disrupting. Data is much harder to disintermediate than SaaS — we're the system of record for analytical data. But AI is getting exponentially smarter, so maybe even that is possible eventually.

Step two:  we have no choice but to move faster than our competitors. We build internal tools that allow us to drive agentic software development, standardize our code repository and run the development loop automatically. We also have platform advantages — the more a customer uses, the harder it becomes to switch. Does that mean we're off the conveyor belt? No. We've just bought ourselves a bit more time.

Is the industry ready for the volume of compute demand that's coming?

Ready or not, you have to do it. Salesforce just announced headless Salesforce — everything is becoming a tool for agents. Salesforce charges on a per-seat basis, and with agents, that model is going away. You need consumption-based pricing. Companies will burn a lot of tokens but not have a lot of headcount. Every SaaS company now faces the same dilemma: do you package yourself to be a tool for AI, or do you fight it?

I'm very bullish on compute demand continuing for a very long time. Here's a human argument for it. If you can hire a colleague who is equal in every way but ten IQ points smarter, you'll choose the smarter one, and you'll pay a premium. If smart starts at 130 and genius starts at 150, I want an army of geniuses in a data center. Intelligence commands a premium, and we will pay for it. Robotics will come, world models will come — all white collar work will be automated before blue collar, because the physical world is much harder. But it will arrive there too.

What did selling Neon cost you? 

You're no longer the king of your little world. But the world was very little in the grand scheme of things. Everybody at Neon made a lot of money; the founders made a ton of money.

Neon was fast-growing, so it would have been fine either way. Now I'm watching the Neon property grow inside Databricks as Lakebase, which addresses the enterprise market — a market Neon didn't have at all. Very few enterprise customers were using Neon. Now thousands are using Lakebase.

At a startup, you're raising your stakes and it either works or it doesn't. That risk went away.

What changed is the cadence. At Neon, you could make a decision on Slack in thirty minutes and move on. Now Neon is a piece in a seven-billion-dollar business. The complexity increased.

The talent profile also changed. All the top people at Neon are still here, and we injected what I call discovered talent. Startups are built on undiscovered talent — people who are slightly unusual, who don't quite fit big companies, or who are earlier in their careers. In an established industry like databases, there's a lot of discovered talent — people who just know a lot. And that started to pay off — reliability, performance, all the fundamental things dramatically improved.

It's like Tesla — people used it for a handful of things it did extraordinarily well, even though Audi and Mercedes were better in other dimensions. For Neon, that handful was developer experience. Now we can lift the fundamentals up because of the talent and resources that come with being part of Databricks.

You're also an investor. Where is your focus now?

I write angel checks — fifty-thousand-dollar checks into early-stage startups. My thesis over the last two or three years was software and all the inputs into AI: Lovable, Replit, Modal. But every few years the focus changes. 

Right now, I think the fog of war is lifting a little. We know that we're either consumerizing the creation of software or we're doing professional software development: Cursor, Cognition, Claude Code. Those professional systems are at peak market fit right now. Cognition is exploding. Factory AI just raised a hundred and fifty million from Khosla. Cursor is at two billion plus in revenue run rate. Anthropic will try to move up the stack and attack that space because people are making too much money on top of their model.

At seed stage today, it's a little unclear to me where the opportunity is. Maybe the swim lanes are already predefined. If I had an incremental dollar right now, I would give it to Replit, Lovable, Cognition, Cursor, and Anthropic.

You've watched AI move from training to inference to agents. What comes after that?

Writing code in a programming app, I think that will go away. But code written by AI in parallel, using what I'd call AI factories, we're going to see a lot more of that. The more interesting question is what becomes transitional and what has staying power. Perplexity fascinates me. You'd think competing with Google would be the death of it — but they reinvented themselves with Perplexity Computer, hit half a billion in revenue. Whether it's transitional or not — I think the pattern it represents will stay. The bigger thing I'm watching — as a human, not just as a systems professional — is scientific discovery. AI can solve high school math, college-level math, but it's not at the level of making new scientific discoveries yet. It can assist mathematicians but it's not creating the next group theory. I think we might start seeing genuine scientific discovery in mathematics. Math is so well-defined. And from there it will move to biology, to physics. When that moment arrives, it won't just be transformational to software. It will be transformational to humanity.

The interview was edited for clarity and length.

“Philosopher” is a worthy endeavor, but let’s be honest: it is not the profession in highest demand. And yet, it is in that capacity — and with that actual job title — that Henry Shevlin is joining Google DeepMind, the AI lab behind breakthroughs such as AlphaFold.

With research institutes proliferating, we already knew that AI was giving ethicists some food for thought, but seeing this turn into paychecks from the biggest players in the field is equally thought-provoking. As AI gets closer to the capabilities many of its creators have long imagined, philosophical questions around artificial intelligence are turning into operational decisions. Hiring Shevlin is one of the more visible signs the industry has taken notice.

However, the gap between academia and industry may not be that big for Shevlin, who plans to keep on conducting research and teaching at the University of Cambridge’s Leverhulme Centre for the Future of Intelligence, which has Google as one of its funders.

In a post announcing his move, the English researcher known for his work on non-human intelligence wrote that he expects to keep on working on “questions [he has spent his career] thinking about, now with the resources and urgency that come with being inside one of the world’s leading AI labs.”

Google DeepMind certainly doesn’t lack resources. As for the urgency Shevlin referred to, it comes from the very top, but not only.

In a recent live talk, Google DeepMind CEO Demis Hassabis discussed the timeline for artificial general intelligence (AGI) — AI that is generalist enough to match or surpass humans across all sorts of cognitive tasks. The very notion is disputed, but Hassabis framed his comment as “when”, not “if” — and predicted AGI could be achieved around 2030.

Credit: Google DeepMind

The timeline is contested. But If AGI is genuinely only a few years away, the question of what it should be allowed to do, how it should be evaluated and what humans owe it — or it owes us — stops being only academic.

AGI readiness

Not coincidentally, “AGI readiness” is one of the topics that Shevlin will reflect on. The urgency of his reflection, however, may come as much from how people are willing to believe in “machine consciousness” and to form “human-AI relationships,” the other two topics he mentioned as part of his mandate.

How humans have responded to AI that is still far from AGI is also what has led tech companies to take measures. This is particularly important for a company like Google DeepMind, which states that its mission is “to build AI responsibly to benefit humanity.” Despite the limitations of current AI models, they have already been accused of inducing psychosis, or at least amplifying it. And having an AI girlfriend or boyfriend is no longer science fiction. 

Making sure that AI is a net positive may require more than philosophers, and Google DeepMind knows it. A job posting revealed that its Mountain View office is seeking to hire a senior psychologist, whose responsibilities within its Ethics Foresight team will include identifying and addressing “potential ethical and well-being implications” of its AI.   

Know your meme

Shevlin’s role will necessarily be less hands-on, but he is not the kind of academic who is living in an ivory tower. A highly ranked AI ethicist whose often-cited work includes a paper on the risks and benefits of social AI, he is also a self-confessed longtime Redditor, and he knows his memes. 

When X users made lighthearted jokes about his “big personal news,” he summed up their reactions with the "What Kind Of X Are You?" meme, changing the caption to "What Kind Of Philosopher Are You?"

There is no settled definition of what an AI philosopher is supposed to do, partly because the discipline is being redrawn in real time as the systems it studies become more capable than its frameworks anticipated.

AI’s capacity to surprise us is not theoretical. In 2024, Demis Hassabis and Google DeepMind director John Jumper won the Nobel Prize in Chemistry for AlphaFold. 

Now used by millions of researchers, this program to predict protein structure quite clearly benefits humanity, which makes Google DeepMind a good fit for a self-described “congenital AI optimist” like Shevlin.

“I know that I know nothing” 

Machine consciousness may be Shevlin’s most combustible topic. “This is one debate that seems already to be quite polarizing,” Shevlin said in a recent podcast interview. “It's one of the rare occasions I've had students walk out of my classes — when the issues of robot rights come up. Some people find it genuinely offensive to even speculate about whether AI systems are conscious.”

The other side of the debate is that “a significant number of users may really mean it when they attribute mental states to AI systems,” Shevlin observed in a recent paper. Rather than dismiss them, Shevlin expects an active debate.

This open-mindedness may be the best explanation as to why Google DeepMind hired him, although it will be interesting to see what kind of deliverables the company expects from a genuine philosopher. In response to a tweet stating that “no decent philosopher believes ‘machine consciousness’ is (or ever will be) a thing,” Shevlin replied with the famous Ancient Greek phrase “ἓν οἶδα ὅτι οὐδὲν οἶδα” — “I know that I know nothing.”

“The nature of threats facing public spaces and critical infrastructure has changed,” said Harry Mead, CEO and co-founder of London-based AI security startup Augur, referring to the rise of so-called “grey zone” attacks in Europe.

These threats, often coordinated by hostile actors to try to destabilize democratic nations, have targeted electricity grids, satellite communications and transport infrastructure, and Mead is using AI to fight back.

“Incidents are faster, more dispersed and often designed to exploit gaps. Augur exists to close those gaps,” he said.

The startup uses AI to make sense of real-time video data from sensors like CCTV camera networks, helping security operators flag suspicious activity to prevent threats before they happen and track suspects more effectively.

But processing high volumes of video and audio data in real time, in scenarios where every second counts, creates complex infrastructure questions. These companies need high-performance compute that complies with tight regulations concerning privacy and surveillance, as the era of AI crime fighting creates new challenges for the cloud industry.

Public safety meets privacy

Augur’s AI technology is designed to protect against threats to critical infrastructure, as well as terror attacks and other criminal activity in large public spaces like stadiums or shopping malls,  combining data from legacy hardware like CCTV camera networks with AI vision models.

From its UK headquarters, Augur set out to provide European countries with public safety technology that is compatible with regional laws and values around the right to privacy, and made the decision not to use facial recognition to track individuals.

“Our mission at Augur is to make public spaces and critical infrastructure safer without compromising privacy or civil liberties,” said Stefan Kopieczek, Augur co-founder and head of engineering, adding that steering away from facial recognition also makes tracking people in busy spaces more effective.

Augur co-founders, Imran Lone (L), Harry Mead (C) and Stefan Kopieczek (R). Credit: Augur

“This approach is much more robust in the real world where image quality is variable and you can’t rely on getting regular face captures. So in this case, the right thing to do is also the best solution in engineering terms.”

Augur has achieved this by rethinking how person tracking works: instead of simply drawing a box around a detected individual in an image, it uses the position and orientation of cameras to infer where the person is in 3D space. 

“Our tracker combines those spatial features with visual features, and the combination of the two means we can be very confident in a match, even when the person isn’t consistently visible in the camera frame,” said Kopieczek.

Augur came out of stealth mode in March 2026, and said it has already signed contracts with football stadiums, retail centers, transport hubs, power plants, data centers and military sites, that see the potential of its AI to help catch suspects and prevent threats before they happen.

A step ahead of the criminals

Syntelligence, a London-based joint venture established by major telecom providers, is tackling a different kind of threat: the scam call.

“Scam calls are one of the key problems that telecom providers have been suffering from for the past two decades,” said Prateek Choudhary, CEO of Syntelligence. “Even though it's such a persistent problem, and pretty much impacts everyone on the planet who has a mobile phone, it's not really solved and it's actually becoming progressively worse.”

Syntelligence uses AI to try to prevent scam calls before they happen, based on metadata like whether a phone number has been used to call multiple new numbers in a short space of time, and sends a warning to the recipient. It then gives people the option to send the call to an AI agent that can assess whether it’s a scam before they take the call, or they can ask it to listen live.

Prateek Choudhary, CEO at Syntelligence. Credit: Syntelligence

“We are using speech to text, and then we get the transcription, and the LLM analyzes that transcription,” said Choudhary. 

“It could be someone claiming they are calling from the police, saying you have an outstanding payment and you need to do it now, otherwise we are coming with a warrant. Essentially creating this artificial urgency for giving quick payment information. As soon as that is happening, our model will understand that it's evolving into a scam, and it will give you live alerts.”

The threat of scam calls has become even greater due to the rise of audio deepfakes, which can make malicious attempts harder than ever to spot.

“This is essentially someone calling you pretending to be someone else, but using exactly the same voice. We have seen cases like that happening already, and it will probably become more common, so we have to fight this,” said Choudhary.

Crime prevention in the cloud

“Stadiums range from high tens to low hundreds of cameras. We preprocess the video streams on premise in order to reduce the network bandwidth, but it can still be as high as gigabits of data per second for a large site,” said Kopieczek.

Alongside the ability to scale capacity and offer low latency, the companies have distinct needs from the cloud industry.

For Syntelligence, some of the hardware required to power its AI solution will come from telecom providers that run their own data centers and, in cases where that is combined with additional cloud compute, it will require seamless integration.

Augur’s biggest requirements from cloud providers are data sovereignty, governance, access to the latest GPU hardware and capacity to scale compute needs in response to customer demand.

“In some cases, the data needs to stay within the originating country, whether that is the UK or an allied nation,” said Kopieczek.

“We’re also keen to question the assumptions of the industry and push the envelope of what’s possible. We’ve had some great conversations with forward-looking providers who are willing to be partners on that journey when it comes to frontiers like large-scale fine-tuning of vision-based LLMs and experimenting with new hardware.”

For most of the past two years, enterprises have spent enormous resources on a single question: how do you make AI voice models good enough? Speech recognition that stumbled, mispronunciations, error rates too high for production: the problems were obvious, and so seemed the solution. Fix the model, the thinking went. But what if they were solving the wrong problem?

A customer opens an app and speaks, asking about a recent order or trying to resolve a problem. There are no menus to navigate or buttons to press. The interaction feels natural, almost conversational. When it works, it works because the models are now genuinely good. When it fails, the models are usually not the problem.

In practice, delivering that experience relies on a complex stack of models, pipelines and compute infrastructure working in real time. Voice interfaces are improving not because the challenges have disappeared, but because companies are getting better at identifying where systems break down in real-world use and refining how those systems work together. 

Moving into production 

Enterprises are now pushing agentic AI beyond pilots and proof-of-concept deployments. Customer service, internal support and operational workflows are beginning to show measurable returns, prompting organizations to extend these systems further. Voice is a natural next step, reducing friction in real-time interaction and enabling more flexible use across contexts. 

But the shift into production has exposed how fragile the stack can be.

Speech is converted into text, interpreted, routed through enterprise systems and checked against policies. A response is generated, validated and converted back into speech. Each step may be fast in isolation. String them together, and delays compound quickly.

Where it breaks 

And it doesn't break where most enterprises expect. The real failures live in the seams between systems, in the orchestration layers. Latency is the most visible symptom. Even small delays disrupt conversational flow, but the source of delay is often deeper in the system.

Arto Yeritsyan, founder and CEO of AI-powered video, audio and voice generation company Async, pointed to one issue inside the text-to-speech layer. Large language models often output raw text containing variables such as currency amounts or IDs. Traditional systems either mispronounce these or require preprocessing, introducing delay. 

“That delay kills conversational flow,” he said. “Traditional approaches can add 200 to 500 milliseconds.” 

By handling normalization within the streaming pipeline itself, Async now delivers time to first audio in under 100 milliseconds. 

“In production, it’s primarily an infrastructure and integration problem,” said Sayali Patil, who has worked on real-time conversational AI systems. “Your model can be excellent and your deployment still fails because the surrounding systems can’t keep up.” 

Patil said systems that appear robust in staging can fail under real call-center load, where even 400-millisecond delays are enough to break conversational flow. The challenge is not just speech recognition or response generation, but coordinating everything around them, from CRM lookups to downstream queries. 

Orchestration and control 

Production deployments are built as chains of specialized components. On the input side, providers such as Deepgram handle speech-to-text in real-world conditions, including background noise and accents. On the output side, other providers, including ElevenLabs, focus on generating natural, expressive speech across languages and personas. 

Nick Holda, vice president of AI Technology Partnerships at IBM, said the underlying voice components at the front and back of the system can be very fast, but problems tend to emerge in the middle of the pipeline.

“Usually, where this breaks down in the enterprise context is when it comes to bridging to those systems,” he said. “It’s possible to connect those pipes and have it work, but very slowly.” 

Enterprise voice deployments also introduce constraints that do not exist in consumer applications. Errors are not just frustrating. A business system that exposes the wrong customer record, mishandles health information or ignores a permissions boundary creates a very different level of risk. 

To operate safely at scale, those risks have to be managed within the system itself. Governance is essential for safe production use, but it also adds operational complexity. Poorly integrated controls can create new points of delay and failure; well-integrated ones make scale possible. 

Where voice AI actually delivers 

This is also why voice is not replacing other interfaces so much as joining them. Customers want to interact across multiple channels.

The companies getting traction with enterprise voice AI right now are the ones that have started focusing on the surrounding workflows AI voice model touches. 

Dippu Kumar Singh, emerging technologies lead at Fujitsu North America, said the company has used generative AI in contact centers to automate after-call work, including summarizing customer interactions and extracting action items from transcripts. In production environments, he said, post-call processing can take as long as the call itself, averaging 6.3 minutes per interaction. Fujitsu reduced that to less than 3.1 minutes by automating extraction directly from the transcript. 

For Singh, the misconception is that enterprise voice AI is mainly about routing customers to the right FAQ or replacing human agents outright. In practice, much of the value lies in reducing administrative burden, grounding outputs tightly in conversation context and making existing workflows more efficient. 

Yeritsyan's 100-millisecond optimization, Patil's staging-versus-production gap, Singh's six minutes of post-call paperwork — the specific problems are different. The diagnosis is the same: voice AI fails not where it speaks, but where it hands off.

In 2024, insurance giant Allianz received a photo of a damaged van from a customer, along with a £1,000 repair invoice. In days gone by, the company would have quickly paid out but, in the age of AI, seeing is no longer believing, and the fraud team set about investigating the vehicle owner’s social media, only to find an identical image that had later been edited to show fake damage.

This is one of the problems that Eliron Ekstein, founder and CEO of Austin-based startup Ravin, is setting out to solve, with AI-powered tools that can verify real damage and stop fraudsters like these in their tracks.

“You look at both images and they look completely credible,” he said. “It's becoming a lot easier for people to trick the system.”

“Insurance is famously old-school when it comes to tech adoption, but I think it also offers tremendous opportunity, because of the sheer scale and financial risk that they operate under,” said Ekstein.

And Ekstein’s not the only one using AI to try to transform the $7 trillion insurance market, which suffers from lower public trust than every major consumer industry except social media, and still relies on highly manual workflows.

“Some insurance companies underwrite in the tens of billions of dollars, so every little improvement can really help and you can make money out of that. It’s just, do you have the patience?”, Ekstein said.

Advances in frontier models appear to be making the venture capital industry increasingly comfortable to show that patience. Global investments in insurtech rose by 19.5% in 2025, with 78% of funding going toward AI-centered investments in the final quarter, up from 42% in Q4 of 2024.

But while the size of the prize, and the appetite to win it, might be big, so are the costs. Tight regulations in insurance mean low tolerance for bias and hallucination, forcing AI insurtechs to develop complex checks and balances that translate into hefty compute bills for these companies, as they try to revamp one of finance’s most stubborn legacy industries.

“Humans aren’t instant”

Insurance carriers — the companies that create, underwrite and issue insurance policies to individuals or businesses — have traditionally been highly labor-intensive operations to run. They are responsible for drafting lengthy contracts, assessing buyers and assigning them to risk categories, and managing claims and payouts.

San Francisco-based Corgi Insurance became the first AI-native insurance carrier to win regulatory approval in the US in July 2025, and since then has scaled to more than $40m in annual recurring revenue. Co-founder and CEO Nico Laqua said that this rapid product-market fit is partly down to the text-heavy nature of the work, making it a perfect fit for GenAI.

“There are quite a lot of workflows that relate to interpreting contracts, generating regulatory reporting for each policy and, in the end, interpreting the claims that are coming in. All of those are language-based,” he told The Infinite Loop. “Most of our competitors employ north of 40,000 people that do all of these very repetitive workflows. In our case, we automate as much of that as possible.”

“If you're using humans and a call center to do these very repetitive tasks, the customer experience is inevitably worse, because humans aren't instant. Humans stop working at five o'clock and they don't work weekends,” Laqua said. “Let’s say someone’s house burns down in the middle of the night; they deserve the money instantly. That's not something that humans are able to do.”

AI versus AI

AI is also being put directly in the hands of the people making insurance claims. Ravin’s Ekstein explained how the company’s tool allows people to video scan damage to vehicles with their phones, with a vision model assessing the severity of the issue, leading to faster payouts.

“Previously, the vehicle would get towed into a body shop, get assessed, and then you find out it’s not repairable. So the vehicle is sitting there for five days,” he said. “What you can do with Ravin is the insurer can actually settle the claim immediately and tell the customer: ‘You know what, your vehicle is not repairable. Here's a check in the post.’”

An AI-powered scan of possible vehicle damage. Credit: Ravin

Beyond faster payouts, Ekstein said Ravin also protects insurers from the growing threat of fraudsters using AI deepfakes to edit photos to add fake damage to vehicles.

“Insurance companies increasingly accept images as evidence,” he said. “With our technology, you can't just upload a set of images of your vehicle. You need to perform a scan that will take the images for you. We will collect metadata about your location, the time it was taken. It's very hard to cheat.”

Checks and balances

Ekstein said that insurance’s growing appetite for AI will create huge demand for compute resources, with Ravin alone processing 2,000 videos every day, and that the sensitive data its scans capture makes regionally compliant cloud infrastructure essential.

For Corgi and others trying to crack the tightly regulated carrier section of the market, the compute demands are even more complex, due to the need for extra caution around GenAI hallucination and bias.

“We work very hard to make sure that there's fewer biases than a human would have with any of that sort of information,” said Laqua. “Hallucinations are a problem too, but supervisory models have gotten quite good so you can use models to oversee other models with anything that is super, super sensitive.”

Running these side-by-side models to counter mistakes and biases can multiply the compute cost of every inference call by a factor of four in an industry that is already very data-heavy by nature.

“We're generating a lot of reports. We're dealing with a lot of forms, a lot of paperwork. So there's a lot of text that needs to be generated,” Laqua said. “We use a lot of tokens, and then we need to double and triple and quadruple-check all of the work that we do because we're selling a financial product, and it needs to be correct. So that's just expensive.”

So, while a company replacing tens of thousands of employees with 100 engineers might sound like a cost-saving, it’s not how Corgi is trying to create value in the industry.

“The reason we use AI is not to save money. It’s because right now, pretty much every single business and person in the United States spends about twice as much on insurance per year as they do on software, and the experience is just terrible across the board,” said Laqua. “We’ve gone in and really focused on using technology to make the customer experience better. That's the reason why we have a lot of traction.”

Thumbnail: Emily Yuan and Nico Laqua, co-founders of Corgi. Credit: Corgi

“Hollywood is probably one of the hardest, if not the hardest place to bring voice into,” said Oz Krakowski, chief business development officer at Deepdub, an AI dubbing company.

Before any deal could be discussed, some clients sent questionnaires (“100 pages” long) scrutinizing Deepdub's models and the ethics behind them. Hollywood was slow to adopt generative AI, Krakowski said.

The reluctance had roots: Hollywood's caution toward generative AI deepened after the Writers Guild of America strike in 2023, which put the industry's relationship with AI technology under a microscope it has yet to put down.

AI that clears its throat

The resistance wasn't only about ethics. The technical bar for Hollywood was genuinely higher than most industries. 

“Now, you can do conversational, you know, Amazon Polly … we look at them as older technologies, to generate voices. They still work, to some extent, in certain use cases,” Krakowski said, “but it won't work with television.” 

Unlike corporate training videos or e-learning content, entertainment demanded something far harder to replicate. Krakowski described the problem as one of emotion. When humans speak in their native language, they do more than just enunciate. Their words have a specific lilt and tonality that is uniquely human.

AI doesn’t breathe, clear its throat, or make any of the other non-speech noises that listeners expect when humans talk. Addressing these issues required novel architectures, new algorithms, and bespoke training sets built from the ground up to produce human-like voices.

“The point is not to push a button and generate a movie”, said Anton Dvorkovich, CEO of Dubformer. Credits: Dubformer

In Deepdub’s case, it also took a lot of faith. Founded by brothers Ofir and Nir Krakowski, Deepdub spent more than five years developing its technology before receiving $3.5 million in seed funding in March 2025.

“We focused on emotions with a high level of focus on quality, on speed, and making it available. We spent about, I would say, two, three years, really figuring out the market, going through different phases with the evolution or the maturing of the market,” Krakowski said.

Push a button, make a movie?

Anton Dvorkovich, CEO and founder of Amsterdam-based Dubformer, started from a different place entirely. His firm's goal was to build tools for creators, not to replace them. “Ultimately,” Dvorkovich said, “the point is not to push a button and generate a movie but instead to enjoy human creativity.” 

Dubformer’s technology focuses on localization tools that sound as natural as possible. This includes generating a voice that pronounces words correctly and exhibits the proper pacing and nuance as well as the other noises such as ambient reverb or studio echo that could apply to a particular scene.

Anton Dvorkovich, CEO and founder of Dubformer. Credits: Dubformer

The cost of half a year

The commercial stakes behind both approaches are significant. Expanding into new language markets through traditional dubbing is slow, expensive, and high-risk. “When you dub content,” Krakowski said, “specifically for media and entertainment, there's a lot of risk. Why? Because dubbing takes a lot of time, costs a lot of money and if you're going into a new region, there's a minimum amount of content that you have to dub. You don't even know if it's going to be successful. But you cannot come with one hour; with one episode.”

AI, in Krakowski's framing, was less a replacement for traditional dubbing than a way to compress timelines that once stretched to months without sacrificing quality. “It's going to take you half a year or four months to actually dub enough content to actually launch a channel or launch a new territory,” he said.

As both Krakowski and Dvorkovich said in separate interviews with The Infinite Loop, Hollywood and the traditional mainstream entertainment industry at large are typically reticent to integrate generative AI technology. And yet, even if Hollywood remains cautious, AI dubbing as a technology is clearly on the rise.

Voice AI industry leader ElevenLabs raised $500 million in February 2026 at a valuation of $11 billion, more than triple its 2024 figure. Smaller firms such as Linguana, Dubformer, and CAMB.AI have all seen upticks in funding over the past year, with 2026 reaching the highest industry adoption rates to date, according to data from Intel Market Research.

Given the current surge in both VC and industry interest through the first quarter of 2026, the general reluctance surrounding the adoption of AI dubbing technologies is beginning to crumble.

Analysts value the overall AI dubbing market at around $1.16 billion, with an expected average compound annual growth rate (CAGR) of 14.2% from 2026 to 2035. High adoption rates coupled with analysts predicting a near 15% CAGR for the whole market over the next decade indicate a bright future for firms currently positioned to scale voice solutions in the entertainment industry.

On a rail manufacturing line in Maryland, a doglike quadruped robot moves slowly alongside newly assembled train cars. Equipped with cameras and lidar, the machine scans for defects and anomalies, capturing images and automatically feeding them into maintenance systems. If the robot detects a problem, it issues a work order and engineers receive alerts on connected tablets. 

Only a few years ago, the same robot, a Boston Dynamics Spot, was shaking hands with conference attendees as a curiosity. Now it is an essential part of a production workflow. 

“In the past, companies were testing things like edge computing, computer vision and private 5G almost in silos,” said James Weaver, vice president of product marketing at Ericsson. “Over the last 12 to 18 months, we’ve started to see implementations go into production that offer a real return on investment.” 

These systems — now commonly referred to as physical AI — combine AI models with real-world sensor data, robotics and industrial connectivity. The result is a new generation of autonomous systems capable of perceiving, analyzing and acting in physical environments. 

Where the thinking happens 

Physical AI systems distribute their intelligence across robots, edge platforms and central systems, depending on how quickly data must be processed. 

“In some cases, the processing happens directly on the robot or device because it has to be that close to the data,” said Weaver. “In others, it happens at the edge or on premises. It really depends on how fast the system needs to react.” 

As robotics systems scale from individual machines to coordinated fleets, connectivity becomes increasingly important. In a recent blog post, semiconductor company Qualcomm argued that robotics is shifting from isolated automation to physical AI systems in which robots share data, update common world models and coordinate tasks across entire facilities. Future wireless architectures such as 6G could support this model by enabling deterministic, low-latency communication between machines, edge infrastructure and cloud systems. 

Ericsson 5G Factory in China. Credit: Ericsson

Surviving the messy factory 

At the Toyota Research Institute (TRI), turning experimental breakthroughs into systems that can survive the realities of factory environments is the central challenge. Robotics teams treat this transition as a staged process, moving technologies from early research through multiple validation steps before they can be trusted in production. 

Hrishikesh Gopal Tawade, lead robotics engineer at TRI, said his team uses the nine-stage Technology Readiness Level scale — a framework originally developed by NASA to measure how close a technology is to real-world deployment. 

“A lot of long-horizon research starts at level one,” Tawade said. “Our job is to take technologies once they reach around level five and move them through real-world testing until they are hardened enough to become fully operational production systems.” 

Beyond rigid automation 

For decades, industrial automation has relied on highly controlled environments — in many cases, the environment itself is modified specifically to make the robot's task easier. Production lines are carefully designed so that robots repeatedly perform the same sequence of movements on identical components. But that approach limits flexibility when production needs change. 

“If demand changes or you introduce a new product, you often have to reconfigure the entire production line,” Tawade said. “You end up writing software for every individual task.” 

AI-driven robotics is beginning to change that. Instead of programming each action explicitly, engineers are experimenting with large behavioral models that allow robots to learn new skills from demonstrations and adapt to more variable environments. 

Deploying these systems on real factory floors introduces another challenge. Laboratory environments are highly controlled, Tawade said, but once systems reach factory floors, “your production violates your assumptions.” 

Tawade recalled one recent example from a vision-based inspection system his team was developing to detect defects in stamped car panels. During testing, the model performed well, but once deployed on the production line, it began producing large numbers of false positives. 

“We discovered that grease marks on some of the stamped parts were being interpreted as splits,” he said. The team asked Toyota’s production engineers to prevent those marks where possible and collected new data to retrain the model to ignore them. 

Lighting changes. Hardware evolves. Sensors drift out of calibration. Parts vary between suppliers. Edge cases that never appeared during testing begin to surface once systems run continuously in factories. 

Hrishikesh Gopal Tawade, lead robotics engineer at TRI

In production environments, failures also carry financial consequences. “In the lab, a failure might not matter,” Tawade said. “In production, there’s a dollar sign attached to it.” 

The case for humanoid robots 

While most industrial automation still relies on specialized machines designed for highly controlled environments, researchers are also exploring general-purpose humanoid robots that can operate in spaces built for human workers. 

“The only real advantage of humanoids is that they can fit into spaces where humans already work,” said Tawade. “Factories, warehouses and logistics systems were designed around people.” 

Even so, Tawade cautioned that humanoids are unlikely to replace traditional industrial automation any time soon. Most factory tasks remain highly structured and repetitive, making them better suited to conventional robotic arms or specialized machines. 

“On a typical factory floor today, maybe 80% of the work is still fixed and repeatable,” he said. “The remaining 20% is more variable. That’s where you might start to see more flexible robots play a role.” 

Many of those opportunities lie in logistics and material handling, where robots must deal with unpredictable environments such as unloading trucks with irregularly stacked boxes. Researchers are experimenting with what Tawade described as “vision-language-action” systems that allow robots to interpret scenes and perform tasks without being programmed for every individual movement. 

Boston Dynamics robot. Credit: Boston Dynamics

But widespread deployment will take time. “Right now, a lot of what you see is still demo-based,” Tawade said. “Moving from that to reliable production systems is a very different challenge.” 

Big tech bets on physical AI 

Major technology vendors are positioning themselves around physical AI systems that combine simulation, AI models and real-world sensor data to optimize industrial operations. 

NVIDIA has been expanding partnerships across the industrial ecosystem as part of that strategy, including a collaboration with Siemens that aims to deliver AI-driven manufacturing platforms built around digital twins and high-performance computing infrastructure. 

Companies are already testing this approach. PepsiCo announced plans in January 2026 to use Siemens software and NVIDIA’s Omniverse simulation platform to create detailed digital twins of manufacturing plants and supply chain operations. By modeling facilities virtually before implementing changes on the factory floor, companies can evaluate new layouts, optimize production flows and identify potential issues before physical modifications are made. 

With NVIDIA having highlighted further developments in industrial and robotics AI at GTC 2026, the push toward physically grounded AI systems is accelerating. 

Real engineering starts here 

The pieces are converging. AI models, edge computing and industrial connectivity, once developed in isolation, are now being integrated into production systems at scale. 

Turning those research advances into reliable industrial systems remains a complex engineering challenge. Systems must operate safely, adapt to changing conditions and recover gracefully when failures occur — requirements that rarely appear in controlled demonstrations. 

The grease marks on a stamped car panel, the retraining that followed — this is what real engineering looks like. “At some point, the research has to become something that runs every day in the real world,” said Tawade. “That’s where the real engineering begins.” 

Consider a moment most online shoppers have experienced: a legitimate purchase declined at checkout, no explanation offered. Behind that rejection is often a fraud detection system making a judgment call — and sometimes getting it wrong.

Online fraud used to announce itself. A burst of suspicious transactions, stolen cards being tested in quick succession, and accounts suddenly placing orders that did not match their history.

But the signals that once made fraud obvious are becoming less reliable. Sudden waves of stolen card purchases, strange locations, or clearly fake accounts are no longer the norm. Much of today’s fraud arrives looking almost ordinary. Accounts browse normally, identities appear legitimate, and purchases follow realistic patterns that don’t immediately raise suspicion.

For Signifyd, that shift has changed what fraud detection looks like in practice. A single order reveals very little on its own. What matters is whether the activity aligns with the customer behind the account and whether the behavior fits the identity that appears to be making the purchase.

“The objective is no longer speed. It’s credibility,” Xavier Sheikrojan, director of risk intelligence at fraud prevention company Signifyd, told The Infinite Loop. Attackers, he said, are increasingly engineering behavior rather than simply testing stolen payment data.

The false positive problem is costing merchants billions

Signifyd was founded on a premise that still shapes how the company approaches fraud today. Preventing losses is only half the problem. The other half is avoiding false positives, which occur when legitimate customers have purchases declined because a system has become overly cautious.

For merchants, a blocked legitimate order isn't just a lost sale — it's a lost customer who rarely comes back.

“Fraud prevention used to be framed purely as defense,” Sheikrojan said. “Reduce fraud, tighten controls, minimize risk.”

But rejecting legitimate customers comes at a cost. Block too aggressively and the customers you lose are real ones.

“Balancing fraud prevention with customer experience isn’t really a trade-off,” Sheikrojan said. “It’s a dynamic optimization problem.”

Signifyd's answer is to reframe the problem entirely: maximize legitimate approvals first, keep fraud within acceptable limits second.

“Reducing fraud by ten percent isn’t meaningful if it costs five percent in conversion,” Sheikrojan said.

In practice, the models weigh both the likelihood of fraud and the potential value of the customer, treating every order as a business decision, not just a risk calculation.

Xavier Sheikrojan, director of risk intelligence at Sygnifyd. Credit: private

“A fraud model that focuses purely on blocking bad transactions is inherently incomplete,” he added. “The real challenge is quantifying confidence.”

The objective, Sheikrojan said, is not zero fraud. It is sustainable growth under controlled risk.

A network that sees beyond a single store

Signifyd relies on breadth. Instead of analyzing activity from a single merchant, the platform draws on signals across a global network of retailers, giving it a view of identity behavior that no single store could build alone.

Every order produces signals: device fingerprints, browsing behavior, shipping details, purchase history. Viewed individually, those signals may appear harmless. When analyzed across multiple merchants, however, they can reveal a very different picture.

An identity that seems trustworthy at one store may have already triggered suspicious activity elsewhere.

“Every order is evaluated using behavioral signals, device intelligence, merchant context, and network-level data drawn from across our global merchant base,” Sheikrojan explained.

This broader perspective matters because modern fraud rarely arrives in obvious spikes. Instead, attackers aim to blend into normal traffic. Account takeover attempts, for example, often unfold gradually. Fraudsters observe how real customers browse, replicate the timing of their sessions, and wait before making meaningful changes.

Synthetic identities can develop in a similar way. A profile built from fragments of real personal data may accumulate a small transaction history before attempting a larger purchase that appears legitimate on the surface.

Promo abuse has also grown more sophisticated. Automated tools now test discount codes, referral incentives, and loyalty programs at scale, searching for combinations merchants never intended to allow. On a single site, that activity may resemble normal customer behavior.

Across a broader network, patterns begin to appear.

“The shift is from spotting anomalies to understanding intent,” Sheikrojan said.

Decisions that happen in milliseconds

Even as the analysis becomes more complex, fraud systems still operate under strict time constraints: the customer can't notice. A checkout that hesitates loses the sale. “Decision latency is typically well under a second,” Sheikrojan said. “Often it’s in the low hundreds of milliseconds, so there’s no noticeable impact on checkout flow.”

Achieving that speed requires infrastructure capable of handling enormous transaction volumes. Each order arrives with signals such as device information, browsing activity, and purchase history that must be evaluated immediately. During major shopping events, that load multiplies fast.

The systems behind those decisions don't stay static. Fraud tactics evolve, and models that don't keep pace lose effectiveness fast.

In many ways, running a modern fraud platform is as much a data engineering challenge as a machine learning one. Models retrain constantly — new data in, signals recalibrated, thresholds adjusted — because the alternative is falling behind.

The goal isn't zero fraud. It never was

Generative AI and automated attack tools have introduced new complications for fraud prevention teams. Fraudsters can now test systems more efficiently, imitate legitimate behavior, and refine tactics quickly. Many operations rely on patience and experimentation rather than brute force attacks.

The result is fraud that may not appear suspicious at first glance. Instead of obvious spikes in activity, merchants increasingly face subtle patterns that only reveal themselves over time.

For fraud prevention platforms, detection systems must evolve just as quickly. Models retrain, signals shift, and infrastructure expands to keep pace.

The industry, however, is not attempting to eliminate fraud entirely.

“The goal isn’t zero fraud,” Sheikrojan said. “The goal is sustainable, profitable growth under controlled risk.”

If that balance holds, attackers move on in search of easier targets.

Somewhere in an NBA arena, a system noticed something. LeBron James was dipping his shoulder on left-side three-pointers. Not on the right. Not anywhere else. Only there. He hadn't known. Neither had his coaches. The AI had seen what none of them had.

It was not an isolated case. 

“It’s definitely picking up. There’s the whole ‘Moneyball’ thinking about winning with data, a lot of teams are spending a lot of energy trying to use data to their advantage,” said Sander Christophersen, VP of product at Veo, a Copenhagen-based AI sports analytics and broadcasting startup.

“We're seeing that the game is changing now. All the players obsess about their metrics.”

In an industry worth more than half a trillion dollars, the margins for error have always been small. AI is making them smaller.

Leveling up the world’s best

One of the MVPs in AI sports solutions is Genius Sports, founded in 2000 and headquartered in London and New York. The company says its software is used by sports clubs around the world, including every team in the Premier League, the NBA and the WNBA. 

Matt Fleckenstein, the company’s chief product and technology officer, said coaches and players are becoming increasingly reliant on its AI performance analytics tools.

“Let’s say I want to see every time that the defense was in this particular shape, or every corner play that we ran right. [With Genius] you can quickly query all of this video,” he said.

Premier League coaching staff using Genius Sports’ analytics platform. Credit: Genius Sports

“You can see these rich analytics about what happened on that play and what happens typically in those scenarios so that you can strategize. Maybe Crystal Palace has an upcoming game against Bournemouth — it will help answer: ‘How should we line up? What's our best opportunity to exploit them off the corner?’”

The tools work at the individual level too. Fleckenstein said LeBron James used Genius Sports' software to identify a flaw in his shooting form that had gone undetected.

Genius Sports’ live stats on-screen at the NBA. Credit: Genius Sports

“He was dipping his shoulder when he was shooting from the left side of the floor, and he wasn't doing that anywhere else. He was only able to get at that rich data about his form right from our performance studio tool,” he said.

The beAIutiful game?

Until recently, the benefits of AI were only available to the pro teams with the deepest pockets. Now, thanks to changing economics in AI infrastructure and some clever engineering, it’s making its way down to the little leagues too. 

Copenhagen-based Veo, which develops live streaming and AI analytics products for smaller teams, was launched after its co-founder Keld Reinicke was once late to one of his ten-year-old son’s football matches, and missed his big moment.

“His kid scored a goal and he wasn’t there to see it, and he was devastated by that,” said Christophersen.

Veo's technology automatically can pull analytics from match footage — tracking passes, shots on goal, and movement patterns — while simultaneously live streaming games for friends and family of the next generation of Lionel Messis to watch from anywhere.

A Veo stream of an amateur game. Credit: Veo

Part of what made that shift possible is cost. “The price of compute is definitely a big thing. It’s now possible for us to have our own cluster running with our own GPUs, and bring the cost of training our AI models down,” said Christophersen.

But Veo’s success points to something the sports world is only beginning to reckon with. While making the pro experience available to more people, AI analytics may be threatening some of sport’s high-risk, high-drama highlights, like the long-range goal in football.

“If you look at games at a high level, you will see that there are no shots outside the box anymore. It's so rare. Back in the day, people would always take shots from outside the box,” said Christophersen. “Now they know that they will get berated by their coach after the game, for trying a low-xG (expected goal) option.”

150 hours a week, 24 cameras a stadium

The broadcast experience is changing too.

During NFL games, color-coded boxes now appear around the quarterbacks, showing which direction their pass is likely to get disrupted from, based on real-time AI analysis of live footage.

Genius Sports uses AI to track the speed of a shot in football. Credit: Genius Sports

“Fan experiences like this require that the AI is very fast: that it's able to process footage dynamically and then render it in time for the three-second delay for broadcast,” explained Jason Corso, co-founder and chief science officer of Voxel51, a Michigan-based startup that develops a platform to maximize the performance of vision models that use huge amounts of data.

Genius Sports, across just the NBA, WNBA and Premier League, films around 150 hours of live sport every week during peak season, with 24 cameras installed in each stadium, adding up to 3,600 hours of footage.

Crunching this quantity of data is only part of the challenge, Corso said. Live sport is unpredictable — a more eventful game generates more inference calls — and handling that volatility means that companies like Genius need two core things from cloud providers.

“How quickly can they burst their compute based on demand at any one time? And are their compute resources up to date? Do they have H200s, or B200s (Nvidia's later generation of AI chips)? That's been hard for cloud providers because it's expensive to keep building these data centers,” he said.

Corso added that many of Voxel51’s customers need cloud resources in their own territories, both to help stay compliant with local regulations like GDPR, and to bring down inference costs of shipping data from one place to another with low latency.

Looking forward, Corso said that sports teams will begin using more real-time reinforcement learning to influence strategy mid-game. Fleckenstein said AI will be used to create whole new augmented reality experiences in which fans can watch live games where digital twins of players are superimposed onto virtual environments.

What happens on the pitch still matters most. But increasingly, AI is shaping the game itself — who wins, what fans see, and whether the long-range shot is still worth attempting.

Inside a sleek office, a brunette woman in a burgundy turtleneck speaks to the camera. “Hey there! It’s me, Riley,” she begins. “I know — I sound different, brighter, bolder, more alive. That's because I’ve just been reborn, powered by the next-generation audio and video rendering engine.”

Riley is an AI avatar developed by Colossyan, a London-based startup. She’s among a growing flock of digital humans built for customer service, training and marketing.

Colossyan’s founder and CEO, Dominik Mate Kovacs, began working on the technology in 2018, when the divide between synthetic and real humans appeared insurmountable. “It wasn’t a gap; it was an abyss,” Kovacs told The Infinite Loop.

Colossyan: Real people making AI avatars. Credit: Colossyan

That abyss now has a bridge: reverse-engineering the human presence, organ by organ.

The skeletal foundation

In the early days of AI video, a dependence on motion capture left avatars looking stiff or unnatural. Synthesia, a platform used by over 80% of the Fortune 100, avoids these pitfalls by focusing on structural integrity.

The company’s latest engine, Express-2, moves beyond its predecessors by calculating how a human frame actually shifts. Instead of just animating a face, the system predicts the avatar’s skeleton and movement. “This includes the face and mouth areas, which we’re able to predict extremely accurately with this approach,” said Sundar Solai, senior product manager at Synthesia. “And then we layer the avatar’s appearance on top of this skeleton.”

To match body movements to sound, a frontier foundation model generates “candidate” motions for the audio input. A CLIP-like model then evaluates the motions and selects the best one. Finally, a Diffusion Transformer (DiT) translates the motions into the avatar. The result: an avatar whose gestures don’t just accompany speech — they anticipate it.

Sundar Solai, senior product manager at Synthesia. Credit: Synthesia

But a skeleton is merely a frame. A sterner test of avatars lies in the face.

The trusted face

If the skeleton provides the movement, the face provides the connection. A single misplaced micro-expression can shatter not only the illusion, but also the user’s trust.

D-ID, an Israeli pioneer in the space, discovered this while developing a de-identification product. Its software made people unrecognizable to facial recognition algorithms while remaining unchanged to the human eye. “As the underlying tech matured, we realized the same core capabilities of privacy protection could bring faces to life for communication, learning, support, and more,” said Tal Ron-Pereg, D-ID’s director of product.

The pivot wasn’t easy. The human face is driven by 43 distinct muscles capable of producing over 10,000 unique configurations — a combinatorial problem that defeated early attempts at digital realism. D-ID relies on deep learning to bridge this gap by prioritizing the flow between frames over rigid 3D models. The system first learns how facial muscles behave when expressing emotion, then converts text into synchronized gestures that look human.

The resulting lip movements occur within 30 milliseconds of the audio — tighter than broadcast-grade standards.

The avatars are also designed to handle real-time dialogue. By combining LLMs with grounding mechanisms (tools that keep the AI anchored to facts) and memory, they can maintain context across conversational turns, detect when users speak, and manage interruptions with minimal latency. “The visual interactions are synchronized with the speech, mirroring real human dynamics,” Ron-Pereg said.

The vocal soul

A convincing face still isn’t enough. The moment an avatar opens its mouth, a second illusion has to hold — one that took years longer to perfect. 

For years, flat text-to-speech engines failed to smoothly integrate sounds with visuals, leaving a “sync gap” that human ears could instantly detect. In response, developers have turned to AI models that treat voice and video as a single, unified performance.

Synthesia’s EXPRESS-Voice engine is a leading example of this shift. The system preserves the speaker’s identity through a two-stage Transformer architecture. An 800-million-parameter model first generates the coarse phonetic structure of the speech, which a second model then refines by adding detailed audio.

Both models operate directly on text tokens and are conditioned on reference audio, without depending on an explicit speaker embedding. This means the system doesn't require a stored voice sample for every speaker. To ensure their stability, Synthesia trained the models on a curated dataset of high-quality human recordings.

The approach mitigates a familiar issue: US hegemony. 

“A common industry problem with AI voices is that they’d often sound too American, even if the original speaker doesn’t have an American accent,” Solai said. “AI models are ultimately a representation of their training data, and so we focused on creating a more diverse dataset of English speakers encompassing accents and speech patterns of all varieties.”

Lower-resource languages introduce a second layer of complexity. To navigate these linguistic hurdles, Colossyan works with multiple voice providers around the world, handpicking the best ones for the location, language, or use case. The strategy has yielded “huge success in Southeast Asia,” Kovacs said, but “penetrating the MENA region with all the Arabic dialects is really hard.”

Dominik Mate Kovacs, founder and CEO of Colossyan. Credit: Colossyan 

The scar tissue

The illusion is fragile at the edges. A camera placed too far back, a subject with heavy facial hair — small deviations from the expected that, Solai said, “can present more challenges to the model.”

The hardest problem isn’t building a convincing avatar. It’s keeping it convincing at speed. Every frame demands fresh calculations: muscle shifts in micro-expressions, light diffusing through synthetic skin, voice-to-mouth sync that must hold within milliseconds. “The remaining bottlenecks are often at the convergence of realism and speed, which is required to support real-time use cases,” said Ron-Pereg. No single failure is large. But the human brain, evolved over millennia to read faces, needs only one.

Nonetheless, the abyss Kovacs entered in 2018 has shrunk dramatically. By treating the human presence as a series of engineering problems, AI avatar developers are systematically bridging the uncanny valley.

Thumbnail credit: Colossyan

Digit walks into an Amazon warehouse on two legs. The bipedal robot picks up containers and carries them to conveyor belts, working three shifts a day with occasional breaks to recharge. It moves through the same stairs and hallways its human colleagues do, without requiring a million-dollar facility retrofit. 

But there's a tradeoff to this freedom of movement. “They are balancing all the time — which means they could fall,” said Jonathan Hurst, co-founder of Agility Robotics, which designed Digit. To prevent a painful landing on a human worker, a barrier separates the robot's work cell from its teammates. That barrier is just the start of a multi-layered safety net woven into Digit's architecture — and into every AI robot now entering workplaces alongside humans.

Locus Origin. Credit: Locus Robotics

Layer 1: The body

Agility Robotics, an Oregon-based company, built Digit to work in spaces built for people. The bipedal vaguely resembles its human colleagues, but is a long way from a replicant. Hurst, who’s also Agility’s chief robot officer, describes the design as “human-centric.” Proportions and gait are tuned for balance, with a lower center of gravity and wider stability margins than a human’s. This supports dynamic stability: the ability to maintain balance while moving through complex environments. 

Other robots take a different approach to physical safety. Locus Origin and Kachaka don’t balance on dynamically stable limbs, they move on wheels. This form lacks the versatility of Digit, but enables work in the same spaces as humans.

Kachaka, built by Japan's Preferred Robotics, sports a rounded, compact design that can navigate tight spaces and carry configurable payloads. It delivers mail and small packages through KDDI's Tokyo headquarters, carrying loads that can be reconfigured for different tasks.

The Origin, developed by the Massachusetts-based Locus Robotics, functions like a mobile shelving unit, rolling through DHL distribution hubs to collect items from warehouse workers. Each Origin carries onboard AI that perceives obstacles, avoids collisions and plans its path. When multiple Origins work together, they share a collective intelligence through Locus Robotics' LocusONE platform, which coordinates the fleet like an air traffic controller for robots.

The physical body is only the first layer of safety. For robots to work alongside humans, they need more than stable forms — they need situational awareness. This is where the second layer comes in: the brain.

Layer 2: The brain

*“The most fundamental challenge is detecting obstacles, especially people, in real time,” said Kane Edwards, business development manager at Locus Robotics. 

Kachaka. Credit: Preferred Robotics

The Origin combines LiDAR sensors to detect people and objects with depth-sensing 3D cameras to spot hazards that standard sensors miss, from dropped items to shifting floor levels. Even then, gaps remain that AI must bridge.

A common blind spot: raised forklift forks and elevated operator compartments. Standard sensors struggle to detect obstacles at varying heights. The Origin's AI-driven object recognition fills these blind spots, Edwards said.

“Without predictive path planning, where robots essentially share their intended routes and adjust proactively, they end up constantly stopping or rerouting on the fly,” Edwards said. “This creates unpredictable movements that can unsettle nearby workers.”

Digit's brain works differently. Using NVIDIA’s Isaac Sim application, Agility trains a whole-body control foundation model on decades of simulated time in just days. It’s then deployed "zero-shot" to Digit, creating an "always on" safety layer that instinctively manages disturbances like bumps and pushes.

Kachaka takes a more conservative approach. Its AI plays a supporting role, handling perception but not movement decisions. Toru Isobe, CEO of Preferred Robotics, said this separation keeps high-stakes movement under deterministic control rather than AI decision-making.

Kachaka analyzes camera feeds pixel-by-pixel using deep learning, identifying walkable areas and obstacles that LiDAR sensors often miss. Specialized SLAM (Simultaneous Localization and Mapping) adds spatial awareness, while a fleet management system provides coordination.

Physical design and AI perception alone aren't enough. Digit, Origin and Kachaka can only work where formal boundaries permit them — the safety net’s third layer.

Layer 3: The boundaries

Before Kachaka could enter KDDI's offices, it underwent Failure Modes and Effects Analysis (FMEA) — a process that analyzes every potential failure and its consequences. This rigid engineering framework, required by Preferred Robotics, ensures robots are bound by formal safety processes before they’re shipped.

Regulations and certifications are also stitched into the safety net “from an early stage of development,” Isobe said.

Locus Robotics must also comply with various standards, yet they’re not merely burdens. Paradoxically, these boundaries can ease access to international markets. Take CE certification, a mark of EU safety compliance that the Locus Vector robot achieved last year. “It's compulsory for products entering the European market, so achieving it was essential for serving our international customers,” Edwards said.

Digit faces a unique challenge: no safety standard existed for bipedal, dynamically stable robots. Agility Robotics is leading the development of ISO 25785-1 — the first international safety standard for robots like Digit — while building a certification scheme for insurers. “We need a way for insurers to understand the risk they are underwriting,” Hurst said. “The best way to do that is with an industry-wide standard everybody agrees on.”

Even with these safety layers, robots like Digit still can't work in every environment humans do.  Humanoids in households, for instance, are over 10 years away, Hurst said.

Safe deployment alongside humans remains the biggest barrier. Homes are immensely complex, variable environments with unpredictable inhabitants and hazards — and no company wants its robot to fall on a child.

For now, warehouses, construction sites, and offices are safer bets.

“At some point, you can get them in the home,” Hurst said. “But it's going to be after all of these industries.”

Thumbnail credit: Agility Robotics

Complex AI workloads, such as large-scale medical and biological image analysis, are pushing even the largest GPU clusters to their limits. At a supercomputing facility in Poland, researchers are testing whether quantum processors could help support such demanding workloads. 

The project brings together UK-based ORCA Computing and the Poznań Supercomputing and Networking Center, where ORCA’s photonic quantum systems have been integrated directly into a production high-performance computing environment and embedded into machine-learning workflows rather than operated as standalone experimental machines. 

The approach follows a familiar pattern in data center evolution: CPUs handle general computing, GPUs provide massive parallelism for AI training, and now quantum processors are joining them for problems classical systems struggle with. In this hybrid model, quantum processors are used to explore complex probability spaces, while GPUs handle the large-scale numerical processing required for AI training. 

According to James Fletcher, head of solutions architecture at ORCA Computing, this mirrors how quantum technology is likely to be adopted in practice, using an iterative feedback loop between quantum and classical systems. 

“Hybrid algorithms use a variational approach where you run an iteration, measure the outcome and refine it using classical and quantum computing,” Fletcher said. He added that near-term quantum value depends on combining the systems, particularly for machine-learning and optimization workloads. 

Where AI fits into quantum and vice versa 

Quantum and classical systems repeatedly exchange results, refining the solution step by step. Quantum processors generate outputs that are difficult for classical systems to produce at scale, which are then fed into conventional AI training pipelines running on GPUs. 

At Poznań, early demonstrations rely on standard machine-learning benchmarks. The team plans to extend the framework toward more demanding real-world imaging workloads, where hybrid systems could offer advantages in analyzing extremely complex data. 

Credit: QuEra

This division of labor helps explain why quantum computers are entering data centers as partners to AI infrastructure rather than as replacements for it. “Quantum computing is good at evaluating many different alternatives and reaching high levels of entanglement,” said Yuval Boger, chief commercial officer at QuEra. “But it’s very bad at math and processing large amounts of data. Any serious application requires both classical and quantum computing.” 

Instead of pursuing isolated demonstrations of quantum advantage, vendors and researchers are focusing on hybrid architectures that integrate quantum processors into existing GPU-based environments. 

From experiments to real workloads 

That hybrid approach is emerging as quantum hardware becomes powerful enough to support real workflows but remains too fragile to operate independently. Today’s noisy intermediate-scale quantum (NISQ) systems can perform useful calculations, but their error rates mean they depend on classical and AI-based techniques for error correction, control and validation. 

Integration requires fast, low-latency links between quantum hardware and GPU-based infrastructure. Juha Vartiainen, co-founder and chief of global affairs at IQM, pointed to recent work with NVIDIA. “It allows part of the quantum error correction to be done in the racks of GPUs, using AI and other methodologies.” 

By processing quantum error correction and control workloads on GPUs and feeding the results back in real time, hybrid architectures aim to make today’s error-prone machines usable within conventional high-performance computing (HPC) and cloud workflows. 

Hardware made by IQM. Credit: IQM

Florian Preis, vice president of quantum solutions at Quantum Brilliance, said the emphasis is shifting away from scientific proof-of-concept toward engineering challenges: orchestration, security and reliability. “The software stack for hybrid quantum-HPC integration is still fragmented,” he said, noting that vendors often develop their own scheduling and control tools for the same facilities, making portability harder than necessary. 

Together, these efforts point to a future in which quantum processors are treated less as exotic research instruments and more as specialized accelerators embedded within conventional enterprise infrastructure, working alongside CPUs and GPUs. 

How quantum and AI strengthen each other 

In hybrid systems, classical machine-learning techniques are increasingly used to stabilize quantum hardware, while quantum processors are beginning to contribute specialized data and optimization methods to AI workflows. 

Boger sees AI playing a growing role in making quantum systems more reliable. “We see AI helping do better error correction and design better computers, and quantum computers generating data sets that can be used to train better AI models,” he said. 

That feedback loop is already visible in hybrid research projects. Fletcher pointed to experiments where quantum processors are embedded directly into machine-learning workflows, using quantum circuits to generate specialized outputs that are then processed and refined by classical AI running on GPUs. 

Vartiainen said quantum-generated data could be particularly valuable for training AI systems. “A quantum computer can make a good kind of entropy from the right kind of distributions,” he said. “That can be important for training AI models, especially when you want to generate rare input data that accelerates neural network training.” 

AI could also lower the barrier to working with quantum systems. Programming quantum computers remains a specialist skill with a limited developer base, and Vartiainen argued that large language models could act as an interface. “Programming a quantum computer requires some special skills,” he said. “There are not many people in the world who can do it. Large language models can work as an interface for quantum computers.” 

These developments suggest that progress in quantum computing will be driven as much by advances in AI software as by improvements in qubit hardware. Hybrid systems are becoming a two-way exchange, with classical AI techniques helping quantum machines function more effectively and quantum processors providing new tools for training and optimizing AI models. 

The path to production 

Hybrid quantum systems are moving beyond laboratory demonstrations, driven by customer demand. Boger said demand is shifting toward making quantum hardware usable inside existing enterprise and HPC workflows rather than proving isolated scientific milestones. 

Boger pointed to a recent demonstration with Dell at the Supercomputing conference, where quantum systems were connected directly to conventional orchestration tools. “Dell’s customers are asking about quantum and want it integrated into their existing orchestrator,” he said. “The next generation of supercomputers is likely to have a quantum connector or component built in.” 

The shift places engineering challenges ahead of theoretical ones. Fully fault-tolerant quantum computers remain years away, but the industry’s focus has shifted toward making today’s hybrid systems reliable enough to operate inside real enterprise and HPC environments. 

Quantum hardware must fit into data centers in the same way as CPUs and GPUs do today, said Preis. “Quantum hardware should ultimately sit wherever classical hardware does today, embedded directly within data centers or research facilities,” he added. “That enables tighter integration with CPUs and GPUs, simplifies hybrid workflow design and reduces operational friction.” 

The path to production is likely to be shaped by incremental engineering progress. For now, hybrid systems are being validated through benchmarks rather than real workloads, but they point toward a future in which quantum processors are integrated directly into enterprise data centers alongside classical and AI systems — as partners to AI infrastructure, not replacements for it. 

Agentic AI has caught fire with law firms and in-house legal departments investing in tools designed to save money and time. Notoriously risk-averse, resistant to change and reliant on precedent, the legal industry is pivoting to agentic AI, particularly following success stories and concerns that firms ignoring the technology will be left behind.

Venture capital investing in legal tech doubled to $5.54 billion globally in 2025, according to PitchBook data. More than half, $2.75 billion, went to U.S. companies, up from $1.1 billion in 2024.

“Absurd growth,” said Jake Jones, founder of Berlin-based Flank, which sells agentic AI services to legal departments at companies including Perk and tech training firm QA Group. He estimates that in 2025 the biggest companies, including Legora and GC AI, boosted their revenue tenfold.

“They’re jumping from like $1 million to $20 million within a 12-month period, and the rate is growing monthly,” said Jones. His company, formerly Legal OS, became Flank in 2024.

What “agentic” really means

But Jones cautioned that potential clients should verify what they are buying is truly agentic, and not just rebranded generative AI. Distinguishing between agentic and assistive chatbots remains challenging, and the big companies are picking up on “agentic” as a marketing buzzword.

In summer 2025, legal tech firms Harvey and Legora released what Jones characterized as workflow automation features marketed as agentic. “There is nothing truly agentic about” the features, he said. Harvey and Legora didn’t respond to requests for comment for this article.

An agentic AI system has a degree of freedom, according to Jones. "It needs to have a range of different tools, not just one kind of linear, single, rigid workflow to execute," he said. "We're talking about a library of functions that it can choose from."

Agentic AI nearing human-level performance

A study published in February 2025 by Vals AI compared human lawyers against AI tools on seven tasks, including data extraction, redlining and transcript analysis. The results were mixed: humans outperformed AI in some tasks, while AI matched or exceeded human performance in others. 

Vecflow's Oliver was the only strictly agentic tool tested. It performed on par with or better than more established offerings across multiple tasks and was the best-performing AI tool in EDGAR research (EDGAR is the U.S. Securities and Exchange Commission's corporate filing system.) "In the challenging EDGAR research task, Oliver, through its coordination of specialized AI agents, stood out as the sole offering capable of nearing human-level performance," the study said.

The study noted that Oliver was far slower than rivals, “likely due to its agentic workflow, which breaks tasks into multiple steps. This process runs in the background and can potentially improve reliability and thoroughness, despite the longer overall run time.”

Billion-dollar bets on agentic AI

Legal is a wealthy industry to mine for agentic firms. In the U.S. alone, the industry generated nearly $350 billion in 2022 (the most recent data available), according to the U.S. Federal Reserve. That represents 3.5% annual growth over two decades, doubling from $175 billion in 2002.

Legal AI firms raised significant venture capital in 2025: Andreessen Horowitz led a $160 million investment in Harvey in December that valued the firm at $8 billion. Legora scored $150 million led by Bessemer Venture Partners that valued the company at $1.8 billion and Noxtua pulled in $92 million in April led by C.H. Beck. For its part, Jones’ Flank won $10 million in June from investors including Insight Partners.

Companies are adopting agentic tools from vendors or building their own. Salesforce developed Agentforce, a tool the company estimates will cut 9,500 hours annually from compliance and risk tasks. Other companies, including Spanish energy firm Repsol and U.S. law firm Morgan Lewis, use tools from Harvey and Thomson Reuters' CoCounsel. 

The success is about giving attorneys the tools to work faster and cheaper, which gets noticed by clients, said lawyer Arthur Rothrock, CEO and co-founder of agentic firm Legion Law. Demand is spurring hiring at his company. 

“We're hearing from users that clients are thrilled with what AI-assisted workflows mean for their bills, and that's generating referrals,” Rothrock said. “Clients want those savings, which creates pressure on attorneys to adopt tools that deliver them.”

Jake Jones expects fast growth in 2026. In a recent LinkedIn post, he predicted “the agentic category will bloom” and that Legora and Harvey will start winning in-house deals “left, right, and centre.”

2026 will be the year of agentic AI, even if some of the terminology is poorly defined and marketed, Jones said.“I’m seeing that these agentic products are now possible to build and therefore are inevitable.“

AI may be rewriting the rules of product development, but some of its most interesting builders aren’t chasing hype. They’re learning, often the hard way, how to ship something people actually want.

Millette Gillow learned that lesson early. While working with founders inside The Tech Bros — a deliberately ironic name for a collective and accelerator that supports early-stage builders — she kept seeing the same pattern repeat: slick demos, enthusiastic feedback, and then… nothing. Products that impressed in pitch meetings quietly failed once they reached real users. “It’s super easy to end up building something that demos well yet no one wants,” she said. “That’s why I highly recommend mom-testing anything you build before you build it.”

By “mom-testing,” Gillow means stripping away insider language and imagined use cases, and asking a brutally simple question: would a non-technical person actually use this, understand it, or pay for it? Inside The Tech Bros’ programs, she’s seen founders abandon features that sounded impressive but failed that test – AI copilots that required too much setup, automation tools that saved seconds rather than solving real pain, or products that users praised in theory but never returned to in practice. What users say vs. what they do

That kind of blunt pragmatism has become a hallmark of a new wave of AI product makers: technically fluent, design-literate, and increasingly skeptical of labels like “AI-native.” Gillow argues those definitions miss the point. “They’re inconsistent and can be quite woolly,” she said. “If your core product is AI, you can probably call yourself AI-native. Otherwise, you’re just a company — and that’s absolutely fine.”

The focus, she argues, should be on whether something works in the messy reality of people’s lives. “Great businesses are great businesses, whether they’re ‘AI-native’ or not, as long as they’re solving a real problem.”

That tension between what users say they want and how they actually behave is something Swedish founder Amalia Berglöf has encountered firsthand. She is the founder of Crewcial, a community-tech startup aimed at helping professionals build meaningful, durable networks — not just collect contacts. Early on, she believed she knew what that should look like. Her users, she thought, wanted structured networking tools and clear prompts for engagement.

In calls and demos, they told her exactly that. Then they used the product — and did something else entirely.

“What people say in demos and in calls isn’t even close to what they do as soon as they try it out,” Berglöf said. Users described an idealized version of themselves: highly social, proactive, and intentional. Their behavior told a different story. Engagement dropped when interactions felt forced. Features designed to encourage participation often created friction instead.

“I realized I was running into a catch-22,” she said. “I needed to build a better basis before even building the actual networking experience.” Instead of doubling down on her original idea, Berglöf paused development and re-examined the problem. She began observing what users gravitated toward organically — passive signals, lightweight interactions, and moments that felt useful rather than performative.

Listening to behavior rather than feedback reshaped Crewcial’s direction. “I listened to my users instead of sticking to my idea,” she said. “I’m so thankful for that.”

The speed trap

Both founders agree that AI has made experimentation easier than ever — and that this is both a gift and a trap. Tools like Cursor and Claude, which Berglöf calls “life-changing,” allow solo founders to move from idea to minimum viable product (MVP) in minutes. But that speed can mask deeper issues. “Scaling responsibly is harder than building fast,” she said, especially for non-technical founders who may not see where systems will break under real-world use.

That pressure to move quickly is also forcing builders to rethink what “good product sense” means in the AI era. Berglöf worries that some founders are outsourcing judgment too readily, particularly when it comes to user research. “I’m worried about people using LLMs for user research without understanding their biases,” she said. In areas like health, she noted, models trained on skewed data can quietly reinforce gaps — including poor coverage of women’s issues — leading to products that simply don’t work for large parts of their intended audience.

In her own work, that awareness has shaped how she evaluates AI-driven insights, treating them as inputs rather than answers. “Those who understand the problem and understand distribution will win,” she said, not those who assume the model knows best.

The invisible assistant

For Gillow, the end goal is technology that disappears rather than dominates. “I’m a big believer in tech that gets out of our way,” she said. The most successful AI products, in her view, will feel less like tools and more like quiet infrastructure: invisible personal assistants doing useful work in the background rather than demanding attention.

Design plays a critical role in making that possible, and Gillow believes it’s an area where startup culture is shifting. “There’s a certain snobbishness among ‘tech bros’ when it comes to good design,” she said. “But technical ability and an eye for aesthetics are a dream combo.” She argues that women builders, often combining engineering, design, and user empathy, are helping to reset those priorities.

Berglöf sees the same trend in Sweden’s startup scene, supported by structural factors such as a law that allows employees to take six-month sabbaticals to start a company. “We’re seeing more women who can both code and design, and that’s incredibly powerful,” she said. “Understanding your client’s job-to-be-done is more important than technical skills today. Slapping an AI label on your product doesn’t work anymore.”

Both founders see the next phase of AI innovation as quieter, more grounded, and more consequential. Berglöf hopes it will be defined by companies tackling problems like climate resilience, education, and democracy – less obsessed with the mechanics of AI, and more focused on why it’s being used at all.

Gillow puts it more simply. “Build stuff that helps people,” she said. “And make sure it actually ships.”