A few weeks ago, I had strong words for conversational finance:

“…people have better things to do than think about the questions they might ask of a new tool…The response to ‘make this easier for me’ should not be a more complicated product.” 

Conversational finance works when it avoids a conversation.

Recap

A copilot, the quintessential embedded conversational AI tool, simplifies and shortens what clicks and keystrokes would make more complex, time consuming, or error-prone. 

For a consumer, it should solve for effort and angst; for a business, it should solve for time and cost. Cut the term “copilot” and the design principles are universal. Conversational AI should make proven features better and faster. 

Blank-slate, open-ended, chat-first or chat-only AI does not; it reinvents a round wheel as a square one.

This article is the third of three in a series. Missed part two? Read it here.

Easy prompts, quick answers

Consumers want a product that just works. Businesses want a product that just works and helps make or save money. Conversational AI frequently augments or adapts proven tools that are tuned to customer needs. Easier. Faster. Better. Less thinking, more doing, unless thinking is the point:

• The user of conversational anything needs context for the tool as much as the tool needs the user’s context to add material value.

• The output must be focused on what the user will want to know or do without failing because a predetermined prompt-and-response pair doesn’t exist.

The most common type of prompt to ChatGPT is to ask for specific information, perhaps expecting it to return specific information. “Embedded” chat gives users context for their prompts. They need answers to specific problems in front of them. The chat box guides users toward the best prompts, and if it needs to clarify, it prompts the user with helpful answers.

Then, it produces results that are quick to digest and act upon with a click or a one-line response.

Semiautonomous banking

The Erica-style chatbot, built upon a library of countless prompt-and-response pairs, is obsolete. Agentic AI supersedes that mechanism; the interactions, responses, and actions it enables vastly surpass what was possible in 2022.

Monzo’s development process for its Ops Agent is an interesting case study: start with a Q&A tool, automate guardrails, and then enable actions (e.g., internal money movement, goal-setting) that are completed semiautonomously from within the chat. Assuming the agentic AI customer safety and compliance nightmare is solved, and the bot works as advertised:

Banking will have a copilot that does everything the traditional app can do. Sometimes better. Sometimes worse. But distinct from any interface that’s been seen before.

Resources

• Intuit scrapped its own AI agent architecture twice in four months (VentureBeat)

• Engineering the Future of Customer Operations: The Monzo Ops Agent (Monzo)

• Conversational Finance: Truth and Friction (Fintech Notebook)

Thanks for reading Fintech Notebook: Weekly intelligence read by fintech operators, strategists, and investors. Published by Tyler Brown. Connect on LinkedIn or visit my website @ tylerbrown.co.

Recap

AI shatters a decades-old channel model. The banal truism “meet the customer where they are” is code for “spend more on mobile.” Vendors have met the demand with a flexible device format, or more likely, tacking on a separate platform. The strategy is wrong; the solution is wrong.

Each app is a channel (a delivery mechanism and interface). An AI tool is a channel and a platform (an infinitely scalable software tool). The average consumer has a new experience; the proficient consumer gains immense power to structure and manipulate information.

Banks are not built for a channel-as-platform. They are limited by by the technology and business siloes created for reaching customers over decades.

Featured asset

US consumers most often use mobile to manage their bank accounts and it's most common with millennials and Gen Z. The Gen Alpha data isn’t real; see the thought experiment below. (Source: American Bankers Association-Morning Consult, Fintech Notebook)

“Mobile is the forever future” is absurd. A thought experiment: What will Gen Alpha’s channel use look like in 2030? 100% mobile?

That would be like saying in 2006 that “online banking will always be the future.” “Mobile-first” and “mobile-native” have been industry shorthand for years. Shorthand today is suddenly “AI-native” and “AI-first.”

Banks are stuck on the mobile era. The 2026 channel model doesn’t exist.

This is the second article in a series that explores how AI will influence consumer expectations and behavior in banking.

“It feels like the world is accelerating, but not to benefit society as a whole…

Society breaks when disruption touches the social fabric, and breaks faster when society has no time to adapt.”

Read Pt.1: Mistrust in AI.

Deep dive

A channel is a delivery mechanism for something. In the digital realm, it’s an app. In banking, it’s an app for financial information and tasks. Market sizing is easy: Age group links directly to device ownership and the adoption of new technologies. Demand changes with demographics. The business case is clear.

With AI, the frame is cracking. Trust in AI varies widely, and the link to adoption is loose at best. Strategies for customer-facing AI aren’t as clear as saying, “our customer strategy is our channel strategy is our mobile strategy,” and link it to Gen Y and Gen Z consumers. And even that gets murky: Gen Alpha hasn’t played out.

Channel-as-platform

Product scope and user control are clear for the average app (a channel; a delivery mechanism and interface). Features are built or assembled by a developer, and the output is purposely constrained.

Chime, for example, is an “average app.” For an AI app like ChatGPT, product scope and user control are nearly limitless.¹ Prompt-and-response pairs are constrained only by the model. Users can customize a system by writing in English.

ChatGPT is also a platform (an easily customized tool that hosts other apps) and a channel-as-platform. Its context vastly exceeds data that’s retrievable from a traditional database, and it it can compete with traditional apps directly. That should terrify companies whose apps just furnish a laundry list of features.²

Channel (mobile) strategy cannot be “build or buy a new app that has more features, or jump on the bandwagon for something flashy and new. The question to address is, “what is a channel in 2026, and what’s its value?

The commercial value of customer-facing AI

The commercial value of an AI tool is hard to measure; it’s unproven as a customer acquisition and engagement tool. Usage can be measured broadly (active users, the frequency and outcome of customer behaviors). But with AI, it’s hard to say what exactly consumers are doing, what it might be worth for a platform participant, like a bank, and the cost of participation in the first place. Without a prompt history and results, it’s a bet. Nearly nobody knows the odds.

Share of consumer ChatGPT messages broken down by high level conversation topic. Values are smoothed over the trailing 28 days (source: OpenAI/NBER). See “Resources” at the end of the article.

The AI labs do a lot of research on AI use and periodically publish findings. The data her can’t be split into financial services use cases but it’s a hint of what consumers are using these tools for.

Where is this going?

Nobody knows how front-end AI will play out in financial services. It will not just be to drop bank statements into ChatGTP.³ Nor will it be a magical PFM tool, or a chat-only interface.⁴ Some fintechs and enterprise software companies have thought through how AI might work as a channel, or part of a channel (a delivery mechanism and interface). AI as a platform is not so clear. Part 3 will focus on what’s being done now, what’s working, and address questions that are still open.

Subscribe to receive part 3.

Resources

• 2025 National Survey of Preferred Banking Methods (American Bankers Association)

• “How People Use ChatGPT” (OpenAI / NBER). Read pps. 13-18.

Thanks for reading Fintech Notebook: A weekly GTM intelligence briefing read by fintech operators, strategists, and investors. Connect on LinkedIn or visit my website @ tylerbrown.co.

AI disclosure: The author used AI to assist chart creation.

“Do I trust AI to be right for me and the world?” That’s a technical question and a human-impact problem staring people in the face. The answer is no, it’s not. Not for me; not for the world. We are living in a techno-dystopia made of emotions.

It feels like the world is accelerating, but not to benefit society as a whole. AI spread nearly instantaneously in 2022: It had no distribution friction via web browsers and mobile apps. Disruption was stratospheric. Society breaks when disruption touches the social fabric, and breaks faster when society has no time to adapt.

This is troubling: 40% of Gen Z weekly AI users say they’re anxious about it; 22% are angry. 24% are excited. That is a trust breakdown. When it bleeds into finance, it threatens how the industry withstands disruption. (Source: Gallup)

AI is a revolution in human behavior because AI is fundamentally a tool for humans. Trust in a technology is that it will help people make themselves better off, feel better off, or reap rewards from others doing the same. 

While builders create the next great model, society is grappling with what is happening in front of it right now.

A techno-dystopia is what’s happening in front of them: Frustration, fear, anger, resentment, for people and the technology. Will the social fabric tear? It’s fraying.

Trust in technology as a tool is the trust that it will work as intended, and increasingly that it will “choose” to work as intended, and not hurt anyone in the process.

A tool sparks anger when it falls short of what its builders claim it can do, or struggling to learn it feels like missing out.

AI has dramatically enabled new, grueling ways of working; it in theory makes us more productive per dollar and unit time.

Power users, and power-companies, are leading new ways of working. They see the upside in AI-driven productivity. They know what they want from an AI, how to achieve those goals, what silently fails, and how to guide it back to a useful answer.

The brutal underside is when the work that AI enables is the root cause, or the public reason, for layoffs, poor job prospects, longer or more intense hours yoked to a screen or on other tasks in the name of greater productivity per hour or per dollar.

AI is a human excuse. That’s why it has a trust problem.

Gen Zers who frequently use AI are more likely to think positively of it. But anxiety in particular still registers. (Source: Walton Family Foundation, Gallup).

How to interpret this backdrop?

The tension won’t go away. Part 2 pivots from the social backdrop of AI to strategic implications for the banking industry.

Resources:

• 2026 AI Insights Report (TD Bank, Big Query)

• Voices of Gen Z Study (Walton Family Foundation, Gallup)

Thanks for reading Fintech Notebook: A weekly GTM intelligence briefing read by fintech operators, strategists, and investors. Connect on LinkedIn or visit my website @ tylerbrown.co.

AI disclosure: The author used AI to assist chart design and layout.

Recap

Why does a personal finance dashboard matter, even when it’s something the world has never seen? Complex, queryable data with nearly instant analysis and advice, gorgeously rendered with nearly infinite options, is the future of… something. But people have better things to do than think about the questions they might ask a new tool.

Featured Asset

A tree map on common AI prompt types, adapted from NBER and OpenAI data. The original, with all categories shown, is in the NBER paper (see below).

The graphic: OpenAI and the National Bureau of Economic Research published a working paper in September 2025 documenting in extreme detail the topics of ChatGPT conversations. Scroll down to “From the Notebook / Resources.”

Deep Dive

An AI can create a remarkable experience, but conversation is hard—it’s not passive.

The Plaid x Perplexity and ChatGPT products released this spring look awesome. But prompting is a complicated way to look at finances because it demands active thinking. The learning curve to get sharp results is steep, and fine-tuning responses adds work. The response to “make this easier for me” should not be a more complicated product.

The user’s path of least resistance is to ask short, easy questions, hope for decent answers, and move on. If the answers aren't clear, or if they are formatted in a way that’s hard to digest, the product fails.

That is an experience problem fueled by four years of hype.

AI works best based on specific, scoped goals and clever prompts. Without them, an AI guesses what the user wants, burns time and tokens, and vomits more words than substance. What is a human user going to do with 500 words of AI bloviating? Publish a Substack?

Banking is a necessary chore; only power users or consumers with a deep-seated need make it more than that.

Most people want quick, digestible answers to concrete questions. Using an AI to get precise, meaningful results for complex issues is exhausting. “How do I optimize my mortgage payoff?” is not a frequent question, and it requires follow-ups.

AI has a ceiling as a consumer product because it underperforms by overperforming for many users.

“Sage,” a mockup mobile banking app. Named, coded, and substantially designed by Claude.

A Philosophical Footnote

Conversational finance is the next wave because it has to be. The technology exists, so it will be built. But a tool does little without an operator, and it’s tempting to over-teach the tool while under-teaching the mental model.

Even after I dropped out of computer science, one concept stuck with me:

Big problems need to be small problems first.

It stuck with me because it matters to code, philosophy, writing, and annual resolutions. A big goal alone will be a mess, turn into a different goal, or be abandoned and conveniently forgotten. A magic bullet exists only in retrospect, because the problem was already solved; reverse-engineering is always easier than creation. And it’s work.

Not wanting to work is a human problem. That makes it a design problem. Things should just work. Consumers shouldn’t need to understand the underlying mental model. Any product that forces critical thought creates a problem to solve one.

The AI power user creates specific, useful results based on specific, iterative prompts. The average consumer won’t spend the time trying.

Thanks for reading Fintech Notebook: A weekly GTM intelligence briefing read by fintech operators, strategists, and investors. Connect on LinkedIn or visit my website @ tylerbrown.co.

AI disclosure: This article used AI to assist chart design and production and design and code the app proof of concept.

Recap

Banks loathe Fiserv (their words, summarized). And there is no love lost with other core providers. Fiserv has a third of the market for community bank cores, and Premier is the dominant core. A few years ago at the American Bankers Association Policy Summit, I heard an executive from a mutual community bank with a delightful Appalachian drawl dryly joke-not-joked about bankers’ love for their core provider.

Featured Asset

The Core Provider Hate Index, adapted from American Bankers Association data.

The graphic: Core market share among the American Bankers Association membership (bubble size); satisfaction is from left to right and likeliness to renew from top bottom to top. The chart is labeled perversely as dissatisfaction and unlikeliness to renew, which are metrics that that ABA measures. Upper right is good. Lower left is bad. Dig into the data further and customer service, or lack thereof, turns up as a crucial factor.

Deep Dive

Decades ago, in-house data-processing units at several banks merged; the merged vendors kept buying other vendors. Now Fiserv supports 16 cores. (The last time I interviewed Dudley White, he barely suppressed a wry grin).

Big incumbents consolidated the market, competition vanished, and small banks were stuck with brittle technology and terrible contracts under the thumb of a B2B oligolopy.

Big banks more or less did what they wanted; Small banks and credit unions got leftovers and terrible contract terms.

Or that’s the story told in 300 pages’ worth of public venting.

Responses to an OCC request for information, in summary:

• Politely: “Core providers are horrible and regulators are complicit” (American Bankers Association)

• Blandly: “We are important; we are also a mess. $3 billion over five years will fix that, but until then you can have a new account manager” (Fiserv)

• Quietly: “The cores work hard to waste everyone’s time; our heartthrob at the CFPB is gone, and our policy team needs something to do” (MX)

• Obviously: “Thanks for the free ad campaign. If you made it this far, please buy our vendor data” (FedFis).

Among community banks, the top four core vendors have 79.9% market share. (Source: FedFis.)  ABA data also surfaces COCC and Finastra with a different methodology.

FedFis

Free data. Please download. An academic concept that sticks out is “primary tech stack” (core, online banking, mobile) and “core strategic,” the banks that have the same vendor for the primary tech stack. I would add that buying the bundle is cheap and easy, not necessarily a strategy. And for a tiny community bank, it’s probably the only choice.

Vendor concentration in key parts of the tech stack is high. About 75% of banks buy a bundled solution. Organic churn is 2-3%. (Source: FedFis.)

Trade Associations & Banks

Words dripping with acid. The dominant core providers abuse their dominance and regulators are complicit (their words, in short): Sales teams make the next solution sound like a masterpiece; it is not. Features that are sold don’t exist; after go-live, support trails off. The big-four oligopoly means that incumbents have no incentive to change.

Fiserv

Absolving oneself of sins. Community banks depend on the evolution of core platforms to support digitization, and the ability to outsource operations to vendors has meant that smaller institutions compete with larger banks (or what once was the benchmark). Fiserv is investing billions of dollars in technology; at the same time, its sales organization has been disorganized and the out for customers is to hope for a better account manager.

Thanks for reading today’s issue of Fintech Notebook. Insights for fintech builders, strategists, and investors by Tyler Brown.

Reach out to me by replying to this email, connecting on LinkedIn, or visiting my website @ tylerbrown.co.

AI assisted research and image generation for this article.

Recap

NCR bungled a 10-year run trying to make itself into a software company. It never committed; private equity and other companies are still picking up the pieces. It tried buying its way out of the ATM’s decline, but it failed to find a direction beyond selling a product it didn’t understand to an audience it thought it knew.

Featured asset

What’s in the graphic: Some major digital banking transactions from 2012-2025 involved private equity players. Digital Insight was passed around a few times over the years before it landed at NCR and was folded into Candescent. D3 Technologies, which NCR later acquired, was never owned by a PE firm.

Deep dive

In 2020 I interviewed Doug Brown, then SVP of digital banking, who pitched NCR’s physical footprint as the foundation for the “operating system of the bank.”

I said at the time that NCR would have to “patch together its new model from both emerging digital capabilities and legacy, lower-margin lines of business.” It never did.

NCR trapped itself by treating technology like a distribution channel, as if it would be a stronger competitor by bundling digital banking with ATMs.

What worked: The National Cash Register company dominated physical commerce for over a century. ATMs were a natural complement to hardware and software for cashiers and tellers.

What didn’t: NCR over roughly a decade assembled a digital banking product line through acquisition selling it through its bank distribution channel.

It stacked a minimum-viable digital banking business with advisory and support services that were built around its ATM business, and operated commerce solutions in another unit.

It wanted to be valued like a software company, but software was secondary.

In a strategic capitulation, in 2021 it bought Cardtronics, the world’s largest nonbank ATM provider and owner of the Allpoint network ($2.5 billion).

It doubled down on a hardware-heavy operation when they had wanted to be a software company but by their actions decided not to be, and the market valued it accordingly.

NCR failed to realize value from a mashup of hardware and software.

The strategy made sense from a business unit and sales organization perspective, but never from a technical perspective.

What’s left? 

• A legacy retail and hospitality business, superseded by vertical SaaS

• A legacy digital banking vendor competing with 20 others

• A division of a cash management company where scale drives success

For NCR, selling distinct products through the right sales channels with no hope for technical integration led to its demise.

• Digital banking didn’t fix strategic confusion.

• Acquisition was not a substitute for innovation.

• The “operating system of the bank” was an impossible dream.

Thanks for reading this week’s issue of Fintech Notebook. Intelligence for fintech operators and investors.

Reach out to me by replying to this email, connecting on LinkedIn, or visiting my website @ tylerbrown.co.

AI assisted research and editing for this article.

The Core’s Original Sin

There is no way to say with confidence which of those records belong to the same person. Many early systems are based on account numbers, and not designed to answer what the bank knows about a customer. To say with confidence that records in dozens of different systems belong to the same customer is a mental leap made best by engineers who are on the verge of retirement who know the system itself, and the technical fix is a colossal engineering challenge.

Was this email forwarded to you? Subscribe here.

This design foundation, built on technology and software that is its own problem, has created the tightly wound, “brittle” systems that the average bank is reluctant or unable to touch. A CIO probably won’t be fired for a status quo that preserves a conservatively run business. But they would certainly be fired for signing off on a decision that had a devastating domino effect on the bank’s critical functions. The exit ramp is systemically and personally perilous, and prone to half-measures.

The exponential problem

That’s five databases. What’s missing? Card processing, loan origination, CRM, bill pay, Zelle, contact center, fraud systems, document management, document imaging, wires, compliance, marketing platforms? There are probably dozens.

Note: Representative at a very small scale. Image: Claude

There isn’t a one-click solution to the database problem, even with a COBOL translator or an AI agent at your fingertips.

The half-measure solution

Necessary half-measures may include the years-long implementation of a middleware-type solution (connecting systems) or the painful and prolonged process of hollowing out the core (decomposing and replacing functions piece by piece). The dream of internal behavioral, transactional, and customer profile data is more of a hack than a reality when new vendors sit in their own siloes on top of the tightly wound system that’s already in place, and depend on or are isolated from other systems. They are another patch on an old quilt.

The basic anatomy of a mind-warping mess

It takes just a few truncated, simulated tables to see how the buildup happened, and got worse over time. The core was designed to track money, not people: The deposit ledger is keyed on account number. The account record points to the customer information file. That was fine for a bank that only offered deposit accounts, and only through a branch. If the customer data didn’t match on account opening and a safeguard wasn’t in place, the core created a duplicate record on account opening.

The deposit ledger (and seemingly why buying a branch network includes the purchase of deposits). Image: Claude.

Database hell begins. The original “source of truth” becomes the source of just “something,” complicated and obfuscating the picture of a customer that didn’t exist to begin with.

The customer information file, which links back to the deposit ledger with the customer information number. Consistency is not guaranteed. Image: Claude.

The ACH system is keyed on account number, which links back to the deposit account ledger, which links back to the customer information file. Enter a new, linear, clearly connected series of dependencies.

The ACH system, which links back to the deposit ledger, which links back to the customer information file. Image: Claude.

Fast-forward to the digital banking platform. There was a day, which may still be today, when after you opened an account you had to key in your account number and some identifying information before you set up your digital banking account. Once the account was linked to digital banking, much that was linked to the account could be resolved to digital banking. The digital banking platform resolves to a user ID, and contains event types, event detail, channel, and details about the session. 

The digital banking system is closer to the customer (via user ID) but can do little to solve the account-first record keeping system. Image: Claude.

The bill pay system links back to at least a few systems. Image: Claude.

The bill pay system resolves to the digital banking user ID and has its own key assigned by the bill pay vendor’s system. That information may appear in the account ledger in its own tab. The data is probably reflected in the account ledger as a check or ACH direct debit.

Despair is not the only destination

In a perfect world, of unified data across the bank, what would it look like? A customer graph. A graph would sit on top of every data source a bank could access and make them work together as one: One place where everything the bank knows about a customer is connected, resolvable, and accessible in natural language.

The “customer 360” is a buzzword in search of a solution that isn’t just a data dumping ground. For a unified view, a bank needs infrastructure built from the ground up. And that may mean building a new bank altogether.

Thanks for reading today’s issue of Fintech Notebook. Deep thinking and color commentary on the present and future of fintech by Tyler Brown.

Reach out to me by replying to this email, connecting on LinkedIn, or visiting my website @ tylerbrown.co.

AI assisted research, image generation, and editing for this article.