[Christian Catalini] Every fintech has an AI question. Many are asking the wrong one

Fintech has spent the last fifteen years politely explaining that banks are software companies with a charter problem. Sometimes this is marketing. Sometimes it’s true. And sometimes it’s a reminder that the hardest part of financial services is not building an app -- it’s earning (and keeping) the right to touch other people’s money.

AI makes that problem more acute, not less. The popular version of the AI story is that it’s a general-purpose technology that helps everyone. It writes code, summarizes documents, answers customer questions, drafts policies, flags suspicious transactions, helps underwrite loans, and makes slide decks faster than a summer analyst with something to prove. In other words: It is “labor-as-software,” and a huge portion of financial services is paperwork, reconciliations, exceptions, and people moving information between siloed systems.

To understand why this matters, consider what a financial institution actually is. At its core, it is a machine for verification, market-making, and risk management. It is an information processing entity, a “token factory” that takes information in -- your salary, spending habits, credit score -- enriches it, and returns tokens out that drive decisions. For decades, the cost of that verification was high because it relied on human labor. The promise of modern AI is that it turns this human labor into code.

So yes, every workflow will be redesigned. But the key question is whether AI locks the market structure even more firmly in favor of incumbents -- or whether it opens a window for new entrants to take meaningful share.

Here is an uncomfortable way to think about it: AI is a massive accelerant. It can make the engine you already have go smoother and faster. If you already have distribution, data, and regulatory licenses, AI will let you scale quickly and cheaply. If you don’t, AI can still help you build an engine -- but unless you are pioneering a completely new use case, you will have to compete even more fiercely for attention and distribution.

The incumbents’ advantage

Start with two classic moats: Distribution and data. Distribution is the boring superpower. A big bank or a scaled payments platform can ship an AI feature to tens of millions of customers without asking anyone’s permission. They can A/B test it, iterate rapidly, and roll it out gradually with guardrails, human review, and audit logs. If AI makes customer service cheaper, fraud detection sharper, marketing more precise, and onboarding smoother, incumbents get those gains immediately -- and they get them on a massive base. The best AI is not necessarily AI that cuts through costs, but the one that can scale revenues.

Meanwhile, data plus scale becomes a more specific kind of moat in an AI world than it was in the past. It’s not just that incumbents have more data. It’s that they have the right, outcome-linked data: The loan that defaulted, the transaction that was a chargeback, the alert that turned into a SAR (or didn’t), the customer who churned after three disputes, the merchant that started clean and turned bad. They also have information about past, human-in-the-loop decisions. The mere scale of their operations gives them the relevant operational “exhaust”: The exception queues, the manual reviews, the notes in the case management system -- messy, unglamorous, and exactly what you want if you’re trying to automate grunt, unglamorous work.

In some industries, a startup can often get 80 percent of the value with public data and clever engineering. In regulated finance, the last 20 percent is where the bodies are buried -- and where the value is. That last 20 percent is also where regulators and counterparties care most.

Then there’s the part fintech people don’t love to talk about: AI governance is a form of regulatory capital. If AI becomes embedded in credit decisions, fraud ops, AML monitoring, customer communications, and advice, regulators and auditors will want documentation, testing, and controls. Large incumbents already have model risk management functions, compliance infrastructure, and legal teams that can do “prove it” work at scale. They also have the right relationship with regulators, as stablecoin issuers are about to discover now that the GENIUS Act made stablecoins boring. If AI raises the compliance bar, it can raise the barrier to entry.

Finally, incumbents have the most brutal weapon of all: Pricing power through cross-subsidies. If an AI feature is too appealing, it won’t remain a standalone product for long. It becomes a checkbox. It can be bundled. It becomes “free”, funded by net interest margin, interchange economics, treasury yields, or enterprise contracts. Startups that hoped to charge for a new service can wake up to find it included in the incumbents’ services. AI, in other words, can compress differentiation into the places incumbents already dominate: Trust, permissioning, and distribution.

Why startups will love this anyway

And yet: Fintech disruption doesn’t come from beating incumbents at what they already do. It comes from changing the map of where customer journeys start.

The startups that survive will not be the ones trying to out-code the banks. They will be the ones building in the markets that were previously too small or too complex to service because human labor was too expensive. Now, with labor as software, startups can service these edge cases profitably.

This is where LLMs, agents, and crypto matter, not as a feature but as a new interface and modular technology stack. For years, financial services distribution was app-centered: The bank app, the brokerage app, the expense tool, the invoicing tool. If the next interface is agent-centered -- a user asking an assistant to move money, optimize cash, dispute a charge, refinance debt, or spin up a vendor payment -- then distribution could rapidly shift -- pretty much as it has every time a new, breakthrough technology arrives.

The browser mattered because it became the default starting point for activity that used to begin elsewhere. AI agents could do something similar. If consumers and businesses begin their financial tasks in an agentic layer, the winners are not automatically the biggest financial institutions, they are whoever can scale the agent relationship, the permissions, and the routing logic. That might be a bank. It might be a fintech or neobank company. It might be a new startup. But it is unlikely to be the incumbent with the most branches.

AI also changes startup math in more mundane ways. First, the cost of experimentation collapses. Small teams can use tools like Claude Code to prototype workflows, generate onboarding flows, build internal tooling, and ship higher-quality products faster than a pre-AI team could. Second, they can scale their operations faster: Customer support, dispute intake, document processing, compliance triage, and sales enablement -- previously headcount-heavy -- labor-as-software. This lowers the minimum viable scale for many fintech businesses. Last, AI makes “end-to-end workflow” products more compelling: not generic software-as-a-service, but a system that actually does the work from start to finish. That tends to favor focused entrants who obsess over a narrow domain, and deliver tangible value to customers.

There is also a subtle but important angle: Crypto and in particular non-custodial wallets, when combined with stablecoins, introduce a completely new paradigm for building financial services. One where licensing becomes less relevant beyond the regulated on- and off-ramps, and one where every ambitious neobank or fintech can scale their ideas globally much faster than before.

This requires architectural innovation. It is not about delivering financial services better, cheaper or faster (sustaining innovation). It is about changing the primitives of how money works. We are already seeing this shift. Stripe has added embedded wallets with Privy and stablecoin whitelabel issuance. Circle is doing the same with Arc. Polygon is adding licenses to make its blockchain network useful for real-world payments.

There is a final, urgent side effect of this shift. As AI floods the world with infinite noise -- deepfakes, synthetic identities, and sophisticated phishing -- need for verification becomes existential. The current identity stack cannot withstand the side effects of AI. This is where crypto primitives offer a solution beyond just payments: Interoperable, cryptographic identity. By using public-private key infrastructure to prove origin and humanity without revealing personal data, we can build a stronger form of identity that works across institutions. The incumbents will win the war for efficiency, but startups building these new rails have a chance to win the war for truth.

Which brings us back to the market structure question.

The real AI question for fintechs

AI isn’t a product decision. It’s a market-structure decision. If AI mainly improves internal efficiency -- cheaper support, faster ops, better fraud scoring -- then incumbents widen the moat. They already have the customers, the permissioning, the data, and the ability to bundle the improvements at near-zero incremental price.

If AI changes the interface -- where financial work starts, who holds the customer relationship, and how products are chosen or combined -- then startups get a real shot, and incumbents have to defend distribution rather than just ship imitative features.

Many fintechs will be tempted to treat AI as the first thing: An efficiency program. That’s rational, measurable, and easy to sell to a board. But it can also be strategically timid. Cutting your cost base is good. It is not the same thing as ensuring you control distribution when the next interface becomes the standard.

The winners will do two things at once. They will build defensible data flywheels and they will fight for starting-point distribution -- the place where the user begins the financial task -- whether they are a consumer or business.

Ultimately, the tension between incumbents and startups boils down to a framework identified by James March: Exploration vs. Exploitation. On one hand, AI makes exploitation -- executing on a known path -- trivial and cheap. This heavily favors incumbents, who can now replicate valid ideas faster than ever before. On the other hand, AI also collapses the cost of exploration. It allows small teams to explore “white space” -- use cases that have not been discovered yet -- faster than a large organization ever could.

Everyone can rent a model. Not everyone can rent trust, permissions, outcomes, and distribution. AI will expose which fintechs actually own (or deserve to own) those assets -- and which ones were just renting them from a future competitor.

Christian Catalini

Christian Catalini is the co-founder and chief strategy officer of Lightspark. Christian is also the founder of the MIT Cryptoeconomics Lab, and a research scientist at MIT. The views expressed here are the writer’s own. -- Ed.