In partnership with

Happy Wednesday, Fintech Listeners!

I’ve got a great podcast for you today, so grab your headphones and your notebook and keep scrolling!

— Alex

Was this email forwarded to you?

Subscribe Here

Sponsored by Plaid

The best financial products have something in common: the infrastructure underneath them is completely invisible to the people using them.

Nobody thinks about bank verification when they donate to a GoFundMe disaster relief fund, or about identity confirmation when they link a bank account to a HealthEquity HSA.

That invisibility is what good infrastructure buys you: the ability to focus on what you're building.

Plaid put together an ebook on 4 companies that got it right: GoFundMe, H&R Block, HealthEquity, and Gemini.

Each company had a different problem: idle charitable dollars, fraud-related service costs that dropped 70% in a quarter once they fixed verification, crypto onboarding friction, fragmented financial data.

The product and engineering leaders who solved these problems walk through how in Plaid's new ebook, aka 21 pages of required reading.

Download it here

3 BIG IDEAS FROM THE PODCAST

Simon Taylor and I are back with Not Fintech Investment Advice, and in the spirit of The Odyssey (which I am so excited to watch on the big screen!) we spent a lot of the conversation talking about what happens when financial institutions let AI labs, agents, models, and workflows inside the gates.

Simon Taylor is brought an agent control plane for large financial institutions trying to figure out what to do with AI agents, a platform for training models on your own financial data instead of borrowing someone else's, and a franchise lending platform he knew I wouldn't be able to resist.

I brought skepticism, and one company. A casino credit card that I haven’t been able to stop thinking about.

We also closed the episode ranting about the only correct reading of the movie Frozen, which is a topic that we, as fathers of young daughters, have very strong opinions on.

Tune in for the full conversation here

And read below for my three big ideas...

#1: A Very Nice Horse

Large financial institutions want to use AI agents. They also need to understand what those agents are doing, what data they’re touching, which models they’re using, how much they’re spending, and whether any of this adoption is actually producing outcomes.

Simon framed the present state nicely: adoption is high; productivity is not.

He cited a recent study across the US, UK, Germany, and Australia where ~70% of firms report using AI, and 89% of executives say it’s had no impact on labor productivity over the past three years. Meanwhile, large companies are getting to near-universal usage quickly. The tools are everywhere. The operating model is not.

That leaves banks stuck between two incomplete modes.

One is giving everyone Microsoft Copilot and calling it transformation. Maybe your employees write slightly better meeting summaries. History trembles.

The other is the token-maxing leaderboard model: let teams build agents everywhere, connect whatever they need, and discover the governance problem the hard way.

Primitive is interesting because it sits between those two modes. It gives regulated institutions a way to build agents, observe them across different models, and keep the governance layer separate from any single AI lab.

That last part is the strategic point.

OpenAI and Anthropic are not forward-deploying engineers into banks and core providers out of the goodness of their algorithmic hearts. They want the workflows and data; they want to understand how financial services operates from the inside. Once an AI lab has mapped those workflows, will it be content letting all the value accrue to banks and their existing vendors? Or will they spin off their own companies to sell directly to banks, or start providing banking services directly?

We don't know how they'll monetize that expertise. But there's no guarantee it'll be good for banks. Nice horse, but why are you just giving it to us?

I can’t help but think about RTP and the clearinghouse: the same anxiety community banks have had about letting shared infrastructure providers too close to the thing that makes them competitively singular. That same logic applies here, at a much larger and more significant scale.

The defensive answer is more practical than it sounds. Open-weight models are now good enough for many internal workloads, especially when the alternative is paying frontier pricing for every use case. (Simon's framework: run most workloads on-premise, call out to frontier models only where they uniquely add value, and use governance infrastructure to firewall the entire thing).

Another option is bringing in engineering talent that takes orders from you rather than from an AI lab. Simon sees McKinsey-style firms rebranding around exactly this model. This way, you get the capability without giving anyone a permanent map of your house.

#2: The Syntax of Transaction Data

The big idea Simon and I have both been circling for years is this: your transaction data is not just data. It has syntax.

Revolut proved it. They built a foundation model called PRAGMA trained not on language but on financial events: every login, every tap, every payment became a token. The model learned the grammar of customer behavior the same way a language model learns the grammar of English prose. The result replaced six separate machine learning systems. Fraud recall improved by 65%. Credit scoring improved by 130%. Simon wrote about it in Fintech Brain Food, and confirmed on the podcast it took six weeks and, he believed, under half a million dollars to build.

Stripe got to a version of this earlier, in payments fraud. Their question going in was whether payments fraud data has a similar internal structure and syntax to written language such that transformer architecture would work on it. They didn't know. The answer turned out to be yes. Nubank acquired a company called Hyperplane (listen to this podcast for more on what Hyperplane does) in 2024 partly to take the same capability off the market before competitors could access it. Affirm has been running its own experiments and found this architecture outperformed their existing gradient boost machine learning models.

A good way to think about it is using a basketball analogy, which someone at one of these companies shared with me.

A player's career free throw percentage says one thing. The fact that they've missed five in a row tonight says something more immediately relevant if you’re trying to decide who should shoot the technical free throw at the end of a tie game. Traditional machine learning models can easily see career statistics. A transformer-based model trained on event sequences can see and weigh both career stats and in-game performance appropriately, in real time.

Plus, there’s a regulatory advantage here. You’re not asking a generative model to approve or decline a loan. You’re using it to find new signals in the data, which can then be fed into more conventional, deterministic predictive models and rule-based decisioning systems.

A regulator doesn’t want a bank saying, “The model generated an answer and we trusted it.” But a bank can more plausibly say, “This model helped us identify a pattern, and that pattern became one input in a controlled underwriting model.”

#3: Fintech + Franchises

Simon brought Exponent to the episode because he knew it was an Alex Johnson-type company. He was right, and I want to explain why.

Franchising is entrepreneurship in a box. Someone else built the playbook, the brand, the operations. You execute it in your geography and own a large chunk of the economics. Simon Googled Shaquille O'Neal's franchise portfolio mid-conversation: 155 Five Guys locations at peak, 50 restaurant franchises, 40 health clubs. He has made more money after his playing career than he did during it, and he had franchise ownership to help thank for that.

We sold homeownership as the default American wealth building story, but owning a franchise might be the better path for a lot of folks.

Usually, the ones accessing this path most often are first-time entrepreneurs, recent immigrants, and people without long established credit histories in a particular community. And the infrastructure serving them is broken. Specialized lenders know they're the only option and price accordingly. Franchise brokers sit at the access point and extract a lot of value out of it. SBA lending exists (loans guaranteed by the U.S. Small Business Administration), but most SBA lenders find simpler deals to pursue.

Exponent is building the financial operating system for this segment: loans, charge cards, AI-assisted bookkeeping, and a deal room for document intake, tracking, and closing. They’re a licensed SBA lender in all 50 states.

The long game is what interests me most. Every franchisee on the platform generates better data on their specific franchise brand. Over time you can build underwriting models not just for franchisees generically but for specific franchises. What does a Crumbl Cookies unit look like versus a Domino's versus a home services franchise? That granular, systematic data doesn't exist today in any organized form. Whoever builds it first will be very hard to dislodge.

The people who end up wealthy in this space are the ones who own twelve Orange Julius locations across three states. Fintech has spent the last fifteen years largely ignoring them. That’s a mistake worth correcting.

WHAT I'M LISTENING TO

#1: SpaceX, AI Bubble Fears, and The Age of the Trillion-Dollar, Zero-Profit Company (Plain English) 🎧

A good episode on a weird potential future for the stock market and the global economy overall.

#2: What Banks Can Learn from Trader Joes (The Community Bank Podcast) 🎧

I like banking. I like Trader Joes. I like cross-industry comparisons. I liked this podcast episode!

*Bonus: Lending, Unbundled (by me, with TruStage) 🎧

Risk doesn't vanish when a loan changes hands; it just changes addresses. In Episode 1, former NCUA Chairman and Acting Comptroller of the Currency Rodney Hood joins me and my co-host for the series, TruStage's Bjoern Nordmann, to explore why a regulatory framework built for vertically integrated lenders no longer fits an industry that's split a single loan across separate players.

*This rec is brought to you by one of our fantastic brand partners.

Thanks for the read! Let me know what you thought by replying back to this email.

— Alex