Happy Friday, Fintech Takers!

As promised, today’s newsletter is a very loosely organized summary of my initial takeaways from this week’s Agentic Readiness Summit.

- Alex

Stablecoins and payments collided

For most of their history, stablecoins were a crypto story. This year, they became a payments story.

Card programs, cross-border rails, and banks building their own tokenized versions all showed up in production instead of pilot decks.

But which of them will last? Some are already core infrastructure. Others still have big problems to solve, starting with the rules.

Come dig into it with me at a live stablecoin year in review on Dec. 15.

Are We Ready for AI Agents?Copy anchor linkCopied

The short answer is “no.”

The longer answer is more nuanced and interesting. I haven’t had quite enough time, since the Summit, to form my thoughts on the question into a cohesive answer. So, instead, I thought I’d just share some of my unfiltered thoughts and observations from the event (adjusted, of course, to comply with the Chatham House Rule).

I hope they’re useful!

1. Where are the regulators?Copy anchor linkCopied

There was a lot of surprise among the Summit participants at just how absent U.S. federal banking regulators have been on AI. Their biggest move this year was a subtraction: April's revised model risk management guidance (SR 26-2) explicitly carved generative and agentic AI out of scope, and the request for information on AI that the agencies promised still hasn't arrived.

The OCC took a small step forward recently, publishing an exam framework for AI that aims to scale regulatory scrutiny to the consequences of an agent failing. But the OCC is careful to say it "sets no new supervisory expectations or enforceable standards."

That should not give any comfort to financial institutions, however. Existing laws are almost certainly being violated by AI agents and tools right now. They just aren't being enforced. That will change. State attorneys general and state regulators, many of them newly staffed up, can enforce parts of federal consumer financial law, as well as their own UDAP statutes, and when there’s a change of administration in 2029, the approach to supervising and regulating AI will almost certainly change as well.

2. Delegated Trust vs. Human in the LoopCopy anchor linkCopied

Agentic finance runs on delegated trust. Consumer protection regulation runs on the opposite assumption: That a human reads the disclosure, gives consent, and initiates the action. These are obviously in tension with each other.

That said, it’s worth asking: Does active human participation actually work that well today? Nobody reads the terms of service.¹ Most consent is reflexive, not considered. In practice, a lot of consumer protection is process theater: A ritual that shifts liability more than it protects anyone. Delegated authority also isn't new. We've had authorized users, powers of attorney, and executive assistants forever. The novel part with agentic AI is delegating authority to something that's automated, non-deterministic, and can't be cross-examined. That might force regulators to focus on outcomes rather than rituals, which, honestly, might not be the worst thing.

3. Activity, Not EntityCopy anchor linkCopied

One of the most useful maps from the Summit sorted agent builders by how regulated they are. Banks are fully supervised. Fintech companies are covered by the CFPB, the FTC, and state law but less frequently examined. Big tech only gets reached when it performs a covered financial activity. And the frontier AI labs, which train the underlying models and deeply shape how agents behave, aren't regulated as financial services providers even as they become application companies themselves. (Anthropic launched Claude Personal Finance this week.)

That's why I think enforcement, when it comes, will be based (mostly) on the activity, not the entity. U.S. bank regulation is mostly entity-based, but consumer protection law mostly isn't, and it's the only lens that reaches everyone in this picture. Take credit repair. The Credit Repair Organizations Act covers any company besides banks and credit unions that, for valuable consideration, provides advice or assistance to improve a consumer's credit, and it requires written contracts, specific disclosures, a three-day right to cancel, and no payment until the work is done. Lots of people are asking AI assistants (that they pay to access) how to raise their credit score. Whether that triggers the law is an open question (and I'm not a lawyer), but the statute was written broadly on purpose. The same goes for UDAAP, fiduciary duty, and Reg E. None of them care whether the advice or action came from a human or a model.

Liability may follow the same path. Agent providers are busy disclaiming responsibility for what their agents do, but product liability usually flows back to the manufacturer when a product does something it wasn't designed to do. The hard part is proving a design defect in a model nobody fully understands.² Either way, "we're just a tech company" likely won't be a defense if your product is being used to facilitate or enable regulated financial activities.

4. Two Policy Analogs: Open Banking and Public Cloud ComputingCopy anchor linkCopied

There are two precedents that policymakers will likely borrow from once they engage more deeply with agentic finance.

The first is open banking. Progress in this area has only been possible because banks and data aggregators reached a stalemate. Aggregators needed bank data, and banks couldn't fully stop screen scraping (though they could frustrate it), so both sides had reasons to deal. The Personal Financial Data Rights Rule is currently stuck in the mud, but the stalemate produced private deals anyway, like JPMorgan Chase's paid data-access agreement with Plaid.

Can banks force AI companies into the same kind of stalemate? I tend to doubt it. The AI companies don't need banks the way aggregators did, consumers are already bringing agents to them, and blocking those agents is a losing game long term (more on that below).

The second is public cloud computing. Regulators couldn't supervise the hyperscalers directly, so they reached them through banks' third-party risk obligations and exams of technology service providers, effectively ring-fencing the parts of Amazon, Google, and Microsoft that serve banks and pulling them inside the regulatory perimeter.

Could that work with the frontier AI labs? Partly. It works for first-party agents, where the bank is the lab's customer and can flow requirements down through the contract. It will be much harder for third-party agents, where the lab's customer is the consumer or a non-bank AI application provider and the bank has no contract to lean on.

5. First-Party Agents vs. Third-Party AgentsCopy anchor linkCopied

This is a very important distinction when talking about AI agents.

A first-party agent is built (or bought) and controlled by the bank or fintech company and works inside their own systems. A third-party agent (Muse, Instinct, ChatGPT, Claude) is the customer's, and it shows up at the bank's front door with the intent and authorization to act on their behalf.

Third-party agents are much, much harder. With a first-party agent, the bank or fintech company controls the state, the APIs, the authentication, and the approval screens, and it hears the customer's request directly. With a third-party agent, the bank or fintech company controls none of that and only sees the agent's interpretation of what the customer wanted. A majority of the hard questions that were raised during the Summit (liability, authentication, consent, blocking) live on the third-party side. Conflating the two is the fastest way to talk past each other.

6. The Missing Ingredient is IntentCopy anchor linkCopied

Why is liability so much harder for third-party agents? Because the bank or fintech company never sees the customer's intent. With a first-party agent, the customer asks the bank or fintech company directly, so if something goes wrong, it's either on the customer or on the institution, and the institution protects itself by requiring an approval before anything happens. With a third-party agent, the bank or fintech company only sees the agent's translation of what the customer wanted.

If ChatGPT gets the translation wrong, who's liable?

Payments might be solvable here, because the scope is narrow. "Send $15 to Jake on Friday" is easy to pass through and easy to confirm, which is why protocols like Google's AP2 are built around signed mandates that carry the user's intent along with the transaction. But a lot of financial services prompts are hopelessly broad. What did you mean by "make the number go up"? How aggressive is "aggressive"? One Summit attendee noted the irony: Payments, the scariest part of the liability problem, may also be the most solvable part, precisely because the intent is clearer.

7. Wells Fargo Wants a ReceiptCopy anchor linkCopied

Wells Fargo's updated online access agreement, which takes effect November 24, says customers are responsible for "properly and timely revoking any authorization provided to an AI Agent," and that they "must retain evidence of such revocation and must provide such evidence to us, at our request." Until the customer notifies the bank, they're on the hook for whatever the agent does.

Here’s what the bank is really saying: We can't see the agents, so you have to manage them. The bank's spokesperson said as much, explaining that Wells "may not otherwise have visibility" into when an authorization ends. It's a perfectly rational position for Wells Fargo to take. It's also a strange one to impose on consumers, who now have to prove a negative about software running on someone else's servers.

8. You Can't Block Them ForeverCopy anchor linkCopied

Right now, big banks can block most AI agents. Big bank security is far more sophisticated than the average e-commerce site's, and when an agent like Muse comes knocking from a known IP address, it's easy to spot and turn away. But that's a temporary advantage. Phone use, where an agent operates a real phone (physical or emulated) in the cloud, with the customer having authenticated into their bank apps on it once, is maturing fast. When the agent is tapping through your mobile app on a real device, it looks exactly like a human.

It reminds me of Morpheus explaining the agents in The Matrix: "Every single man or woman who has stood their ground, everyone who has fought an agent, has died." Blocking buys time. It isn't a long-term strategy.³ The banks that use that time to figure out what a sanctioned front door looks like will be in much better shape than the ones that assume the current wall will hold.

9. Receive, but not send?Copy anchor linkCopied

Remember how FedNow rolled out? Many banks joined FedNow to receive first because receiving is relatively easy and low-risk, and held off on sending, because sending is where fraud, liability, and rapid deposit flight live. As recently as 2025, a survey found only 4% of financial institutions had both send and receive enabled for FedNow.

I wonder if we'll see the same pattern with banks and AI agents? You can open an account with an agent, and fund it with an agent, but you can't move money out with one. It would be a self-serving asymmetry, but an understandable one. Deposit flight was a recurring worry at the Summit, with one attendee warning that agents could turn "all money into hot money."

Customers will notice the asymmetry eventually. But it may be how AI agent access starts.

10. Agentic Commerce Needs Everyone at the TableCopy anchor linkCopied

The protocols (Visa Intelligent Commerce, Mastercard Agent Pay, Google's AP2, OpenAI and Stripe's ACP) only work if issuers, merchants, AI companies, and wallet providers all opt in, and so far they mostly haven't. JPMorgan Chase, for example, isn’t enrolled in Visa Intelligent Commerce, which is a big deal given that JPMC is the largest card issuer in the U.S.

So Stripe built a workaround. When an agent like Meta's Muse buys something, one option for payment is a single-use virtual card through Stripe Link (Cross River is the issuing bank). The virtual card is charged for that one purchase and then Stripe charges the consumer’s debit or credit card separately. It works great for Stripe, which arbitrages the difference between commercial card interchange rates (which it earns on its virtual cards) and consumer card interchange rates (which it pays when it charges the consumer). It works less well for everyone else. The consumer's issuer sees very little information on the transaction. Link's terms say the consumer "will not receive your underlying card offers, benefits, or rewards." And the merchant has no real idea an agent was involved. Not ideal.⁴ But until everyone agrees on something better, the workaround is the market.

11. Plaid in the MiddleCopy anchor linkCopied

Plaid is now sitting squarely between the banks and the AI companies. This week it became the connection layer for Claude Personal Finance (read-only), and it already powers bank connections for ChatGPT and a number of other AI assistants.

That's a valuable place to be.

It's also a risky one.

Banks expect Plaid to hold the AI companies to the same restrictions it applies to every other fourth party⁵, usually as spelled out in bilateral data access agreements: What data can be pulled, how often, and for what purpose. But the AI companies aren't like other fourth parties. They're enormous, fast-moving, and used to setting terms rather than accepting them. They also use data in ways that are fundamentally different from the ways in which even the most sophisticated pre-AI fintech companies use it.⁶ Can Plaid be the bouncer the banks want and the best friend the AI companies want (and, quite frankly, expect) at the same time? If it leans too far toward the AI companies, banks will tighten the pipes. If it leans too far toward the banks, the AI companies may just start working with the banks directly.

In fact, that’s already happening! The AI companies are talking to the big banks directly. They are currently exploring read access, and eventually write access, with the banks’ permission. And that poses an interesting question: If an AI company can get direct API connections to the top 10 banks, how much does it still need an intermediary? Especially if that intermediary’s superpower in these types of negotiations — the ability and willingness to scrape banks’ websites directly — is a superpower that the AI companies already have, to an even greater degree?

Plaid's best defense will be to own something the AI companies can't replicate bank by bank: A direct relationship with the consumer. Stripe is showing what that looks like on the payments side with Link, a consumer-facing wallet that agents plug into and that travels with the user across merchants. Plaid needs the equivalent for financial data: A place where I, the consumer, see every agent I've connected to my financial data and can grant, scope, and revoke access in one spot.⁷ Plaid Portal is a start, but it needs a Stripe Link-level investment to become a mainstream consumer-facing product.

12. Clippy Was Right (But Too Early)Copy anchor linkCopied

There's a fun irony in where we've landed. Ask Jeeves bet in the late '90s that people wanted to ask questions in plain English and get answers back.⁸ Clippy bet that people wanted an anthropomorphized assistant that noticed what they were doing and offered to help. Both got the form factor exactly right. Both were mocked mercilessly, because there was no brain behind the form factor.

Now we're circling back. Today's agents are Ask Jeeves (answer my question) and Clippy ("It looks like you're traveling in 12 days and haven't bought travel insurance") with an actual brain attached. The lesson for banks and fintech companies: The interface was never the problem. Quality was.

13. Quality is the Whole ProductCopy anchor linkCopied

Speaking of quality, here’s a simple but important point: The value of an agent is its quality.

That's it.

Can it get the job done for you, correctly, without you having to check its work? Everything else (the model, the interface, the protocols) is a means to that end. One speaker at the Summit put it bluntly: How an agent is built determines its quality, and quality is the whole product.

If that's true, the pricing model that follows is outcome-based pricing: You pay when the job gets done, not per seat or per token. Sierra already prices this way for customer service ("Sierra gets paid only when we complete a task for you"). It's the most lucrative model for a provider that's actually good, and the most customer-aligned model for everyone else.⁹ The hard part is agreeing on what counts as an outcome, which, in financial services, is often a question with real regulatory weight.

14. At Minimum, Play DefenseCopy anchor linkCopied

Across the internet, AI agents have already passed humans. In June, Cloudflare said agentic bots accounted for 57.4% of global web traffic, and 68.6% in North America. Banks aren't there yet. Based on data from several institutions, one speaker estimated that agent traffic to banks is still under 10% overall, but closer to 10–20% in customer service.

That makes sense, because agents dramatically lower the cost of complaining. One attendee described their agent reading their inbox, noticing that a promised refund hadn't posted, and offering to email the bank about it. Multiply that by millions of customers, and a support team built for 25,000 tickets a month could see that number triple overnight. You can (for the moment) block an agent from logging into your app. You can't block it from sending an email, filling out a complaint form, or calling your contact center.

That's why, whatever a bank or fintech company decides about letting third-party agents through the front door, at a minimum it needs to build first-party agents. Not as an innovation project, but as a defensive necessity. The only way to absorb agent-scale volume is with agents. And the work isn't wasted: The plumbing you build to answer your customers' agents (clean APIs, measurement, traces) is the same plumbing you'll need if you eventually decide to let those agents in or to go on offense, more meaningfully, with your own first-party agents.

15. How to Build AI Agents That Actually WorkCopy anchor linkCopied

We had a great presentation and discussion on how to build first-party agents that actually provide value. Here are a few of the most important pieces of advice.

Start narrow, and start somewhere you can measure. Customer service is the obvious place, not because it's the most valuable use case, but because you've already built the measurement apparatus: Resolution rates, reopen rates, and NPS for every ticket type, which gives you a human baseline to beat.

Keep the agents narrow, too. One speaker shared research showing that if you expose just 50 APIs to a model, it picks the right one for a given customer request only 13% of the time (43% if you narrow the tools first). Then the errors compound. Most financial services workflows are 10 or 25 decisions in a row, and a model that's right 95% of the time on each step gets all 10 right only about 60% of the time. At 90%, it's about 35%. In workflow automation, mistakes don't add up, they multiply. The fix is one job-to-be-done API for each agent, which runs every step, approval, and audit trail under the hood, rather than 50 different tool-centric APIs that the agent has to sequence on its own.

Don't chase perfection. Getting to 80% takes days; the last 20% takes a year, and that's where nearly all the cost lives. If the AI agent is at least as good as a typical human, ship it, route the hard cases to people, and move on to the next use case.

Finally, trace everything. And I mean everything: The customer's request, the model's reasoning, the tools it called, the final response. Traces make evaluations possible, they're the training data you'll use to move from expensive frontier models to cheaper fine-tuned open-weight ones, and they're exactly what regulators will ask for. The Financial Stability Board's October report floated log tracing of every agent action, and the OCC's new framework puts the consequences of agent failure at the center of exams. Done well, a traced agent is more auditable than a human whose notes are incomplete and whose memory is fallible.

16. When AI Agents Beat AI-assisted Human EmployeesCopy anchor linkCopied

The most common way that community and regional banks are using AI today is Microsoft Copilot or something like it: AI-assisted human employees. It's the safe choice, and it usually comes from a vendor the bank has already approved.¹⁰ But for customer-initiated workflows, it's the wrong form factor, and AI-assisted employees may actually be riskier than AI-enabled customers.

Think about the chain of translation. A customer explains their problem to a banker. The banker interprets it and types a prompt. The model interprets that prompt and generates an answer. The banker relays it back. That's three layers of probabilistic interpretation stacked on top of each other, and each is a chance for intent to get lost. A first-party agent that talks to the customer directly removes a layer, and it can be measured, traced, and improved in ways that a thousand individual employee prompts can't.

Copilot is a strong option for internal productivity. It's not a substitute for building the agent your customers (and their agents) will actually talk to.

17. What do we actually mean by "relationship"?Copy anchor linkCopied

Ask a community banker how they compete and you'll almost always get one word: Relationship.

The problem is that "relationship" is vague, and vague words are hard to stress-test. One of the speakers at the Summit broke it into five more specific things:

1. Trusted Expertise
2. Direct Access
3. Personalized Service
4. Local Context
5. Personal Friendship

Once you get that specific, the question of how AI disrupts relationship banking gets a lot more interesting.

Take trusted expertise. Is the average community banker a better expert on any given financial question than Claude? Honestly, often not, and if I were a teller at a bank, I would constantly be thinking about whether the answer I just gave a customer will be different from the answer that Claude will give her once she leaves the branch.

That said, trust isn't just about getting the right answer. It's about accountability when the answer is wrong. The model can't be fired. The banker can. Meanwhile, other dimensions, like direct access, could actually get stronger with agents handling the routine stuff, freeing up bankers for the conversations that matter.

18. How do you sell a feeling to a robot?Copy anchor linkCopied

AI Agents are very good at comparing things that can be measured: Rates, fees, limits, approval odds. They're much worse at weighing things that can't: "They picked up the phone when my business was in trouble," or "they care about my community." If your value is abstract, emotional, or otherwise hard to quantify, how do you convey it through an agent? And if you can't, how do you avoid getting dragged into a race to the bottom on price?

The conversations at the Summit uncovered a few partial answers.

One is to make your value legible. Publish a plain-text, agent-facing version of your website with the actual product details agents look for. Another is to specialize. Be the best bank in the country for teachers, or for SBA lending in Nashville, so that an agent with deep context on its user has a concrete reason to pick you. Agents know far more about their users than Google or NerdWallet ever did, which makes niche product-market fit more discoverable than it's ever been.

However, I don't think anyone has figured out how to put "we'll take care of you" into a format an agent can parse.¹¹

19. The Miracle on 34th Street TestCopy anchor linkCopied

In Miracle on 34th Street, Macy's hires a Santa Claus¹² who starts telling parents to buy toys at Gimbels when Macy's doesn't have what they want, or when Gimbels has it cheaper. Management initially panics, but then realizes customers love it. Honesty turns out to be the best marketing strategy Macy's ever had.

That's the bar for first-party AI agents. If you want customers to actually use your agent, it has to tell the truth, even when the truth is inconvenient for your bottom line. "You'd earn more on that money at another bank" has to be a sentence your agent is allowed to say.

The agentic finance wars to come will boil down to a contest over honesty. First-party agents will be tempted to push their own products, and a regulator could view that as steering under UDAAP's "abusive" standard, especially when the agent talks to you in the first person like a trusted human advisor would. Third-party agents face the mirror-image temptation: They're expensive to run, and the obvious way to pay for them is advertising, which is, inherently, a little dishonest.

Whoever convinces consumers they're actually on their side wins, and the easiest way to do that is to just tell the truth.

WHERE I'LL BECopy anchor linkCopied

Two more to go! We can do this!!!

✈️ AFC Policy Summit | November 17 | Washington D.C.Copy anchor linkCopied

I’m now moderating a very interesting fireside chat at AFC. Make sure to stick around until the end, if you’ll be attending!

✈️ Fintech NerdCon | November 18-20 | San DiegoCopy anchor linkCopied

Still some room at the dinner that I’m hosting on Wednesday night at NerdCon. The focus is on lending, so if that’s an area you spend time in, let me know and I’ll get you on the list!

WHERE I'VE BEEN Copy anchor linkCopied

*The Science (and Fiction) of Friction (by me, with Spinwheel) 📚Copy anchor linkCopied

Okay, technically this is where I was back in September, but the conversation was too good not to resurface. Two decades of stripping friction out of financial services assumed a human on the other side, and AI agents break that assumption. Based on our (riveting!) conversation, we created a report with a test for any product: could an agent handle it, or does it warrant explicit human authorization? Check it out!

*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