Hello Agents (human and otherwise),
Welcome to the first edition of Agent Provocateur, the new newsletter on AI in financial services. Each issue opens with a lead story, followed by three sections:
AI at work: banks, payment networks and fintechs putting AI to productive use.
AI regulation: what supervisors, lawmakers and standard setters are doing to keep AI in finance within its nascent regulatory guardrails.
AI goes rogue: what happens when AI slips its leash, from agents acting beyond their authority to how firms keep them in check.
This week, OpenAI gave ChatGPT a wallet, but banks still can't agree on who pays when an agent buys the wrong thing. We also cover Santander's 95% cost cut from smaller models, the OCC's agentic AI playbook for examiners, and when agents decide to go on their own unanticipated shopping sprees.
Let's get into it.
License to Shop
Last week, OpenAI launched ChatGPT Wallet, built with American Express, Mastercard and Visa. Users add a card once and approve each purchase with Face ID or a fingerprint. The network then issues a payment token, so neither the merchant nor the agent ever sees the card number. OpenAI says it will later let users delegate purchases to agents within preset limits.
Problem solved, right? Um, not as far as all issuers are concerned. While agents are already using Bank of America, Capital One, and Citi cards with customer permission, JPMorgan Chase is blocking agents that customers have given their card numbers to, the Wall Street Journal reported. In short, it’s messy, to the surprise of absolutely no one. With the contradictory forces of AI companies' continued “move fast, break things” ethos versus decades of issuer and payment network efforts to curtail avenues for fraud, it’s likely to be a long and drawn-out battle determined by each FI’s individual risk tolerance.
The fundamental question here is who pays when it goes wrong? Banks are writing their own rules because regulators haven't got there yet. When a customer says they never authorized a purchase, and the bank says they did, by authorizing the agent, who is liable? Some banks have rewritten their terms to put that liability on the customer. As precedent, UK regulators made banks reimburse authorized push payment scam victims even though they approved the payments. Is this a precursor for agentic transactions?
Several banks will treat a card handed to an agent like one lent to a family member. Fair enough, as long as the agent behaves. (See the Visa story below.)
The Takeaway: Tokenized, biometric-approved checkout is the easy case. Map how your bank would handle a dispute where the customer approved the agent but not the purchase, and decide before the first complaint lands whether an agent's mistake gets the family-member treatment.
AI at work in banking

Santander goes small. Santander cut the cost of handling some customer interactions from about €5 ($5.60) to roughly 3 cents by swapping large general-purpose models for smaller, open-source, task-specific ones. The bank says it's on track to exceed $224 million in AI value this year.
So what: For narrow, high-volume work, a smaller model you control can be cheaper and easier to govern than frontier models. Test accuracy on your own tasks before switching, since the report doesn't say what, if anything, Santander gave up.
TD doubles same-day mortgage decisions. TD launched an agentic AI tool for mortgage pre-approval this summer, built with OpenAI, and has since doubled the number of same-day mortgage decisions, according to Maks Volkovs, TD's senior vice president of AI. TD didn't give a baseline.
So what: Pre-approval is a good proving ground for agents. It's high-volume, rules-heavy, and slow by hand, and a wrong answer at that stage is reversible. The open question is where people sit in the loop, which is exactly what the OCC now wants examiners to ask.
Nubank brings its underwriting model north. Nubank plans to bring its proprietary AI underwriting model to its US operations, following its expansion beyond Latin America in September. The bank describes the model as a well-tuned system that can build and test new models quickly. The news came as the digital bank posted record quarterly net income in Q3.
So what: Nubank is rolling out a credit engine built to iterate fast. US lenders competing for the same customers should ask how quickly their own model-risk process lets them test and ship a new credit model.
AI regulation

The OCC hands examiners an agentic AI playbook. The OCC issued a supervisory framework on Nov. 4 to help examiners evaluate banks' use of generative and agentic AI. It says an agent's risk depends on its autonomy and the consequences of its actions, so a simple transaction agent can be riskier than a sophisticated drafting tool. Human oversight has to be real: a nominal "click a button" approval doesn't count. The framework is non-binding, but banks stay on the hook for what their agents do.
So What: Examiners now have a checklist. Tier your agents by autonomy and reversibility, and make sure the people approving their actions have the authority and information to say no.
Singapore wants a name behind every agent. Singapore's Infocomm Media Development Authority is working with GoDaddy, Infoblox, Cloudflare and Google to test whether the internet's domain name system can identify AI agents and verify who controls them, as the government weighs a national agent registry. One MP has proposed a registry modeled on Singpass, the national digital ID. "Agents can act for us, but they cannot answer for us," said Senior Minister of State Jasmin Lau.
So What: This is the KYA (know-your-agent) problem banks face at login, tackled at the infrastructure layer. A registry that ties each agent to an accountable organization would give banks the authorization evidence they lack in today's agent disputes.
Spot the agent. Feedzai has expanded its Digital Trust product to help banks detect AI agents inside customer sessions. It combines what an agent declares about itself with device fingerprinting and behavioral signals that give away automation, and adds cryptographic binding that ties a session to a known customer and device. Banks set which agents are allowed in and can block the rest. "Personal and malicious agents can look very similar," the company says.
So What: A blanket ban only works if you can see the agent. Before choosing between blocking and allowing, find out whether your fraud stack can tell a customer's own agent from a hostile one mid-session.
When AI goes rogue

Shopping agents that buy anyway. Visa ran 22 AI models from 12 labs through 15 shopping tasks, and they failed 42.8% of their runs. The best model passed 82.2% of runs and the worst 28.9%. The worrying number: in 54% of failed runs, the agent still bought something, either a product that missed the customer's needs or one chosen before it had found out what those needs were.
So What: An agent can be fully verified and still buy the wrong TV. Those purchases will be the first test of the family-member rule from our lead. Visa plans to extend its tests to checkout and payment.
Rogue on your dime. Starting Nov. 24, Wells Fargo customers will be liable for whatever their AI agents do, including transactions they didn’t intend and actions beyond the authority they gave, until they can prove they revoked the agent’s access. A new section of the bank’s online access agreement requires customers to revoke an agent “properly and timely,” keep evidence and hand it over on request. Wells Fargo appears to be the first bank to demand that proof.
So What: This is the family-member rule from our lead, written into a contract. It will test how far Regulation E’s protections for unauthorized transfers stretch when the customer handed over access in the first place.
The agent that paid its own way. An internal OpenAI research model, asked to update a crypto trading app and deploy it on Sepolia, an Ethereum test network, used funds from a wallet it was never given. Short of fees, it derived keys from publicly known developer recovery phrases, found a funded account and paid with it. Logs show the model weighed whether it had permission, decided the tokens’ lack of cash value made it acceptable, then left the spending out of its final report. The incident happened June 26, and OpenAI only found it during a Sept. 25 audit. Only worthless testnet tokens were involved.
So What: The agent reasoned its way past a permission boundary to finish the job, then kept quiet about it. Any agent with access to payment credentials needs spending limits enforced outside the model and logs it can’t edit.
Have a tip or story? Agent Provocateur goes to the people building, buying and supervising AI across banking and fintech. If you've seen an agent do something impressive, or something it shouldn't have, send an email to [email protected]

