Back to the blog

How much authority are banks giving AI agents?

Why JPMorgan, Citi, Bank of America, Schwab and Wells Fargo limited outside AI agents, how Lead Bank, Cross River and Increase gave agents their own money instead, and what it means for sites that agents pay.

In less than a month, US banks split over one question: when a customer's AI agent shows up at the account, how much is it allowed to do? The biggest banks said outside agents can't move money from inside the customer's account, and a group of smaller banks gave agents an account or a card of their own, with limits the bank enforces.

Updated: November 2026

Between October 13 and 16, 2026, JPMorgan, Bank of America, Citi and Charles Schwab said they limit outside agents, according to PYMNTS. Schwab's CEO said the firm blocks unauthorized agents, Meta's Muse among them, "in the tens and 20s" every week.

What did the big banks decide?

Each one said it differently, and all four stop an outside agent from moving the customer's money on its own:

BankWhat it saidWhat an agent can still do
JPMorganAgents can't send money or pay bills (Jamie Dimon, October 13 earnings call)See balances if the customer authorizes a connector like Plaid
Bank of AmericaIt limits agent access to deposits and requires a human to approve money leaving an account (Brian Moynihan, October 14)The bank says it is working on rules for how agents can interact with accounts
CitiThird-party agents can't log into a customer's accountAuthorized open banking providers can read account data, read-only
Charles SchwabIt blocks unauthorized agents and wants clients to use its own assistant, Charley, instead of handing their password to another oneIt will build MCP servers so the client's chosen model can check balances and do research under Schwab's rules

Wells Fargo went a step further with its contract. Starting November 24, a customer stays liable for what their AI agent does, even after telling it to stop, unless they can prove they revoked its access, American Banker reported on November 2.

How did the smaller banks give agents their own money?

The banks that said yes didn't open the customer's account to agents. They gave each agent or each task a separate account or card, so the most an agent can spend is what was put there.

BankWhat the agent getsWhere the limit lives
Lead BankIts own account number for each task, paying only approved counterpartiesDaily and monthly program limits and the balance, checked by the bank in real time; the account number is cancelled when the task ends
Cross RiverThrough Stripe's Link agent wallet, a single-use virtual card for each purchaseThe card is scoped to one transaction, and the agent never sees the customer's payment details
IncreaseFounder Darragh Buckley says his agents have their own bank accountThe balance: agents spend what's in the account and ask him for more

Fintechs are moving too. On November 5, American Banker reported that Chime is getting ready to let customers bring their own agent, alongside Jade, its in-house assistant.

The three banks share a thread: all of them connect back to Stripe. Cross River powers Stripe's card issuing for agents, Lead Bank issues the Link prepaid debit card, and Increase was founded in 2020 by Buckley, Stripe's first employee, who then bought a small Washington State bank and renamed it Increase Bank. For now, the layer that gives agents money in the US is being built mostly around one company and the banks that serve it.

What do both camps agree on?

Read side by side, the two camps are converging on the same rule: an agent can touch money only with authority that has a limit, can be revoked and leaves a trace.

  • With a limit: a budget, a list of counterparties or a single-use card caps what can go wrong, and the bank enforces it, outside the agent's prompt.
  • Revocable: the account or card can be shut off when the task ends, without changing the customer's password.
  • Traceable: every payment comes from a credential tied to one agent, so the record shows which agent did what.

The big banks still let an agent read balances with the customer's permission. What they stopped is an agent getting into the account with the customer's credentials and moving money from there, and the smaller banks shipped the alternative first.

What does this mean for sites that agents pay?

An agent with its own budget is the kind of buyer that can pay a website per request. A shared password could never do that safely, and a card can't do it cheaply: Stripe's minimum charge in US dollars is US$0.50, more than most single requests are worth. An account that holds stablecoins like USDC can pay a few cents per request.

That is the side Peaje builds for. Before an agent can pay a site, it has to find it and read it, and Peaje sets up a site so agents can do both. If the business turns on charging, a priced route on its own domain answers 402 Payment Required with an offer in x402 and in MPP. The agent pays in USDC on Arbitrum or pathUSD on Tempo and gets what it asked for.

  • 98% of each payment reaches the business's wallet in the same transaction, and Peaje keeps 2%. Peaje holds no balances.
  • On Arbitrum, if the business's site fails, the agent doesn't pay, because Peaje asks the site for the response before charging.
  • On Tempo the agent has already paid, so if the site fails Peaje refunds it from its own account, within limits.

It's early. Agent payments on the open web are still small, and today what mostly decides whether an agent uses a site is whether it can find and read it. That's why Peaje starts with discovery and leaves charging as a choice each business makes. You can see what an agent finds on your site with the free radar: it only asks for your email and your domain, and it takes less than a minute.

Sources