JPMorgan Chase Deposits Grew 8% and Wells Fargo 11% in the Quarter AI Agents Got Blocked — Wall Street Sold the Wrong Half

Between Oct. 13 and Oct. 23, six large U.S. banks restricted third-party AI agents operating inside customer accounts, and the market traded the news as a deposit event. Read bank by bank and the pattern is the opposite of a run: closed the write path and left a narrow read path open.

What is missing from every one of those statements matters more than what is in them. No bank quantified a loss, and Reuters could not independently verify the Axios reporting that triggered the sweep. Citigroup said it has seen no reports of agents accessing client accounts at all, and Morgan Stanley disputed the premise outright. Supports a control problem rather than a loss event.

On Oct. 13, on its own earnings call, Citigroup was asked by Mike Mayo about the “AI scare trade” — the idea that better cash sorting and software agents could pull deposits out of banks. CEO Jane Fraser answered that she is not worried about deposit flight risk in Services, citing a network that is hard to replicate, a very high win rate and very little attrition. Nine days later Axios reported agents inside customer accounts. Was live before the story that supposedly justified it

Two numbers dissolve the scare. JPMorgan Chase and Wells Fargo grew average deposits 8% and 11% in the same quarter they were writing agent-blocking policy, and neither reported a single dollar of agent-caused outflow. The discount being applied to bank franchises therefore prices a risk that no bank has quantified.

JPMorgan Chase's digital services agreement lets it treat instructions and transactions initiated through a customer's own credentials as evidence of authorization, and makes account holders responsible for supervising software that misinterprets instructions or exceeds them. Wells Fargo's updated access agreement, effective Nov. 24, 2026, reaches the same result from the opposite direction: customers answer for agents they authorize, including unintended activity, and the bank may suspend an agent's access. Consent now lives in the credential, not in the channel

The rails being assembled in October are not deposit rails. Visa, Mastercard, Cloudflare and Okta are all selling the right to approve a payment, while the balance stays where it already was. Every one of them sells permission rather than float

Follow the money and the aggregator loses while the edges gain. Plaid keeps the data role but has no path to deliver write access, while Cloudflare and Okta sell exactly what banks now need to tell an agent from a customer. Mastercard projects one in ten consumers will routinely buy through agents by 2030 and that global agentic commerce revenue could reach $5 trillion by the end of the decade. Shrinks the aggregator's mandate and hands fraud compliance to the identity layer

Over the next two quarters the catalysts are legal and disclosure events, not product launches, because JPMorgan Chase has said there is no timeline for letting agents operate accounts directly. Turns on contract dates and disclosure language, not shipping features

Over one to three years the question is not whether deposits leave, but how the authorization toll is split. If agents hold scoped tokens, the economic center of retail payments shifts from the deposit relationship to the credential issuer, pulling value toward Visa, Mastercard and the identity and visibility layers around them. Shifts the center of retail payments from the deposit to the credential

The Q3 numbers were not hurt by agents. Citigroup said outright that it is not worried about deposit flight, and its own automation is the larger story: technology vendor spend down 30% year over year, product launches 30% to 40% faster, and about 100 of its most manual processes reengineered. Marked down the balance sheet and ignored the fee that is being repriced