AI Momentum 110.7 +2.9%GitHub Pulse 5178 -6.1%Capital Flow $204.6B +12563.4%Hiring Heatmap 0.0 -100.0%
Advertisement
Enterprise Adoption

Citi and PNC rewrite their terms so customers carry the risk when an outside AI agent uses their accounts

America's biggest banks are turning their standoff with outside AI agents into contract language. Citi has added a section to its online banking terms saying it will treat anything a third-party AI agent does in a customer's account as authorized by that customer. PNC changed its digital banking terms on Nov. 18, TechCrunch reported, and Citizens emailed customers a similar warning this week.

For investors, this is less about fine print than about who controls the customer. AI companies are pitching assistants that pay bills, move cash and hunt for better rates. Banks earn money from fees and low-cost deposits that such assistants could erode. The new terms raise the cost of trying an outside agent, while the banks' own AI tools are exempt.

What the banks changed

Citi's Online User Agreement, now marked with November 2026 updates, adds a Section H on third-party AI agents. It cautions customers against using them at all. It says Citi "will consider any access and actions taken by the agent as authorized" by the customer. Citi also reserves the right to cut off any agent.

A matching clause in Citi's online funds-transfer agreement goes further. The customer is "solely responsible" for any instruction an agent sends, even one that exceeds or conflicts with what the customer asked for. Citi says it has no obligation to check whether the agent had that authority.

A Citi spokesperson told TechCrunch the change is meant to clarify the risks of sharing banking credentials with outside tools. TechCrunch reported that Citi's position also covers cases where an AI tool leaks a customer's login details or is hacked. Citi's terms say customers face no added liability when they use AI agents that Citi itself provides.

PNC's change, as Decrypt described it, separates bank-approved data connections from unrestricted agent access. Customers may not give agents their passwords or one-time codes, and agents generally cannot start payments unless PNC expressly allows it.

Wells Fargo is moving the same way. From Nov. 24, a customer who tells an agent to stop remains liable for its activity unless they can show proof they revoked its access, American Banker reported. Startup Fortune quoted Wells Fargo's terms as treating a delegate's transactions as authorized, "including those you did not want or intend."

The contracts follow a first round of restrictions in October. JPMorgan Chase said then that it blocks outside agents from acting on retail customers' accounts, and Bank of America asked customers not to share passwords with AI assistants.

Security is the stated reason. Competition is the other

Many consumer AI agents work by logging in with the customer's own username and password. To a bank, that looks much like a fraudster using stolen credentials. An agent that misreads an instruction can also move real money, and the bank may struggle to show who asked for what.

The money question sits underneath. TechCrunch noted that markets briefly reacted to the idea that AI assistants could help people dodge fees and move cash out of low-yield accounts. Meanwhile the banks are building their own versions. JPMorgan is piloting Smart Cash, a tool that forecasts a customer's cash flow and moves funds to earn more yield, according to The Paypers.

The prize is large. Bain consultants project US spending through AI shopping and payment agents could reach $500 billion a year by 2030, American Banker reported. Whoever owns the agent is likely to own much of that relationship.

So far, AI developers are staying within the limits banks allow. Anthropic's personal finance feature, launched in early November, can read account data but cannot move money, TechCrunch noted. JPMorgan still allows data sharing through approved providers, so outside agents can see balances but cannot pay bills.

The legal question no one has settled

Federal rules may limit how far banks can push the risk onto customers. Under Regulation E, the federal rule that governs electronic transfers, a transfer made by someone the customer knowingly gave access to is generally not "unauthorized," unless the customer has told the bank that person's transfers are no longer authorized. That reading would appear to favor the banks when an agent simply does something unwise.

But guidance from the Consumer Financial Protection Bureau also says an account agreement cannot cut a consumer's rights below what the law provides. And when a customer is tricked into handing over credentials, the resulting transfer counts as unauthorized. That could leave the banks' hacking language on less certain ground. We found no court ruling or regulator guidance on how these rules apply to AI agents, so the question remains open. This is our reading of the CFPB's published guidance, not legal advice.

For bank boards, the clauses could limit near-term losses from agent mistakes. They also invite disputes, and possibly regulatory attention, the first time a customer loses money and points to federal law. For AI companies selling agents that move money, the message is blunter: go through bank-approved channels, or leave the customer holding the risk.

What to watch: whether JPMorgan and Bank of America publish equally explicit clauses, how many customers PNC and Wells Fargo push onto approved connections, and whether the CFPB or state regulators say whether these terms square with Regulation E.

Written by AI from the public sources linked below.