News analysis · Banking and AI agents
Banks now say your AI agent's actions are yours. What that means for the office that pays the bills
This week Citi changed its terms to make customers responsible for what a third-party AI agent does in their accounts, even if the agent is hacked, and PNC and Citizens followed with similar changes and warnings. Here's what Citi's new terms say, why business accounts already carry more risk, and five rules for any AI tool near your bank login.

In brief
Citi added a section to its online banking agreement this month saying that if you give a third-party AI agent your login, Citi will treat what the agent does as authorized by you, including what anyone does with credentials the agent leaks. TechCrunch reported on Nov. 18 that PNC changed its terms the same day and Citizens emailed customers a similar warning, after JPMorgan and Bank of America announced similar policies in October. These are consumer online banking terms. Business accounts have less protection to start with. The rule for any office: no AI tool gets a bank password. Use read-only feeds for data, and keep every payment release with a person.
- Citi's new Section H treats a third-party AI agent's actions, and any use of leaked credentials, as authorized by you.
- TechCrunch reports PNC and Citizens followed this week; JPMorgan and Bank of America moved in October.
- These are consumer terms; business accounts already carry more risk under UCC 4A and lack Reg E protection.
- No AI tool gets a bank password. Use read-only feeds and keep payment release with people.
- Banks have a real security point, and a competitive motive: Citi exempts its own agents.
What changed this week
On Nov. 18, TechCrunch reported that Citibank had changed its terms of service to tell customers who use third-party AI tools that anything that goes wrong is their problem, even if they get hacked. It reported that PNC changed its terms that day, that Citizens emailed customers a similar warning, and that JPMorgan and Bank of America announced similar policies last month.
We read Citi's terms ourselves. The Citi Online Banking User Agreement now carries a notice that it was updated in November 2026 to add a new Section H, "Use of Third-Party Artificial Intelligence ('AI') Agents." In plain English, Section H says:
- Citi cautions you against using any third-party AI agent to access your accounts, give Citi instructions or make transactions.
- If you give an agent your login anyway, Citi treats the agent's access and actions as authorized by you.
- If the credentials you gave the agent are shared or leaked to other people or agents, Citi treats that access as authorized by you too.
- Citi has no duty to investigate, even if the agent's actions conflict with your instructions. You're responsible for the agent's errors, misreadings and actions beyond what you allowed, except where the law says otherwise.
- It's your job to monitor the agent. Citi can cut off any third-party agent at any time.
- Citi's privacy notice doesn't cover the agent, and AI tools that Citi itself provides are excluded from the section.
The line that matters most: "Citi will consider any access and actions taken by the agent as authorized by You."
How we got here
This didn't come out of nowhere. In October, Banking Dive reported that Citi, JPMorgan Chase and Bank of America had said customers' outside AI agents, such as Meta's Muse, were blocked from accessing accounts, and that Fifth Third wasn't letting agents start transactions. Citi's head of AI for its services business said no one was connecting to Citi through agents yet because the bank didn't know what the models could do.
At the same conference, PNC CEO Bill Demchak said no bank would let agents transact in accounts without a human in the loop, and that AI companies would have to take on the liability themselves. And in early November, American Banker reported that Wells Fargo had started requiring customers to prove they had turned off an AI agent's access, or else be liable for its activity.
Why a business should care more, not less
The new terms are in consumer online banking agreements. A business owner might read this and think it's a personal finance story. It's the other way around.
Consumers already have protections that businesses don't. Regulation E, the federal consumer rule for electronic fund transfers, defines an account as one established mainly for personal, family or household purposes. A company's operating account isn't one.
For wire transfers, the rules in UCC section 4A-202 say that if a bank and its customer agreed on a security procedure, a payment order the bank accepts in good faith under that procedure is effective as the customer's order, whether or not the customer authorized it, as long as the procedure is commercially reasonable. Your business banking agreement sets out what that procedure is. If an AI tool logs in with your credentials and passes your security steps, the bank can say it followed the procedure.
So the banks' message to consumers this week is, for a business, close to where things already stood. Ask your banker and your lawyer how your own treasury agreement handles it. But don't assume the bank will absorb an agent's mistake.
If an AI tool has your bank login, the bank sees the tool as you.
Where this shows up in a real office
Few controllers plan to hand an AI agent the bank password. The risk is quieter than that. A browser assistant runs in the same browser where someone is signed in to the bank portal. An always-on agent like Microsoft's Copilot Autopilot works within the permissions its user already has. A bookkeeper asks a chatbot to "log in and pull last month's statements" to save time. Each of these can put an AI tool between your company and your bank without anyone deciding it should be there.
Five rules for any AI tool near your bank
- No AI tool gets a bank password. Not in a prompt, not in a browser profile it can use, not in a shared password vault it can read. Write that into your AI use policy.
- Use read-only feeds for data. If you want AI to help with cash reports or reconciliations, feed it data through a read-only connection your bank or accounting system supports. Anthropic's new Personal Finance feature for consumers shows the pattern: its launch post says access is read-only through Plaid, it can't move money, and Claude never sees the bank password.
- Payments stay with people. An AI tool can prepare a payment run or flag invoices that are due. A person reviews it and releases it inside the bank portal, with a second approver for anything above your limit.
- Keep browsers separate. Do banking in a browser profile with no AI extensions or assistants. Our guide to third-party apps with mail access covers the same idea for email.
- Know what to revoke, and keep a record. List every tool that touches financial accounts. When one is removed, note the date and who did it. Wells Fargo is already asking customers for proof.
The banks have a point, and a motive
Be fair to the banks. AI agents do make mistakes, and an agent holding a customer's password is a new target for criminals. Asking the bank to cover every loss from a tool it didn't choose and can't see would be a hard deal for any business.
There's also a business reason. TechCrunch notes that banks see agents as a competitive threat, since tools that help customers avoid fees or move deposits cut into bank margins. Citi's terms exempt agents that Citi provides. Expect banks to offer their own AI tools, and their own approved connections, while warning customers off everyone else's.
For an office, the practical answer is the same either way. Let AI read and prepare. Let people approve and pay.
Common questions
What did Citi change about AI agents?
In November 2026 Citi added Section H to its Online Banking User Agreement. It cautions against using third-party AI agents and says that if you give one your login, Citi will treat the agent's access and actions, and any use of leaked credentials, as authorized by you.
Which banks have changed their terms on AI agents?
TechCrunch reported on Nov. 18, 2026 that Citi and PNC changed their terms and Citizens emailed customers a similar warning, after JPMorgan and Bank of America announced similar policies in October. American Banker reported that Wells Fargo asks customers to prove they revoked agent access.
Do these rules apply to business accounts?
The changes reported are to consumer online banking terms. Business accounts already have less protection: Regulation E covers accounts held mainly for personal, family or household purposes, and wire transfers fall under UCC Article 4A, where a payment order accepted under an agreed security procedure can bind the customer. Check your own treasury agreement with your bank.
Can we let AI read our bank data?
Use a read-only connection your bank or accounting system supports, never a password. Anthropic's consumer Personal Finance feature, for example, connects through Plaid with read-only access and doesn't see the bank password.
Can an AI agent pay our vendors?
Let it prepare a payment run, but have a person review and release payments in the bank portal, with dual approval above your limit.
What should we do this week?
List every tool that touches a financial account, remove any AI tool's access to bank logins, separate banking browsers from AI extensions, and record when access is revoked.
Sources
- TechCrunch, More banks say customers using AI agents are on their own, even if they get hacked (Tim Fernholz, Nov. 18, 2026)
- Citi, Citi Online Banking User Agreement, Section H (updated November 2026)
- Banking Dive, Citi takes cautious approach on AI agent access (Oct. 23, 2026)
- Banking Dive, PNC CEO: Agent-only transactions 'will never happen' (Oct. 21, 2026)
- American Banker, Wells Fargo wants customers to prove they've revoked AI agents (Nov. 2, 2026)
- Cornell LII, 12 CFR 1005.2 (Regulation E definitions)
- Cornell LII, UCC 4A-202, Authorized and verified payment orders
- Anthropic, Introducing Personal Finance in Claude (Nov. 5, 2026)
Microsoft's Copilot Autopilot gets its own login. What to decide before it reaches your office
Vendor bank-detail changes: a callback process your AP team can actually follow
Giving an AI agent its own login: what access it should and shouldn't have