AI Agent Liability: What Citi and Wells Fargo's New Bank Terms Mean for Finance and IT
Citi and Wells Fargo have rewritten their online banking terms so that customers carry the liability for AI agents, and other large US banks are moving the same way. Citi's version even covers credentials that leak from the agent. This guide explains what the clauses say and where Regulation E may push back. It also sets out a control checklist for finance and IT teams to put in place before any agent touches a company account.
Over a few weeks, several of the biggest US banks rewrote their online banking terms and pushed AI agent liability toward the customer. Citibank, the third-largest commercial bank in the US, changed its terms so that customers who use third-party AI tools for personal finance carry the consequences when something goes wrong, even if they get hacked. Wells Fargo's upcoming online access agreement takes a similar line, treating transactions by a delegate's AI agents as authorized by the customer, including ones the customer did not want or intend. JPMorgan, Bank of America and PNC have moved the same way.
The practical burden lands on finance and IT. If someone on your team has connected a bookkeeping tool, AI assistant or browser automation to a bank login, the new language decides who has to watch that tool and who pays when it misbehaves.
Coverage of Citi's change focuses on personal finance. Whether any of this binds your company accounts depends on the agreements your business signed, as the section on business and non-US accounts below explains.
What Citi actually changed in its online terms
Citi added a section on third-party AI agents to its Online User Agreement. It opens by warning customers not to give their login credentials to agents, then explains what happens if they do anyway.
Agent actions and leaked credentials
The central clause applies when a customer gives their Online Credentials to a third-party AI agent, or otherwise lets it access accounts or transact on their behalf. In that case, Citi will treat any access and actions taken by the agent as authorized by the customer. That wording is broad enough to cover the agent's mistakes too, including misread instructions.
The clause also reaches credentials that escape the agent. If the credentials the customer gave the agent are further shared, leaked or otherwise passed to other people or AI agents, Citi will treat any access and actions taken with them as authorized by the customer, which turns a breach at the agent vendor into the customer's problem.
Monitoring and privacy
Citi's terms say the customer, not Citi, is responsible for monitoring the agent's activities to make sure they are accurate, complete and within the authority the customer gave. Someone has to own that review, and they need logs to do it.
The agreement also states that the Citi Privacy Notice does not apply to the third-party AI agent. Citi does not control how the agent collects, uses, stores, protects or shares the customer's data, so once data reaches the agent, the vendor's own terms are what protect it.
Funds transfers and how Citi frames the change
As Unite.AI reports, Section 4 of Citi's Online Funds Transfers agreement makes customers solely responsible for instructions an agent gives Citi on their behalf. Agent-initiated transfers are deemed initiated and directly authorized by the customer without further investigation by Citi.
A Citi spokesperson told TechCrunch that the bank updated its Online User Agreement to give greater clarity about the risks and to reinforce that customers who grant account access to third-party AI agents are responsible for what those agents do. Calling it clarity undersells the change, since the agent section is new liability language in the agreement rather than a restatement of an existing rule.
PNC, Wells Fargo and the wider bank pattern
According to TechCrunch, JPMorgan and Bank of America, the two largest US banks, announced similar policies the previous month. PNC, the eighth-largest, changed its terms on the day TechCrunch reported the Citi update. None of the coverage we reviewed quotes PNC's clause, so its exact scope is worth pulling from PNC's own agreement.
The upcoming version of the Wells Fargo Online Access Agreement applies the same logic to delegated access. It states that all transactions an Authorized Representative or Delegate performs, including those of their AI Agents and including those you did not want or intend, are transactions authorized by you.
At the same time, the largest institutions are keeping outside agents out. As PYMNTS reported, JPMorganChase CEO Jamie Dimon said outside agents would not get inside the bank, and Citigroup said it does not let third-party agents log into accounts or transact. Charles Schwab CEO Rick Wurster said the brokerage blocks unauthorized AI agents a double-digit number of times every week. Bank of America CEO Brian Moynihan added that a human should stay in the loop to approve money moving out of an account.
PNC CEO William Demchak was more candid: for agent transactions to happen inside accounts, he said, laws on liability shifting would have to be passed.
Why would a bank that refuses agent logins also write contract terms for them? Our interpretation, which no bank has stated, is that the new terms fill Demchak's legal gap in the meantime. Citi does not offer agent logins as a matter of policy, and its contract is the fallback that says who takes the loss when a customer shares credentials anyway.
So the line banks are drawing is around agents that log in with customer credentials, and that risk stays with the customer until the law changes or approved access channels appear.
The legal logic, and where it may not hold
The authority doctrine
The banks' position has a footing in existing rules. The CFPB's Regulation E definitions commentary deals with a consumer who gives an access device to someone else, such as a family member or co-worker, and authorizes them to make transfers. If that person exceeds the authority given, the consumer is fully liable for the transfers unless they have told the financial institution that transfers by that person are no longer authorized.
The banks' clauses appear to assume that an AI agent holding your login is no different from a co-worker holding your card. If that analogy holds, revoking an agent should include written notice to the bank as well as deleting the app.
The credential theft problem
The leaked-credentials clause stands on weaker ground. The CFPB's Electronic Fund Transfers FAQs address account access information obtained from a third party through fraud such as computer hacking. When a hacker uses that information to make an EFT, the transfer is an unauthorized EFT under Regulation E. That is hard to square with a clause treating any use of leaked agent credentials as customer-authorized.
The distinction comes down to who acted. An agent that misreads an instruction or goes beyond its scope fits the authority doctrine, because the customer chose to hand it access. A criminal who steals credentials from the agent vendor looks much more like the CFPB's hacking case, where the customer authorized nothing at all.
The anti-waiver rule and the consumer caps
The Electronic Fund Transfer Act (EFTA) also includes an anti-waiver provision: no agreement between a consumer and another person may contain a provision that waives a right or cause of action the Act creates.
The consumer liability caps often get summarized briefly. Tech Times, for example, described them as $50 if a fraudulent transfer is reported within two business days and $500 if reported within 60 days. Wells Fargo's own agreement restates the first tier as the lesser of $50 or the total unauthorized EFTs that occur before notice to the bank.
Summaries like these are a useful starting point. We suggest reading the rule itself with counsel rather than relying on press summaries, and treating statement review as time-sensitive, because late notice can widen exposure.
A clause that relabels a hacked transfer as authorized could be challenged as an attempt to get around these protections, though nobody can yet say how such a challenge would end.
Untested, with a regulatory gap
According to Startup Fortune, no court or regulator appears to have tested this reading, and no US agency has publicly commented. Regulators have not filled the gap either. As Tech Times reported, the Federal Reserve, OCC and FDIC issued revised Model Risk Management Guidance in April 2026, and it explicitly excluded generative and agentic AI from its scope.
Business and non-US accounts
We suggest not assuming that the consumer protections discussed here carry over to company accounts. Those accounts may fall under different rules and under the commercial agreements your business signed, so confirm the applicable law with counsel before relying on any of the protections above.
Outside the US, teams should expect local rules on third-party access and liability to differ. Those local rules, not the US positions described here, are the ones to check, ideally with local counsel.
A control checklist for AI agent liability on company accounts
The bank terms hand monitoring and authority decisions to the customer, so on company accounts that duty needs a named owner in finance or IT.
Catalogue agents and limit their access
Start by listing every AI agent, aggregator, browser automation and script with bank access. Record the owner, scope, credential type (shared password, token, API key) and the accounts each one can reach, because you cannot revoke or monitor something nobody wrote down.
Default every entry to read-only access through bank-supported APIs and leave payment initiation to people. Most reporting, reconciliation and forecasting work needs data, and very little of it needs the power to move money.
Replace passwords and rehearse the off switch
Wherever a scoped, revocable connection exists, use it instead of any tool that asks for your bank password. Startup Fortune notes that startups offering scoped, revocable access through an open-banking API, with no passwords handed over, are in a stronger position than ones that log in as the user. For agents your own developers build, we suggest routing secrets through a credential proxy so the agent never holds raw API keys; our credential proxy playbook for AI agents walks through one approach.
Cutting access needs its own written procedure: rotate credentials, revoke tokens, and notify the bank in writing that the agent is no longer authorized. That written notice is what cuts off your liability for later transfers by that agent. Practice until the drill takes minutes; our AI kill switch and agent runbook guide covers how to structure the stop procedure.
Reconcile agent activity
Give a named person the job of reconciling agent activity against its expected scope. This is the duty Citi's terms place on the customer. How often is a risk call, but for any connection that can move money, we recommend a daily review.
Where Regulation E applies, the liability tiers described above reward fast notice. Under the CFPB commentary, you also stay liable for an agent's overreach until you tell the bank, so slower reviews let exposure build before anyone can act. For read-only connections, weekly may be proportionate. On company accounts, use the notice deadlines in your commercial agreement instead.
For agents built in-house, we suggest tracing each tool call so the reviewer has a record to work from; our guide to GenAI tracing for agent observability covers one way to set that up. For vendor tools, you depend on the activity logs you require the vendor to provide, which is one more reason to ask for them before signing.
Keep people on payments and recheck the terms
Require a human to approve any outbound money movement, and set bank alerts for new payees and unusual amounts. This matches the human-in-the-loop principle Bank of America's CEO described publicly, and it gives you a checkpoint that does not depend on the agent behaving as intended.
Recheck both the agent inventory and your bank terms whenever a bank updates its agreements, as Citi and Wells Fargo have just done, and on a fixed schedule in between. A new clause or a forgotten connection should not sit unnoticed for months.
Questions to ask fintech and AI agent vendors
Get answers to these before you connect any company account:
- Do you connect through an open-banking or bank-approved API with scoped tokens, or do you log in with our credentials?
- Can access be limited to read-only, and can we revoke it ourselves, immediately?
- How are credentials or tokens stored, and what happens if they leak? Under Citi's clause, a leak at the vendor is your loss.
- Which privacy terms govern our data once it reaches you?
- What activity logs do you provide, so we can carry out the monitoring duty?
- What liability do you accept if your agent exceeds its scope or is compromised, and is it capped?
Vague answers to the first three are a good reason to keep the tool away from any account that can move money.
Caveats and open questions
- This is an operational reading of published terms, not legal advice. Take the account-type and jurisdiction points from the business and non-US section to counsel.
- Enforceability is untested. The leaked-credentials clauses may not survive a dispute involving genuine credential theft.
- Terms differ by bank, product and version date. Wells Fargo's updated agreement is published as an upcoming version, so check its effective date.
- Regulators have not addressed agentic AI directly, and banks may later offer approved agent access routes. Either would change these trade-offs.
Key Takeaways
- Citi's updated Online User Agreement treats all actions by a customer-permitted third-party AI agent as authorized by the customer, including actions taken with credentials that leak from the agent.
- Citi puts the duty to monitor agent activity on the customer, and its privacy notice does not cover data held by the agent.
- Wells Fargo treats AI agent transactions by delegates as customer-authorized even when unwanted. PNC has also changed its terms, but its clause has not been quoted in coverage, so read it directly.
- The banks' position draws on Regulation E's authority doctrine, but the CFPB treats hacker transfers using stolen access information as unauthorized, and the EFTA bars waivers of its rights.
- The clauses are untested, and the April 2026 Model Risk Management Guidance excluded generative and agentic AI, so regulators have not yet filled the gap.
- Before any agent touches a company account: catalogue it, limit its access, replace shared passwords, rehearse revocation, reconcile its activity, keep humans on outbound payments and recheck terms regularly.
Conclusion
Banks are moving toward a model where the customer authorizes the agent, watches it and absorbs the loss when it goes wrong. Citi and Wells Fargo have written that into their terms, and Citi's version reaches as far as credentials that leak from the agent. Part of that rests on settled Regulation E doctrine about people who exceed their authority. The leaked-credentials language is harder to defend against CFPB guidance on hacking and the EFTA anti-waiver rule, and no court has ruled on it yet. In the meantime, any team that keeps an agent connected to a bank account needs someone who knows what that agent can do and how to switch it off.
Two things are worth doing now. First, ask whoever owns your finance systems to list every tool that holds a bank login or token, and flag any that can move money. Second, send the vendors behind the flagged tools the questions above, ask for written answers, and have counsel review how your own account agreements apply.
If it would help, Optijara can work with your team to map which tools currently hold bank access and draft the revocation runbook.
Frequently Asked Questions
Am I liable if a third-party AI agent connected to my Citi account gets hacked?
Under Citi's updated Online User Agreement, if credentials you gave an AI agent are further shared, leaked or distributed to other people or agents, Citi will treat activity using those credentials as authorized by you. The CFPB, however, says that transfers made by a hacker using access information stolen from a third party are unauthorized EFTs under Regulation E, and the EFTA limits waivers of consumer rights. No court has tested how these interact, so the contract may not be the final word. If you face a real loss, speak to counsel.
Does Regulation E still protect me if an AI agent makes an unauthorized transfer?
It depends on why the transfer happened. If you gave the agent access and it exceeded the authority you gave, Regulation E commentary treats that like a family member or co-worker exceeding authority, which leaves the consumer fully liable until the bank is notified. If an outsider stole credentials and moved money, the CFPB treats that as an unauthorized EFT, and the consumer liability caps come into play. Tech Times summarizes those caps as $50 if you report within two business days and $500 within 60 days. We suggest confirming the details with counsel rather than relying on press summaries, and reporting suspicious transfers as quickly as you can. As the article explains, these consumer rules may not apply to business accounts.
What is the difference between an agent exceeding its authority and credential theft?
Exceeding authority means the agent you knowingly connected does something you did not intend, such as paying the wrong amount. Regulation E treats that much like a trusted person misusing access you gave them. Credential theft means someone else gets hold of the access information, for example by hacking the agent vendor, and uses it.
Are read-only, open-banking API connections treated differently from sharing my login with an agent?
Citi's clauses are triggered when you give your Online Credentials to an agent or otherwise let it access accounts and transact. Scoped, revocable API connections that do not involve handing over passwords reduce the exposure those clauses describe, and analysts note that vendors offering this kind of access are in a stronger position. Still, check each bank's terms and each connection's actual permissions, because read-only scope is what limits your risk.
How do I revoke an AI agent's access to my bank account quickly?
Deleting the app is not enough. Rotate any credentials the agent held, revoke its tokens or API connections, and notify the bank in writing that the agent is no longer authorized. Regulation E commentary keeps the account holder liable for a trusted party's overreach until the bank has been told that party is no longer authorized, so the written notice matters for later transfers. Rehearse the steps ahead of time so the whole thing takes minutes.
Does this apply to accounts outside the US?
The bank terms and regulations discussed here are from the US. If your accounts are held elsewhere, we suggest checking how local rules treat third-party access and liability, and confirming the position with local counsel rather than assuming the US approach carries over.
Sources
- https://www.citi.com/tandc/view-printable-version
- https://www.wellsfargo.com/online-banking/online-access-agreement/upcoming/
- https://techcrunch.com/2026/11/18/more-banks-says-customers-using-ai-agents-are-on-their-own-even-if-they-get-hacked/
- https://www.hunton.com/insights/legal/major-banks-move-to-block-or-limit-customer-directed-ai-agents
- https://natlawreview.com/organization/hunton-andrews-kurth-llp
- https://www.consumerfinance.gov/rules-policy/regulations/1005/2/
- https://www.consumerfinance.gov/compliance/compliance-resources/deposit-accounts-resources/electronic-fund-transfers/electronic-fund-transfers-faqs/
- https://www.pymnts.com/news/banking/2026/banks-crack-down-on-unauthorized-ai-agents-operating-inside-customer-accounts/
- https://www.unite.ai/citi-makes-customers-responsible-for-third-party-ai-agent-activity/
- https://startupfortune.com/citi-pnc-and-citizens-tell-customers-they-own-any-loss-an-ai-agent-causes/
- https://www.techtimes.com/articles/329947/20261103/wells-fargo-just-made-you-legally-responsible-every-transaction-your-ai-agent-makes.htm



