top of page
davydov consulting logo

DAVYDOV CONSULTING BLOG

HOME  >  NEWS  >  POST

AI Agent Liability Is Now Your Problem: What the Banks' New Terms Mean for Your Business

3 minutes ago
6 min read

This week three more large US banks quietly changed the rules for customers who use AI agents. Citi added a whole new section to its online user agreement, PNC updated its digital banking terms on Wednesday, and Citizens emailed its customers a warning. The message is roughly the same in all three cases: AI agent liability now sits with you, the account holder, for anything an agent does after you let it in.

That includes the ugly scenarios. Under Citi's new terms, if an agent leaks your credentials or gets hacked and a criminal moves your money, the bank still treats those transactions as authorised by you. JPMorgan and Bank of America made similar moves back in October, so this is quickly becoming the default position of the biggest banks in the country.

For business owners this is more than consumer news. Agentic payments, AI bookkeepers and autonomous procurement tools are being pitched to companies everywhere right now, and the liability question behind all of them has just been answered in the least pleasant way possible. Let's sink in!

Business person checking a banking app on a smartphone in front of a glass bank tower

Who changed what, and when

Citi's change is the most detailed one. During the week of 16 November it added a new Section H to its Online User Agreement, covering anyone who gives a third-party AI agent access to their accounts through Citi Online or Citi Mobile. The section says that actions taken with credentials you handed to an agent are treated as authorised by you, even if the tool ignores your instructions, goes beyond the permissions you meant to give it, or passes your credentials to someone else. A separate November update to Citi's funds-transfer agreement adds that any transfer an agent starts counts as authorised even when it differs from what you actually told the agent to do, and that Citi has no obligation to investigate whether the agent followed your original prompts, subject to applicable law.

PNC moved on 18 November. Its updated digital agreement is less about liability and more about access: customers cannot give third-party AI agents their passwords or one-time security codes, and agents generally cannot start payments unless PNC expressly allows it. The bank draws a line between approved connections with limited access and handing a tool the keys to everything.

Citizens took the softer route and emailed customers that they could be responsible for what their AI assistant does. And this is on top of October, when Bank of America told customers on 27 October not to share banking passwords with AI assistants, and JPMorgan put the consequences of credential sharing on the customer while also, according to a CNBC report, blocking some outside agents from reaching accounts directly. One report also says Wells Fargo has a new Online Access Agreement taking effect on 24 November that treats every transaction made by your delegates, AI agents included, as authorised, including those you did not want or intend.


The fine print that actually matters

The headline is scary enough, but the details are worth reading slowly. In Citi's version, monitoring the agent is your job. If the tool misreads an instruction, buys the wrong thing or exceeds the authority you thought you gave it, that mistake is yours except where the law says otherwise. Citi can also cut off any third-party agent's access at any time, without notice.

There is a privacy angle too. By granting an agent access, you agree that Citi may share your personal information, including sensitive personal information, with that outside tool. Citi's own privacy notice does not cover the agent, and the bank says it cannot control how the tool stores or shares your data. So the agreement is not only about who pays when something goes wrong, but also about what happens to your data on the way.

A Citi spokesperson told reporters the change was meant to provide clarity about the risks, and that customers who grant agents access are responsible for activities conducted through those agents. Notably, the new section does not apply to AI tools Citi builds and offers itself. The bank's own agents stay inside the bank's own responsibility, which tells you something about where this is going: banks want AI in banking, as long as it is their AI.


Close-up of hands going through the fine print of a bank agreement before signing

Why banks are drawing this line now

The timing is not random. In late October several major banks reportedly found third-party agents logging into customer accounts in ways that appeared to go beyond what users had asked for. To a bank's fraud systems, an agent armed with your real username and password looks exactly like an account takeover, because technically it is one, just a consensual one. The banks' answer is to say: fine, but if it goes wrong, that was your login.

There is also a quieter commercial motive. An AI agent that shops around for better savings rates, moves idle cash out of low-interest accounts or negotiates fees is working against the bank's margins. Deciding which third-party tools may connect, and on what terms, lets banks control who gets to build on top of their customer relationships.

And banks are not alone in this. Ride-hailing apps, delivery platforms and retailers have been pushing back against agents acting on users' behalf all autumn. The pattern repeats everywhere: companies want to serve AI agents through official, controlled channels, not through an agent pretending to be you with your password.


What regulators have and have not said

The legal mechanics here are genuinely interesting. Under US rules on electronic transfers, a consumer's liability for unauthorised transactions is capped, and banks normally have to investigate disputes. The new agreements work around that by definition: if everything an agent does with your credentials is authorised by contract, then there is no unauthorised transfer to dispute in the first place.

Whether that holds up is another question. Reports suggest the CFPB's position is that these protections cannot simply be waived by an agreement, and the OCC has issued examiner guidance that touches on external AI agents and payment controls, though it is described as nonbinding. No court appears to have tested any of this yet. We are not lawyers and this is not legal advice, but it is fair to say the banks have planted a flag on contested ground.

Until a regulator or a court says otherwise, though, the practical reality for customers is the terms as written. If your dispute starts from a clause that says the transfer was authorised, you are fighting uphill from day one.


AI agent liability for business accounts

Here is the part that matters most for our readers. Business accounts generally enjoy fewer statutory protections than consumer accounts to begin with, so a liability-shifting clause bites harder when it is your company's operating account on the line. If your bookkeeper, your virtual assistant or an automation tool logs into business banking with stored credentials, you should assume that whatever happens next is legally your doing.

A few sensible moves follow from that. Read your bank's online access agreement this month, because these updates are landing quietly and yours may already have one. Make an inventory of every tool, plugin and agent that currently holds banking credentials for your company, and revoke anything you cannot justify. Prefer read-only connections for analysis and bookkeeping; it is no accident that Anthropic markets its personal finance tool as read-only, saying it cannot move, manage or invest your money.

If you do want automation that pays, isolate it. A separate account with a limited balance and no overdraft gives an agent room to be useful while capping the damage of a bad day. That is the same logic as giving a new employee a company card with a low limit rather than the corporate chequebook.


Business team around a table reviewing how their company uses AI agents and payment tools

Takeaways for teams building digital products

If you are building products in this space, the direction of travel is now very clear: credential sharing is dead as a design pattern. Any fintech feature that asks users for their bank password is asking them to accept unlimited personal liability, and banks are actively blocking that style of access anyway. Build on open-banking APIs and OAuth-style connections, where the user grants scoped, revocable permissions and never types their banking password into your product.

Design the permissions to be visible and boring. Show users exactly what the agent can see and do, keep an audit log of every action the agent takes, and make revoking access a one-click affair. When something disputed happens, that log may be the only evidence your user has, because the bank has already said it will not investigate on their behalf.

Finally, treat payment initiation as a privilege your product earns through official channels, not a trick it performs with borrowed credentials. The protocols for declaring what agents may do on a site or service are arriving fast, and the companies that integrate properly will be the ones still standing when the credential-based shortcuts get shut off. For agencies like ours building sites and custom tools for clients, that means scoping agent access per feature from day one rather than bolting it on later.


Final notes

None of this means AI agents in finance are over. It means the free-riding phase is over. The convenience now comes with a price tag and a named owner, and that owner is you unless the integration is done through channels the bank has blessed.

Our advice is simple: check your bank's terms, audit which tools hold your credentials, keep agents on read-only or sandboxed accounts, and if you are building agentic features into your own product, do it through proper APIs with scoped permissions. The businesses that take an afternoon to sort this out now will save themselves a very unpleasant conversation with their bank later.


 
 
 

Comments


​Thanks for reaching out. Some one will reach out to you shortly.

CONTACT US

bottom of page