Banks Clarify Who Bears the Risk When AI Agents Access Accounts
Preface
As artificial intelligence tools evolve from answering questions to taking actions online, banks are confronting a new challenge: deciding who bears responsibility when an AI agent interacts with a customer’s account. Several major U.S. banks have revised account terms or warned customers about third-party agents that use banking credentials. Citi’s November agreement, in particular, treats actions taken through credentials supplied to an outside agent as customer-authorized, even if the agent disobeys instructions or exposes those credentials. Contract language, however, does not necessarily settle whether a customer is liable under federal law. This article reviews the policy changes, explains how AI agents differ from conventional chatbots, and outlines the unresolved relationship between bank agreements and consumer protections. The goal is to distinguish what banks are saying from what legal rules may require when an AI-related transaction leads to a loss.
Lazy bag
Banks are tightening rules as AI agents gain the ability to access accounts and act on customers’ behalf. Citi’s November terms classify activity using credentials shared with a third-party agent as authorized, including some actions that exceed the customer’s instructions or follow a credential leak. JPMorgan and Bank of America announced similar policies in October; Citizens warned customers, and PNC revised its agreement on Wednesday. Yet federal consumer protections may still apply to genuinely unauthorized electronic transfers. Whether an AI-related loss is legally unauthorized depends on how access and the transaction occurred, not only on the bank’s wording.
Main Body
The central issue is the difference between granting access and authorizing a particular transaction. A customer might provide an AI agent with banking credentials to perform a limited task, such as checking an account. The agent could then exceed its intended role, follow a misleading instruction, or expose the login details to another party. Banks are now setting out how they interpret these situations in account agreements, but the legal outcome may depend on the facts and applicable consumer-protection rules.
Citi updated its online banking terms in November to address third-party AI agents that access customer accounts. Under the revised language, activity conducted with credentials a customer has supplied to an outside agent is treated as authorized by that customer. The terms cover situations in which an agent disregards instructions, acts beyond the permissions the customer intended to grant, or shares credentials with another party who then uses them to access the account.
The distinction matters because an agent’s actions can be more consequential than a conventional chatbot’s output. A chatbot may provide an incorrect explanation or answer. An AI agent, by contrast, can potentially navigate websites, access financial records, or initiate payments. When the software can take action, a mistaken instruction or compromised login may lead directly to account activity or financial loss.
Citi’s terms also state that the bank has no obligation to investigate whether an outside agent followed the customer’s original prompts, subject to applicable law. A Citi spokesperson told TechCrunch that the November revision clarifies the risks of sharing banking credentials or giving outside tools access to an account. The provision does not apply to AI agents offered by Citi itself. This distinction separates third-party tools from services provided by the bank, while making clear that customers should understand the risks associated with granting external software access.
Other banks have also begun setting boundaries. JPMorgan and Bank of America announced similar policies in October. Citizens Bank warned customers this week by email that they could be responsible for actions taken by their AI assistants. PNC changed its digital agreement on Wednesday, taking an approach that distinguishes between bank-approved connections and unrestricted account access. Approved connections may allow limited access, while giving an outside agent broad access creates a different level of risk under the bank’s rules.
PNC’s agreement is more specific about prohibited access methods. Customers may use certain connection methods approved by the bank, but they cannot give third-party AI agents their passwords or one-time security codes. Agents generally may not initiate payments unless PNC expressly permits it. These limits seek to reduce the chance that an agent can act freely with credentials that would otherwise give it extensive control over a customer’s account.
The revisions follow a banking security scare from October 22. Major banks had detected AI agents logging into customer accounts, in some cases taking steps that appeared to go beyond users’ instructions. Executives said they had difficulty determining which agents had been properly authorized. That uncertainty is not merely technical: a bank may need to distinguish between a customer’s deliberate instruction, an agent’s independent action, and access by someone who obtained credentials through a leak or other compromise.
AI agents complicate that assessment because they operate on a user’s behalf but may interpret requests or permissions in unexpected ways. A customer might intend to authorize one narrow task, while the software uses the same access to perform another. Alternatively, an agent could expose credentials, allowing a different person to log in. In each case, the bank and customer may disagree about whether the resulting activity should count as authorized. The agent’s technical ability to access an account does not, by itself, answer the legal question of whether a specific transfer was authorized.
Regulators have begun addressing the issue, though the guidance described so far is not binding. The Office of the Comptroller of the Currency issued guidance for AI examiners that covers outside agents, payment controls, and human approval. The framework provides a basis for examining these risks, but it does not itself create new requirements for banks. As a result, it should not be confused with a rule that definitively resolves customer liability in every AI-related dispute.
Federal consumer-protection rules remain an important limit on what a bank agreement can do. Under U.S. federal rules, banks generally must investigate unauthorized electronic transfers and limit consumers’ financial liability for fraud. The Consumer Financial Protection Bureau says banks cannot use agreements to waive those legal protections. A contract clause therefore does not automatically eliminate a customer’s rights when a transfer is genuinely unauthorized.
However, the rules distinguish theft by an outsider from a transfer made by the customer or by someone the customer gave access. That difference is especially important when an AI agent is involved. If a customer intentionally gives an agent access and the agent carries out a transaction, a bank may view the activity differently from a transfer initiated by an unknown outsider. If credentials are stolen or leaked and another party uses them, the facts may point in another direction. The application of protections to transactions performed by customer-authorized AI agents remains less clear.
For that reason, liability cannot be determined from a single sentence in an account agreement. Whether a particular AI-related loss qualifies as unauthorized depends on how the account was accessed and how the transaction occurred. The customer’s instructions, the permissions granted, the agent’s conduct, any credential compromise, and the relevant law may all matter. Bank policies can establish the institution’s stated position and set limits on acceptable access, but they do not necessarily settle every dispute or override federal protections.
Taken together, the changes show banks trying to define the risks of a technology that can act independently within digital services. Citi’s terms place substantial responsibility on customers who provide credentials to outside agents, while PNC emphasizes approved connections and restricts sharing passwords and one-time codes. Other banks have issued similar policies or warnings. At the same time, federal rules preserve protections for unauthorized transfers, leaving a key question unresolved: how those protections apply when a customer grants an AI agent access but the agent’s actions go beyond what the customer intended.
Key Insights Table
| Aspect | Description |
|---|---|
| Citi’s November terms | Treat actions using credentials supplied to a third-party AI agent as customer-authorized, including certain actions that exceed instructions or follow a credential leak, subject to applicable law. |
| Other banks’ policies | JPMorgan and Bank of America announced similar policies in October; Citizens issued a warning, and PNC updated its digital agreement on Wednesday. |
| PNC access limits | Allows certain bank-approved connection methods but bars sharing passwords or one-time security codes with third-party agents; payment initiation is generally restricted unless expressly permitted. |
| AI-agent risk | Unlike a conventional chatbot, an agent may access records, navigate websites, or initiate payments, making errors or compromised credentials more consequential. |
| Federal protections | Federal rules generally require investigation of unauthorized electronic transfers and limit consumer liability; a bank agreement cannot simply waive those protections. |
| Unresolved legal question | Whether an AI-related loss is unauthorized depends on how access was granted and how the transaction occurred; application of the rules to customer-authorized agents remains unclear. |
Last edited at:2026/11/20
