News

AI banking agents and liability: what changed, what rights remain

AIR
November 22, 2026

Several US banks are restricting how outside AI agents access customer accounts and assigning customers responsibility for agent activity. The changes are real. The claim that giving an agent access automatically removes every right to recover a loss is too broad: contractual terms, the authority actually granted and applicable consumer-protection law are separate questions.

As of November 22, 2026, the banks have also taken different steps on different dates. Some provisions are current; others take effect later. For businesses building agent-led commerce, the practical lesson is to make identity, data access and permission to move money independently understandable and enforceable.

AI-generated photorealistic scene of an adult reviewing an AI account-data access request on a laptop alongside third-party AI terms; generic illustrative interface, not a real bank or AIR product screen.
Cover: AI-generated illustrative scene with a conceptual interface.

What the bank documents actually say

The strongest evidence comes from the agreements themselves. A social post can summarize the direction of travel while missing a future effective date or presenting a bank's contractual position as the final legal outcome.

BankDocumented changeStatus on November 22, 2026
CitiNew Section H addresses third-party agents, customer responsibility and access using subsequently leaked credentials.The agreement identifies a November 2026 update, without an exact day.
PNCUpdated agreement restricts credentials, connection methods and agent transactions, with responsibility provisions subject to law.The available updated PDF is effective December 19, 2026.
JPMorgan ChaseCredential-sharing restrictions expressly include outside AI; approved application programming interfaces (APIs) remain a route for authorized account-data access.The current document dates these restrictions to October 27, 2026.
Wells FargoUpcoming Section 15 covers reliance on agent instructions, customer responsibilities and withdrawal of authority.The upcoming agreement takes effect November 24, 2026.

Sources: Citi Online User Agreement, PNC updated Digital Services Agreement, Chase Digital Services Agreement, and Wells Fargo upcoming agreement.

Citi's Section H is particularly explicit. It treats customer-enabled agent activity as authorized and extends that treatment to access using credentials subsequently leaked or passed on by the agent. It addresses incorrect instructions, unintended actions and actions beyond the authority the customer intended. The responsibility clause also includes the qualifier “except as otherwise permitted under applicable law.” These are the bank's stated terms, not proof that every disputed loss will have the same outcome.

PNC's forthcoming wording emphasizes access controls: agents must use methods PNC permits, customers must not give them passwords or one-time codes, and agent transactions require express permission for the relevant service. Chase's cover separates its October credential rules from a January 2027 change concerning mobile-app and trusted-device identity checks. Combining those dates would misstate the timeline.

Citizens is another distinction. TechCrunch reported on November 18 that it sent customers a warning email that week. Its public digital agreement contains general third-party credential-sharing provisions. That supports describing a warning alongside existing terms; it does not establish a newly dated AI-specific amendment.

Why “no refunds” misses the legal distinction

US consumer electronic-transfer rules distinguish someone acting with furnished access and authority from someone obtaining access through fraud or theft. That difference matters more than whether software appears somewhere in the chain.

The official interpretation of Regulation E, section 1005.2, addresses a consumer who supplies an access device and grants another person authority to make transfers. It states that the consumer can be liable when that person exceeds the authority, until the bank is notified that further transfers are no longer authorized. This is an existing delegation rule, not a new AI-specific finding.

The Consumer Financial Protection Bureau (CFPB) electronic-transfer FAQs separately discuss credentials stolen from a legitimate third party and information obtained through fraud. Those circumstances can produce unauthorized transfers. The guidance also says negligence cannot justify liability beyond Regulation E's limits and rejects contractual waivers of the protections. A hacker stealing information from a service and a service executing an overbroad instruction therefore should not be collapsed into one scenario.

Electronic Fund Transfer Act (EFTA) anti-waiver provision prevents agreements from simply removing rights created by the law. Determining how these rules apply to a particular agent transaction still requires its facts: who initiated it, what authority existed, how access was obtained and whether authority had been withdrawn.

These protections concern covered transfers involving consumer accounts. They should not be generalized to every business account, wire transfer, brokerage trade or credit-card purchase. Nor does calling a transfer unauthorized establish automatic full reimbursement: Regulation E's liability provisions include conditions and reporting requirements. Product copy should promise neither universal recovery nor universal loss of protection.

A bank-provided tool has a different contractual context

Citi expressly excludes tools and features it provides through Citi Online or Citi Mobile from the new Section H. That is a defined exclusion from this section. It is not a statement that every bank-operated AI tool carries unlimited protection, or that other providers cannot offer a bank-approved connection.

For a commerce platform, the useful procurement question is which access method the bank supports and what its agreement covers. A customer consenting inside an agent application does not establish that the bank accepts the application's method of access. A functioning login is similarly weak evidence of a permitted transaction route.

The commercial issue is where permission becomes enforceable

A shopping request can contain several intentions at once: compare products, apply a membership benefit and place an order if the price is acceptable. Each receiving business needs to know which part it is being asked to perform.

Consider the failure case. An agent finds a suitable product but changes the payment source to complete the purchase. The merchant may see a valid customer identity while the payment provider sees a technically valid request. Neither observation, alone, shows that the customer approved changing funding sources.

The operating design should preserve those distinctions before a dispute. Product, payments and support teams need a shared record of the offer, the approval and the resulting order. This is why agent identity and authorization deserve separate decisions rather than a single reassuring badge.

Where AIR fits, and where the payment provider must act

AIR's technical commerce architecture, reviewed on November 22, states that current AIR Agent delegation concerns identity. It documents a user-bound agent connection, fixed identity-related scopes, short-lived sessions, credential verification and a checkout identity token for a supported merchant audience. The documented checkout handoff currently supports Pivota.

That identity binding does not grant wallet-signing or payment authority. The documentation also distinguishes unreleased fine-grained, per-task delegation and wallet delegation from the current capability. Spending limits must be enforced by the payment integration; an identity token does not enforce a budget.

The commercial value is customer recognition. A business can evaluate consented credential evidence to determine whether an arriving customer qualifies for an offer. AIR Identity belongs in that decision. Bank-approved access, purchase approval and payment controls require their own supported integrations. No part of this architecture establishes a partnership with the banks above or determines who bears a disputed loss.

An illustrative purchase flow with clear boundaries

Suppose a retailer offers a benefit to members of a participating professional community. A customer asks an agent to compare desk equipment and prepare an order. The following is a proposed operating design, not a claim that all its components are available through AIR.

First, the identity integration establishes the agent's connection to the customer. The merchant then evaluates the membership evidence shared through the supported credential flow. The result answers whether the customer qualifies for the benefit. It does not permit the agent to inspect an unrelated bank account or move money.

Second, the retailer produces the exact basket, delivery terms and total. The chosen payment integration obtains the required approval and enforces any supported transaction restrictions. If the total or funding source changes, the flow follows that integration's rules for new approval. An instruction remembered by the model should not be the only place the restriction exists.

Third, the order and payment systems record their outcomes. Support can distinguish an accepted order from a failed payment, identify the approval used and see which service should handle cancellation or a dispute. Withdrawal of identity access and cancellation of an already accepted payment remain separate operations.

Before launch, test a changed basket, declined payment, expired identity session and a repeated transaction request. Our agent permissions guide describes proposed controls and their ownership. The goal is a purchase the business can explain and service, with authority enforced where the relevant action occurs.

Frequently asked questions

Are customers always liable if an AI banking agent makes a mistake?

No universal answer follows from the agreements. The applicable account and transfer rules, actual authority, access method and facts of the event matter. Some terms assign broad responsibility while preserving applicable-law limits.

Can a bank contract remove Regulation E protections?

An agreement cannot waive rights provided by EFTA. That does not make every agent-initiated transfer unauthorized: the rules distinguish furnished access and transfer authority from fraud or theft.

Did all the named banks change their terms during the same week?

No single timeline is supported. Citi identifies a November update; Chase dates relevant restrictions to October 27; Wells Fargo's upcoming terms start November 24 and PNC's updated PDF starts December 19. Citizens' email was reported separately.

Does using a bank's own AI guarantee reimbursement?

No. Citi's exclusion applies to a particular section of its agreement. The relevant service terms and applicable law still determine the treatment of an error or disputed transaction.

Does AIR enforce an agent's bank-account spending limit?

The current AIR Agent identity APIs do not enforce spending caps or grant payment authority. AIR supports customer-agent identity and credential checks; the payment integration must separately approve and enforce the transaction.

Partner with AIR

Discuss how customer recognition and credential-based eligibility can fit your agent commerce flow, alongside the payment controls your integration requires. Talk to AIR.

Moca Network, the identity network of Animoca Brands.

Partner with AIR

AIR is one integration suite for identity, money and loyalty. AIR Identity lets trusted issuers issue verifiable credentials and businesses verify them, so acquisition starts with users who already qualify, with minimal data custody. AIR Money embeds compliant global financial infrastructure, opening new markets without building new rails. AIR Loyalty gives brands a programmable, stable-value points engine that keeps users coming back.

Partner with us to grow your business.

AIR is built by Moca Network, the identity network of Animoca Brands.

Stay updated on AIR launches
Product updates, partner launches, and research across digital identity, fintech, and loyalty. Unsubscribe anytime.
By subscribing, you agree to our Privacy Policy and consent to receive updates.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
In this article
Blog

Read more articles

View all
AI-generated photorealistic scene of two merchants scanning and weighing one test parcel beside a limited batch and a physical pilot test-order card.
News
How to launch an agentic commerce pilot that earns its next phase
Plan an agentic commerce pilot with a focused customer task, four launch gates and a commercial case based on contribution, support and returns.
AI-generated editorial illustration of a white shopping cart suspended above a silver coin stack on a white plinth.
News
How to measure agentic commerce: a merchant KPI framework
Measure agentic commerce from AI visibility to accepted orders and retained contribution, with clear denominators and an illustrative merchant funnel.
AI-generated photorealistic scene of a customer handing a labeled return parcel to a courier at home, with an illustrative AI return-ready confirmation on a nearby phone.
News
Returns and refunds in agentic commerce: who handles what?
A merchant guide to agent-assisted returns: customer authority, partial refunds, order evidence, support ownership and escalation.