Why Banks Can't Verify AI Agent Authority Over Customer Accounts

A bank's core systems can now confirm that a valid session, a valid token, and a valid API call all line up, and still have no idea whether a human meant for the resulting action to happen. That gap, not a model capability gap, is what two of the world's most consequential financial regulators spent the week of November 16, 2026 talking about in public.

Federal Reserve Governor Michael Barr told an audience in Milan that major banks are already encountering AI agents attempting actions in retail customer accounts without the bank being able to verify authenticity, consent, or alignment with the customer's actual instructions. He did not name the banks. He did not put a number on the incidents. He also warned that agents could accelerate deposit withdrawals during a stress event, compressing a run that once took days into a run that takes minutes. On the same day, Hong Kong's Financial Secretary Paul Chan delivered a line that deserves to be quoted in full. "An algorithm has no license to lose and no reputation to protect. The institution has both." Two days earlier, Hong Kong's Securities and Futures Commission had conceded something banks rarely hear from a regulator, that reviewing every intermediate step an agent takes is not realistic, and that guidance will have to assume it.

Put those three statements next to each other and a pattern appears that goes well beyond Hong Kong or the Federal Reserve's jurisdiction. Regulators are no longer asking whether an AI agent behaved reasonably. They are asking whether the institution can prove, for one specific action, that a real, consenting, authenticated human intended it. A session token proves a login happened. It says nothing about whether the person who logged in, four steps and two tool calls ago, meant for an agent acting on their behalf to move money, close an account, or change a beneficiary today.

This post is written for the CISO and the Chief Compliance Officer who will be asked to answer for that gap, not for the engineering team that built the agent. The question a regulator asks after an incident is rarely "was the model good." It is "show me the record that proves this specific action was authorized."

Key Takeaways

  1. Identity is not authority. A valid session or API credential proves an agent was technically able to act. It does not prove a human meant for that specific action to happen, and regulators are now treating that distinction as the actual compliance question.
  2. Liability sits with the institution, not the agent. Hong Kong's Financial Secretary made the point explicitly. An algorithm has no license and no reputation on the line. The deploying bank and its overseers carry both, regardless of how autonomous the agent was.
  3. "A human was in the loop" is no longer sufficient evidence. The Securities and Futures Commission has acknowledged that reviewing every intermediate agent step by hand is unrealistic, which means institutions need a different kind of proof than a reviewer's signature.
  4. Adoption is outrunning verification. A Hong Kong Institute for Monetary and Financial Research report found 13 percent of surveyed financial institutions have already implemented agentic AI and 56 percent plan to, well ahead of the authority and identity frameworks regulators are still drafting.
  5. Evidence has to be captured at the moment of action, not reconstructed afterward. A permission that technically allowed an agent to act is not the same record as proof that a specific human consented to that specific action, and only the second one holds up to a regulator or a court.

What Barr, Chan, and Leung Said on the Record

Start with the specifics, because the vagueness in the public record is itself informative. Barr's remarks, delivered at the Seventh Annual Federal Reserve speech on artificial intelligence in finance in Milan, described banks finding AI agents attempting actions in retail accounts that the bank could not verify as authentic or consented to. He called for strict authentication, explicit per-action consent verification, and technical controls that actually limit what an agent can do rather than merely instruct it to behave. He did not say which banks. He did not say how many incidents. For a Federal Reserve governor speaking at an international conference, that is itself a signal. The absence of a number is not an absence of a problem; it most likely reflects that banks are still working out how to count incidents they were not built to detect in the first place.

Hong Kong's response arrived from two directions at once. The Project SAGE launch on November 2, 2026 set up the Hong Kong Monetary Authority's Agentic Finance Expert Group to work through agent identity, delegated authority, interoperability standards, and liability for agentic payments. It is exploratory rather than binding, and it does not yet assign specific obligations. What it does establish is the vocabulary regulators intend to use going forward, including "Know Your Agent" practices that mirror Know Your Customer obligations banks already run.

Financial Secretary Paul Chan's remarks, covered in regulatory forum coverage of the Bloomberg Global Regulatory Forum, put the liability question to rest in the sentence quoted above. That is not a rhetorical flourish. It answers a question CISOs have been asking privately for two years, who absorbs the consequence when an autonomous agent does something wrong. Hong Kong's answer is unambiguous, and the Federal Reserve's framing points the same direction without ever using the word liability.

Securities and Futures Commission CEO Julia Leung had already previewed the hardest part of this problem on November 2, at Hong Kong FinTech Week. Reviewing every intermediate step an agent takes, she said, is not realistic. Rather than demand it, the SFC plans principles-based guidance that assumes humans cannot approve every action and asks instead for a graduated model, more autonomy for reversible, low-impact internal processes, and stronger safeguards for client advice, portfolio decisions, and market execution. That is a meaningfully different posture from "add a human checkpoint," and it is the posture every other regulator watching Hong Kong's approach is likely to converge on, because the alternative does not scale.

The Authority Gap Banks Cannot Close With Logins Alone

Here is the specific, concrete failure mode regulators are circling without quite naming it this way. When an AI agent acts on behalf of a customer or an employee, the institution needs to answer one question before the action completes. Does the agent's claimed authority to act trace back to a real, consenting, authenticated human intent for this specific action, right now?

Every one of those failure modes widens what Bonfy has called the trust boundary an institution must account for, because each connected tool, document source, and delegated permission is a new edge where authority can be claimed without being real. Four things can break that chain, and none of them show up in a standard access log. A compromised session can hand an attacker a legitimate, authenticated path into an account, and every downstream agent action will look perfectly authorized because the credentials are real. A prompt-injected instruction can arrive through a document, an email, or a connected tool the agent trusts, and the agent will treat it as if the customer asked for it. A stale permission grant, set up for one purpose months ago, can authorize an action the customer never contemplated and would not approve if asked. And an agent can simply misread ambiguous intent, interpreting "help me manage my cash position" as license to initiate a transfer nobody requested.

Standard identity and access tooling answers a narrower question than the one regulators are now asking. It confirms that a credential was valid, that a token was issued, that an API call matched an allowed pattern. None of that is the same claim as "a human meant for this." Bonfy has described this as the gap between identity and intent: identity tells a bank who acted, not why the action was requested or whether a customer wanted it. The distinction sounds academic until a bank tries to answer a regulator's follow-up question after an incident. The question is not "was the agent authorized to access this account," but "did a human consent to this specific withdrawal." Those are different questions with different evidence requirements, and most institutions today can only answer the first one.

Why "Someone Reviewed It" No Longer Counts as Proof

The instinct inside most compliance functions is to solve this with a human checkpoint. Leung's comments make clear why that instinct will not hold up at scale. An agent that completes a multi-step task touches dozens of intermediate decisions before it reaches a customer-visible action, and asking a human to review each one defeats the purpose of deploying the agent at all. Most institutions will end up with a human reviewing the final output, not the reasoning chain that produced it, which means the reviewer is approving a summary, not the decision.

That matters because an audit trail built around "a person clicked approve" collapses the moment a regulator asks what the person reviewed. Did they see the agent's reasoning? Did they know which data sources it consulted? Could they have caught a subtly wrong interpretation of the customer's intent, or were they rubber-stamping a system they had no practical way to second-guess? A reviewer's signature proves a click happened. It does not prove the click constituted meaningful oversight, and regulators across multiple jurisdictions are now explicit that a human present in the workflow is not automatically a human in control of it.

This is the same distinction Kiteworks has made in the context of human oversight compliance for other regulated workflows. Oversight that cannot be evidenced is not oversight a regulator will credit, and the Hong Kong forum made the financial-services version of that argument on the record for the first time this explicitly.

Project SAGE and the Shape of What Comes Next

Project SAGE is worth reading closely because it tells institutions what regulators will eventually ask them to produce, months before any rule requires it. The Hong Kong Monetary Authority's Expert Group is working through four areas, use case exploration across consumer, business, and machine-to-machine payments; agent identity and "Know Your Agent" verification; interoperability standards so agents from different institutions can transact safely; and legal frameworks covering liability and consumer protection.

Agent identity deserves the most attention from a CISO sitting outside Hong Kong's jurisdiction, because it reframes a technical access-control problem as a compliance obligation. Know Your Customer rules did not simply ask banks to confirm a customer exists. They asked banks to establish who the customer is, what the relationship permits, and how the bank would prove that relationship under scrutiny. Know Your Agent extends the same logic to a non-human actor, which agent is this, under whose delegated authority is it acting, and what can the bank show a regulator about that delegation chain. This connects directly to the work Kiteworks has published on non-human identity governance, which makes the same point for enterprise data access generally. An agent identity that resolves to a shared service account or a borrowed human credential cannot satisfy a Know Your Agent standard, because there is no agent there to know.

SAGE is deliberately exploratory rather than prescriptive today. That will not last. Regulators who stand up expert groups and publish frameworks for agent identity and delegated authority are previewing the shape of binding requirements, not floating an idea they might abandon.

Adoption Is Outrunning the Governance Catching Up to It

The urgency here is not theoretical. A report from the Hong Kong Institute for Monetary and Financial Research, released November 17, 2026 and covered in the HKIMR agentic AI report, found that 13 percent of surveyed financial institutions have already implemented agentic AI in some form, and 56 percent plan to. That is nearly seven in ten institutions either live or committed, against a regulatory and technical standard for verifying agent authority that both the Fed and Hong Kong's regulators described, in the same week, as still under construction.

That sequencing is the actual risk, not the technology itself. Institutions are not waiting for Know Your Agent standards, consent-verification architecture, or principles-based SFC guidance before deploying agents that touch customer accounts. They are deploying now, on the assumption that existing identity and access tooling is close enough. Barr's remarks suggest that assumption is already being tested in production at scale, inside institutions large enough to be the ones briefing a Federal Reserve governor in the first place.

Evidence Captured at the Moment of Action, Not Rebuilt Afterward

The response regulators are converging on, even without using identical language, is a shift from proving authorization existed to proving authorization was exercised for this specific action. That shift matters because most institutions' current evidence is backward-looking. A permission record shows an agent was allowed to initiate transfers generally. It does not show that a specific customer consented to this transfer, today, for this amount, to this recipient. By the time a regulator or a litigant asks the question, the only honest answer many institutions can give is "the agent was technically permitted to do this," which is not the same claim as "the customer meant for this to happen."

Barr's call for explicit, per-action consent verification and technical controls that limit agent capability describes an enforcement point that sits between the agent and the customer's data and systems, evaluating each proposed action against the specific entity, relationship, and context involved rather than against a standing permission. A customer asking an agent to summarize their statement and the same agent proposing to move funds between accounts are different events even when one broad grant of access technically covers both. Contextual Data Enforcement is the architecture for making that distinction in real time, rather than discovering it only when an auditor reconstructs the sequence weeks later.

This is also where the record itself becomes the deliverable, not a byproduct of it. An enforcement decision made at the moment an agent attempts an action, logged with the entity involved, the authority claimed, and the outcome, is evidence a regulator can use. A quarterly access review is not. That is the practical meaning behind Bonfy's argument that continuous control produces audit evidence as a side effect of enforcement, instead of requiring a separate, after-the-fact exercise to manufacture it.

Financial Services Has Its Own Version of This Problem

Retail banking is the headline case because Barr named it directly, but the authority gap applies across the financial services sector, not only consumer accounts. Wealth management, broker-dealer operations, and insurance claims processing all depend on the same underlying assumption, that whoever or whatever accessed a customer's data did so within the bounds of an actual customer relationship, not merely within the bounds of a technical permission. Bonfy has written about this gap directly in the context of customer relationship obligations under GLBA, Regulation S-P, and FINRA rules, all of which assume a human-scale definition of relationship that an agent's technical access controls do not automatically respect.

Kiteworks has made a related argument about why financial services AI governance cannot rely solely on model-level controls. The regulations that govern financial data, SOX, GLBA, SEC rules, and now the emerging agentic frameworks out of Hong Kong and the Federal Reserve, regulate the data and the relationship, not the technology that touched them. An agent that technically has access to lending decision data under a broad grant is not automatically authorized to use that data for every purpose a customer relationship could theoretically justify, and examiners are increasingly aware of the difference.

What Institutions Should Do Before the Next Regulatory Cycle

None of the steps below require waiting for the SFC's principles-based guidance or a finished Project SAGE framework. They are things a CISO or CCO can start now, with the authority gap framed explicitly rather than treated as an access-control footnote.

Build an inventory of every agent that can touch a customer account, with the specific actions each one can initiate, not just the systems it can reach. Most institutions can list their agents. Far fewer can list, per agent, which irreversible customer-facing actions it is capable of triggering without a fresh, per-action consent check.

Separate technical permission from action-level authority in writing, for every agent on that list. If the honest answer for a given action is "the agent can do this because nothing stops it," that is the gap a regulator will eventually ask about, and it is cheaper to close before the question arrives than after.

Build the capability to produce a per-action evidence record, not a quarterly access report, for any agent action a regulator or a customer disputes. The question will be "prove this specific transfer was authorized by this specific customer," and an access log that only proves the agent had standing permission will not answer it. Non-human identity governance is the starting discipline for making that distinction enforceable rather than aspirational, and a data protection impact assessment is a reasonable place to document where the gap currently sits.

Treat Project SAGE and the SFC's forthcoming principles-based guidance as a preview, not a deadline. Hong Kong is not the only jurisdiction working this problem, and institutions that wait for a finalized rule before building verification capability will be building it under examination pressure instead of on their own schedule.

Frequently Asked Questions

1. We are rolling out an AI agent that can move money between a customer's accounts; how is its technical permission to do that different from its actual authority to do it?

Technical permission means the agent's credentials allow the action; authority means a specific human consented to this specific instance of it. A bank can grant an agent broad transfer permissions while no customer ever authorized a particular transfer the agent initiates. Regulators including the Federal Reserve now expect institutions to verify authority per action, not just confirm permission exists at the account level.

2. How can a bank verify that an AI agent's action on a customer account came from a real instruction, rather than a prompt-injected command or a hijacked session?

A bank verifies this by checking the specific action against the entity, context, and relationship involved at the moment it is attempted, not by trusting that a valid session proves valid intent. Session validity confirms a login occurred. Explicit per-action consent verification, the control Federal Reserve Governor Michael Barr called for in November 2026, confirms a human actually meant for that transaction to happen.

3. Is it safe to let an AI agent initiate withdrawals or account changes for retail customers without a human approving every individual action?

It is not safe without technical controls that verify authority at the moment of action, because human review of every step is widely acknowledged as unrealistic at scale. Hong Kong's Securities and Futures Commission has said reviewing every intermediate agent decision by hand does not work, and plans graduated guidance instead, stronger safeguards for higher-impact actions like withdrawals, lighter review for reversible, low-risk tasks.

4. If a regulator asks our bank to prove that an AI agent's account action was properly authorized, what evidence do we need, and how quickly must we produce it?

A bank needs a record created at the moment the agent acted, showing which agent acted, under what delegated authority, against which customer relationship, and what consent verification occurred, not a reconstructed narrative built afterward from scattered logs. Regulators increasingly treat speed and completeness of that record as evidence of control maturity in its own right, separate from whether the underlying action was ultimately proper.

5. What is the difference between how Hong Kong and the US Federal Reserve are approaching AI agents acting on bank customer accounts?

Hong Kong, through the Hong Kong Monetary Authority's Project SAGE and the Securities and Futures Commission, is building named frameworks, including agent identity verification and principles-based guidance for higher-autonomy use, while the Federal Reserve has so far raised the risk through Governor Barr's public remarks without yet proposing a specific rule. Both agree the deploying institution, not the AI agent, carries the liability.

To learn more about agent authority verification, schedule a custom demo.