Executive Summary
The American central bank has put on the public record that AI agents are already moving inside other people's accounts. The occasion was a speech by Federal Reserve Governor Michael Barr on November 16, 2026, at a finance AI conference in Milan. He said large banks are reporting AI agents that try to do something inside the accounts of individual customers — and he attached one condition to that. The bank cannot verify that the agent is genuine, that the customer consented, or that the action matches what the customer asked for.
Barr argued that if the gap stays open, the speed at which deposits leave a bank changes. Once an agent takes over the step a person used to perform — the transfer itself — it can move the money on a difference in rates the moment it reads one, and an outflow that fast outruns both the bank's liquidity management and the point at which a supervisor could step in. The agents inside the accounts are a fact banks have reported. The run that could finish in minutes has not happened. Elsewhere in the speech sit the debt figures of the AI companies.
Sections 1 through 4 are what the speech says. Section 5 is this article's reading. It takes up the warning where the record of delegation is empty.
Key Figures
The numbers in the speech split into two kinds. One is the speed of a scene that has not happened; the other two are debt that has already piled up.
Source: Federal Reserve — Governor Michael S. Barr, "Some Potential Financial Stability Implications of Artificial Intelligence" (November 16, 2026, Milan).
Minutes
How short a deposit run could become
A warning conditioned on agents reading the signal and moving funds autonomously. It is not offered as an event that has occurred
$800 billion
On-balance-sheet debt of the seven largest AI firms
It grew sharply through 2026 and was expected to pass this line before year end. The speech does not name the seven
$3 trillion
Potential notional maximum of off-balance-sheet obligations
An estimate from the Fed's Financial Stability Report, and the speech keeps the words "potential notional maximum." It is not measured on the same scale as the debt above
The One Sentence a Central Bank Put on the Record
The speech is titled "Some Potential Financial Stability Implications of Artificial Intelligence," and the venue was the seventh ACM International Conference on AI in Finance. The body breaks into six parts: the AI investment boom, leverage in the data center buildout, unintended consequences of AI in wholesale securities trading, market herding and a possible stampede, AI and cyber risk, and AI in consumer finance. The passage this article holds on to is in the last part.
That part opens with the good side. Alternative data sources such as rental payments, utility bills, or employment patterns may expand access to credit for people whose thin files keep pushing them out of underwriting. Then the risks arrive, and they catch on the same data. Those alternative sources may inadvertently serve as proxies for protected characteristics like race or national origin, Barr wrote, and systems leaning on incomplete or biased data may generate inaccurate information. The warning that a recommendation may rest not only on the investor's own circumstances and objectives but also on factors that advance a firm's commercial interests belongs here too. Every risk this part names points back at data.
After the data, a single sentence reports what banks are already seeing.
"Major banks now report finding AI agents attempting to take actions in retail customer accounts, without the bank being able to verify authenticity, consent, and alignment with customer instructions."
Governor Michael S. Barr, Milan, November 16, 2026, in the section on AI in consumer finance.
The weight of the sentence rests on "now report finding." A clause looking ahead at what might happen would have carried a different tense. Barr wrote that banks are living through this and reporting it. The agents turn up in retail customer accounts rather than corporate ones, which is to say in the accounts of individuals. And the verb is "attempting." Nothing in the sentence says the action ran to completion. What the bank saw is the trace of a try.
Barr went on to set out what would have to be in place. AI in financial services can only succeed long term with appropriate guardrails, he said, and he counted four: strict authentication, explicit consent verification for specific actions, technical controls limiting agent capabilities, and constraints grounded in both rules and values. Saying that four things must be built points at the same spot as saying that all four are currently empty.
A similar scene reached paper first in another industry. United Airlines amended its contract of carriage on November 11 to bar unauthorized personal AI assistants from booking and ticketing flights (related article). That, though, was one company drawing a line in its own terms. This time a governor of the central bank that runs monetary policy and supervises banks wrote the same gap into a public speech, and the subject was the banking system as a whole.
Three Questions the Data Cannot Answer
The three words Barr lined up in one sentence are authenticity, consent, and alignment. Most coverage has carried them as a single bundle, but each is a different question and each asks for different data.
Authenticity asks whether the program that sent this request really is the assistant that customer uses. In front of a bank server today, a person and a program arrive in the same shape. An assistant handed an ID and a password is indistinguishable from the customer at the login screen. The moment one set of credentials changes hands, everything that account can do travels with it.
Consent asks whether the customer agreed to this action. The qualifier Barr attached when he listed the guardrails, "for specific actions," catches here. A blanket agreement clicked once at sign-up explains nothing about the single transfer that went out this afternoon. The question only gets an answer if consent is kept as an event that happens per action rather than a setting parked on the account.
Alignment asks whether the action stayed inside what the customer asked for. "Move the spare cash to whichever account pays more" and "send today's entire balance to another bank" are different instructions. If an assistant that heard the first did the second, authentication was not broken and consent was not missing. It is a case that cannot be sorted out because nothing on record says how far the delegation reached.
Set the three side by side and the common thread shows. Not one of them asks about model performance. An assistant that compares rates better, or calculates more accurately, leaves these three fields exactly as empty as before, because the request that reaches the bank server does not carry that information.
Speed Removes the Time to Respond
A few lines later in that part, Barr sketched the scene that follows. When a crisis hits, the time a bank has to work with gets shorter.
"AI agents may accelerate deposit outflows during stress periods. Where customers once needed to execute transfers, AI agents could instantaneously detect yield differentials or stability signals and autonomously move funds at high speed, potentially accelerating runs that unfold in minutes, overwhelming traditional liquidity management and regulatory intervention timelines."
The speech again, a few sentences after the quotation above.
The sentence does not enlarge the fear. What it shrinks is the length of time anyone has to act. Even in the recent episodes where deposits left through a phone, somebody still had to open the app, type an amount and press a button. The interval in which a person hears the news, makes up their mind and moves a hand was the time a bank and its supervisors were given. Barr's sentence takes that interval out.
Two places have changed. The step where the signal gets noticed, and the step where the money gets moved.
The speech separates them too. Picking up a yield differential or a stability signal is the first step; shifting the funds without anyone's hand on it is the second. The word "minutes" holds only when both steps go automatic together. If only the noticing is automatic, the money waits for someone to press the button. If only the moving is automatic, somebody still has to read the signal first. As long as a person remains on one side of the pair, the bank keeps some of its time.
| Step | Before | What Barr described |
|---|---|---|
| Noticing the signal | The customer watches rates and news | The agent detects it instantaneously |
| Moving the money | The customer had to execute the transfer | The agent moves funds autonomously, at high speed |
▲ Pebblous original summary. The lower cell under "Before" is the speech's "customers once needed to execute transfers", and the upper cell spells out the state that sentence assumes. The right column is the clause that follows it.
One line is worth keeping separate. What Section 1 describes is something banks reported. The run in this section is a hypothesis growing out of that report. The modal verbs Barr used are "may" and "could," and nowhere does the speech say that such a run has happened anywhere.
More of the Speech Went to Debt Than to Accounts
The agents-in-accounts story is one passage of the speech. By length, far more room went to the financial structure of the AI investment boom, and that is mostly what coverage at home and abroad pulled for headlines. For this article, what matters is that both stories live in the same speech.
Take the debt that is on the books first. On-balance-sheet debt at the seven largest AI firms grew sharply through 2026 and was expected to pass $800 billion before the year was out. Which seven is not written down in the speech. A tally naming five hyperscalers and two chip companies circulates through the coverage, but that list came from the reporting rather than from Barr.
The side that is not on the books is bigger. The Fed's Financial Stability Report estimates that obligations of this kind, meaning off-balance-sheet ones, now exceed $3 trillion at their potential notional maximum. Note the phrase attached to it — this is an undiscounted figure that mixes in commitments subject to conditions, so measuring it on the same scale as the $800 billion and adding the two produces an overstatement.
The report also writes down where it places these numbers. It sets the cumulative capital expenditure and bond issuance the hyperscalers have run up since 2020 against what the whole telecom industry did during the broadband investment cycle that began in 1995. That comparison brings something to mind clearly enough without being spelled out.
None of which puts the speech on the pessimistic side of AI. Barr said first that AI's role in consumer finance will grow, and he drove the argument toward thinking through the risks and implications before any new technology is integrated into trading activity or market infrastructure more broadly. The condition he attached runs like this: when the technology is still not fully understood, and its development may be galloping ahead of our understanding, some caution is warranted. That is closer to laying the guardrails first and letting understanding catch up than to arguing against the technology.
Why Pebblous Is Watching This Speech
Reading this speech only as an event in financial supervision would waste one paragraph of it. The three things Barr counted all ask about records, not about models. Authenticity asks whether the party that sent the request is written down. Consent asks whether the moment of agreement to that action survives anywhere, and alignment asks whether the reach of the delegation was ever recorded. The reason a bank cannot verify is not that it cannot judge, but that the material to judge on never arrives.
In most service logs today, a request from a person and a request from a program acting on their behalf pile up looking exactly alike. The account that logged in, the time of access and the balance that changed are all kept; who caused that line to be written is not. On logs like these, after-the-fact verification does not hold up. When something goes wrong there is no basis for telling a customer's instruction apart from an assistant running loose, and without that, responsibility cannot be divided either.
When the record is empty, the choices narrow to two: block everything or allow everything. The airlines have spent recent weeks swaying between the pair. A bank standing in the same place would pay more for the swaying. Blocking everything costs customers an assistant they have started using, and allowing everything moves the minutes Barr sketched one step closer to being real. The third path lies only on the side where the record gets filled in.
That is also the part a data team can reach today. Write one field that separates a request coming from a person from a request coming from a program acting for them, and when a program is acting for someone, keep who that someone is and how far the delegation reaches alongside it. Keep consent as an event stamped per action rather than as a state value in the account settings. Cleaning data up and recording who touched that data under what authority are two different jobs. In an organization that has deferred the second, doing the first well adds no questions it can answer afterwards.
When my AI assistant moved my money, where is the evidence that it was me? Take that question to the transaction history today and what comes back is a line in a log. The line says when, how much and where to. It does not say who asked.
References
- 1.Barr, M. S. (2026). "Some Potential Financial Stability Implications of Artificial Intelligence." 7th ACM International Conference on AI in Finance (ICAIF), Milan. Board of Governors of the Federal Reserve System.
- 2.Board of Governors of the Federal Reserve System. (2026). "Financial Stability Report." November 2026.
- 3.Richter, W. (2026). "Top 10 Risks of the AI Boom to Financial Stability, According to Fed Governor Barr." Wolf Street.