On November 16, 2026 a Federal Reserve governor said in a speech that major banks report finding AI agents attempting to act in retail customer accounts. The speech names no bank that found such agents, describes no case and sets no rule. It lists four safeguards. A footnote calls the views the speaker's own. In our reading, the safeguards are questions for any owner whose software acts for customers.
On November 16 a Federal Reserve governor gave a speech about AI and the financial system.
One paragraph covers ordinary customers' accounts. In it the speaker said that major banks report finding AI helper programs trying to act in customer accounts. The gap, in the speaker's words, is the bank not being able to "verify authenticity, consent, and alignment with customer instructions".
Think of a stranger at a counter holding a note that says "act for me" (our comparison): the teller would want to know the note is genuine, that the customer agreed, and what it allows.
On November 16, 2026 a Federal Reserve governor gave a speech at a conference in Milan. In one paragraph the speaker said that "major banks now report finding AI agents attempting to take actions in retail customer accounts". This page reads that paragraph. It is a decode of a speech, not of an event: the speech page names no bank as having found such agents and describes no specific case.
The speech's first footnote reads "The views expressed here are my own" and adds that they are not necessarily those of the speaker's colleagues on the Federal Reserve Board or the Federal Open Market Committee. Below, "the speaker" means that one governor, and we do not present anything here as a Federal Reserve position.
What did the governor say about AI agents in customer accounts?
An AI agent, in our plain definition, is software that takes actions for someone instead of only answering questions. The speech page we read does not define the term.
The speaker's sentence has three parts. First, who reports: "major banks", unnamed. Second, what they report: AI agents "attempting to take actions" in "retail customer accounts" (our gloss: accounts held by ordinary customers). Third, the gap. The speaker put it this way: "without the bank being able to verify authenticity, consent, and alignment with customer instructions".
In plain words, the speaker names three things the bank is not able to verify in the cases it reports. Is the action authentic? Did the customer agree to it? Does it match what the customer asked for? The three questions are our reading of the three terms. The sentence does not say what is authentic: the agent, the customer's sign-in or the request. In our reading, not being able to verify consent is not the same as knowing there was none: the sentence does not say that any customer objected or that any action went through.
The paragraph sits in the section headed "AI in Consumer Finance" and opens with "New risks continue to surface."; the banks' report is an example. That banks "now report" this is the speaker's relay. The page does not say how the banks found these agents, how the speaker knows, or how many reports there were. We read it as the speaker's statement, not a survey or a Board finding.
Which safeguards does the speech list?
The next sentence says "AI in financial services can only succeed long term with appropriate guardrails" and lists four, in this order. Guardrails is the speaker's word; we use safeguards as a synonym. The fourth item names two kinds of constraint, so a reader could count five.
| As the speech prints it | What it could mean in a business (our reading, not the speech's) |
|---|---|
| "strict authentication" | Each request shows who is behind it: the customer, the agent, or the agent acting for the customer. The speech does not say which. |
| "explicit consent verification for specific actions" | The customer's yes is checked and recorded for the specific action being asked for, not one broad yes for everything. |
| "technical controls limiting agent capabilities" | Limits built into the system on what the software can do: which actions, how much, how often. |
| "both rules- and values-based constraints" | The speech defines neither kind, and we do not guess what either would look like. |
The speaker adds that technology is needed so AI agents follow established rules and values, "especially critical as agentic commerce expands". The speech page does not define agentic commerce. It does not tell any business to adopt the four safeguards: it speaks of AI in financial services.
What can a business check this week?
The paragraph speaks about banks and retail customer accounts. This section is our translation, labeled as ours, for any owner whose software, or a supplier's, acts for customers or staff. We are not lawyers or financial advisers; this is not advice about any account or provider. Three of the four questions below follow the speech's list. The fourth builds on one line from the paragraph before it (see the table) and is ours.
| Question an owner can ask (ours) | Closest words in the speech's list |
|---|---|
| Who may act as a customer, or as a member of staff, through an AI agent, and how does each request prove who sent it? | "strict authentication" |
| What consent is recorded for each specific action, and where is it kept? | "explicit consent verification for specific actions" |
| What is the most the agent can do in one step and over a day, and who set those limits? | "technical controls limiting agent capabilities" |
| Who is told when the agent acts, and how can a customer or a colleague dispute what it did? | Not one of the four. The paragraph before it says institutions must "provide accessible dispute mechanisms" to use AI responsibly (said of AI in general, not of agents). Applying that to notice and to an agent's actions is our addition. |
For one vendor's way of giving an agent its own identity, see our teardown of IBM Agent Identity and Vault. For approval steps, limits and logs around software that acts for you, see our guide to AI agents in production.
Do this week
In our reading, not the speaker's, three steps if your software, or a supplier's, acts for customers or staff. We do not advise anyone to buy, avoid or switch any product.
- List every AI agent, yours or a supplier's, that can act as a customer or as staff, and what each can change.
- For each, write down how a request proves who sent it, what consent is recorded per action, and its limits per step and per day.
- Decide who is told when it acts and how an action is disputed.
What does the speech not say?
We read the speech page and nothing else on the subject. The page we read states none of the following:
- The name of any bank that found an AI agent, how many reports there were, when they were made or what they cost anyone.
- An example of a reported agent action, whether it went through or was stopped.
- A rule, a deadline or a supervisory action on AI agents in accounts. The page is a speech.
- What an AI agent, a rules-based constraint or a values-based constraint is.
Our October piece on banks restricting third-party agents asked a related question from press reports. This page reads one primary text instead.
What did we leave out, and why?
The speech is long and mostly about other things. The speaker says at the outset "I am an optimist about the long-term promise of AI", and ties that hope to managing today's challenges well. It then covers the AI investment boom, trading, cyber risk and consumer finance.
We leave those sections out. Of the six-sentence paragraph we read, we use the first four and leave out the last two, which are about deposits in stress periods: they are the speaker's scenario, not a report from banks. The speaker presents the investment-boom sections as a recap of the Federal Reserve's Financial Stability Report, which we did not open, so we use none of its figures.
How do we know?
On November 18, 2026 at 11:10 PST we fetched the speech page on the Federal Reserve Board's website, which prints "Last Update: November 16, 2026". We read it in full, matched every quote word for word against a saved copy, and also quote one line from the paragraph before the one we decode. We did not read the Financial Stability Report, the PDF version, any press coverage or any bank's terms, and we tested nothing. This page analyzes one speech, not the speaker's views as a whole.
Read next
The Signal / Noise digest of November 18 carries a section from an online user agreement on third-party AI agents, a different kind of document from a speech. The AI Failures hub lists our other decodes. For AI agents with limits, approval steps and records built in, see our AI agents page.