Sooner or later, someone outside your firm will ask how you know your AI did the right thing. It may be a board member, an examiner, your outside auditor, or a buyer doing diligence. None of them will accept that the system checks itself.
Picture a commercial real estate firm at the leading edge of AI. Its brokers and analysts build their own prompts, skills, and models, what we call “Citizen AI“. A few of those helpers have been promoted into production tools that draft offering materials, rank likely bidders and summarize diligence files. Someone has already profiled which people and which tools may see which data. Part of its capital markets work involves securities, so it also runs a registered broker-dealer that FINRA and the SEC can examine.
Our argument is that this firm has built almost everything except the part outsiders trust. For the rest of your business, that part has a name: internal audit. Good AI governance plays the same role for AI. It is a human layer, small and well equipped, but human. Trust in a business has never come from the machinery. It comes from accountable people who test the machinery, sign their names and keep the record.
Trust in a business has never come from the machinery. It comes from accountable people who test the machinery, sign their names and keep the record.
If your firm is earlier on this road, our 90-day pathway to an AI governance audit covers the basic evidence folder. This piece is for the firm that already has the tools and the controls, and now needs the credibility.
Why Can’t the Tools Vouch for Themselves?
Because a system’s report on itself is a claim, and outsiders want proof from someone they can question. That is our view, and the regulators’ own words point the same way. In its oversight report for 2026, published in December 2025, FINRA described the risks of AI agents, the kind of AI that takes actions and does not only draft. An agent may act without a person approving the step. It may act beyond the authority its user meant to give it. And its multi-step reasoning, in FINRA’s words, “can make outcomes difficult to trace or explain.” Among the things FINRA suggests firms consider is where to keep a human in the loop.
The SEC makes the point about itself. On October 23, 2026, its Division of Examinations published its priorities for fiscal year 2027. The document, which uses the term super intelligence for AI, says the Division is bringing the technology into its own examinations, “supplementing, but not replacing, staff judgment.” Our reading is plain. If the examiners keep a person in charge of their own AI, they will look for the person in charge of yours.
What Does Internal Audit Teach About Trust?
The Institute of Internal Auditors describes the purpose of internal audit as giving the board and management independent and objective assurance. In our summary, four habits make that assurance believable. The auditors are independent of the work they review. They test instead of taking anyone’s word. They keep a written record of what they tested and what they found. And they have a direct line to the board. None of the four is a technology.
Internal audit is also a useful model because it is changing its tools without changing its nature. In a survey released in February 2026 by the Institute’s research foundation and an audit software vendor, 83 percent of audit leaders expected their teams to use more AI within a year. Treat that with care: it covers about 370 North American leaders, it is self-reported, and one of its two publishers sells audit software. Our reading is that the profession is buying better tools for its people, not replacing the people who sign.
The same foundation shows how far the trusted layer still trails the technology. It surveyed 3,285 chief audit executives and directors in 132 countries and reported in September 2026 that 58 percent rank digital disruption, including AI, among their five highest risks. Only 42 percent rank it among their top five audit priorities, and 11 percent report full audit coverage of it. These are self-reported figures from audit leaders worldwide, so read them as a signal. The signal is that the checking has not caught up with the building, and a firm at the leading edge cannot wait for it to.
What Does That Layer Look Like in a regulated context?
In a finance broker-dealer the comparison is close to literal, because the rules already assume accountable people. FINRA Rule 3130 requires the broker-dealer’s chief executive to certify every year that the firm has processes to establish, maintain, review, test and modify its compliance policies and supervisory procedures. Those processes must be “evidenced in a report” that goes to the board. A person signs. We made a similar point about a different signature in our piece on a $13 million cyber settlement.
FINRA’s supervision rule requires that each registered person be assigned to an appropriately registered supervisor who is responsible for that person’s activities. Our recommendation is to give every AI agent the same treatment: a line in a registry that says who built it, which principal supervises it, what it may see and do, when it was last reviewed and when it will be retired. No rule requires this. It is our judgment about the cleanest way to put a name next to every tool, and it matters most for the helpers your own people build, a balance we covered in The Best of Both Worlds.
Then there is the testing. A separate FINRA rule requires the firm to test and verify its supervisory procedures and to report the results to senior management at least once a year, and FINRA has said that an internal audit process may be used for that testing, in whole or in part. In our view, the people who already test how your firm supervises its brokers are the natural people to test how it supervises its AI.
How Small Can the Human Layer Be?
Smaller than you fear, if the tools do the recording. Think of a flight recorder for AI work: a set of records the tools write as they run, covering what the tool was asked, what data it pulled, what it produced, which model version it used and who approved the result before it left the firm. The access profiles you have already built belong here too. A profile says what an agent is allowed to see. A log shows what it did see.
A profile says what an agent is allowed to see. A log shows what it did see.
That leaves people with three jobs the tools cannot do for themselves. They choose what gets tested, and FINRA allows that choice to rest on risk and sampling. They look hard at the exceptions. And they sign. In a midsize firm, our working assumption is that this is a few named people with the right tools, not a new department. The catch is that the checking has to be real, which is the subject of Human in the Loop or Rubber Stamp?
Here is a test of whether the layer exists. We call it a replay test. Pick one offering memorandum that AI helped write last spring. Ask two people who did not produce it to show, within two business days, how each figure got there and who signed off. Two days is our illustrative bar, not a regulatory one. If they cannot, every control may have worked and you still cannot prove it.
Who Has to Trust It?
Three audiences, and they want the same thing. Your board receives the certification report and, as we wrote in The 9% Problem, wants evidence an outside party can check. Examiners and auditors read the record and ask who supervised the work. In September 2022 the SEC announced charges against 16 Wall Street firms, with penalties totaling more than $1.1 billion. The charges were about records and supervision, not fraud: employees had discussed business by text on personal devices, the firms had not kept most of those messages, and the firms admitted the facts. Those firms are far larger than yours and the dollar figure will not transfer. The lesson does.
The third audience is a buyer, whether that is an acquirer, an investor or an institutional client deciding whether to trust your numbers. Diligence is a search for what can be verified. In our view, a firm whose AI work can be traced to named, accountable people is easier to buy from and easier to buy. A buyer cannot interview your software. A buyer can interview the person who signed for it, and read what that person kept.
A buyer cannot interview your software. A buyer can interview the person who signed for it, and read what that person kept.
What Do You Decide, and by When?
Three decisions are yours. First, name the layer. Decide which people vouch for the firm’s AI and where they report. Borrow the auditors’ rule here: the Institute’s standards presume objectivity is impaired when someone gives assurance on work they were responsible for within the previous 12 months. So keep the people who check apart from the people who build, and give the checkers a line to the board.
Second, fund their tools as part of the AI budget, not as overhead added later. When the record is a byproduct of the tool, your producers never fill out a form. When it is bolted on afterward, it becomes one more load on your best people, on top of the verification tax that already slows AI work in this industry.
Third, decide what must be on the desk before the next certification is signed. We would ask for the agent registry, the list of procedures changed because of AI in the past year, and the result of at least one replay test like the one above.
The deadline is not abstract. Rule 3130 requires each certification no later than the anniversary of the last one, and the SEC’s fiscal year 2027 began on October 1.
What This Will Not Do
This is not legal advice. Whether your procedures meet FINRA and SEC rules is a question for your counsel and your chief compliance officer, and nothing here changes what those rules require.
The comparison with internal audit is an analogy. We are not saying a governance team is an internal audit function or meets its professional standards, and a firm that has internal auditors should decide with its board where this layer sits. A human layer that glances and approves adds a signature, not assurance. The registry and the supervising principal are our judgment, not a rule. The surveys we cite are self-reported, and one was co-published by a software vendor. Regulators’ positions are still moving, so check current guidance before you rely on any of this.
Where to Start
Run the replay test in the next 30 days. Choose one AI-assisted document that went to an investor or client last quarter, hand it to two people who did not produce it, and time how long they need to show how it was made and who approved it. The gaps they hit are your first work list, and the names of the people who closed them are the start of your human layer.
If you want an outside view of where your firm stands, our assessments start with what you can prove today, and you can reach us through the contact form on our About page.


