The question your board should ask about AI is not “what are we doing?”
A list of AI projects tells a board what management chose to launch. It does not show where AI already shapes decisions nobody approved.
A director asks the obvious question: what are we doing with AI? Management answers with a slide: a few experiments, a vendor or two, a steering group, a policy being drafted. The board nods and moves on.
Nothing in that exchange is dishonest. It is simply an answer about the wrong object. It describes what the company decided to launch. It does not describe where AI output is already influencing what the company does to customers, employees, suppliers and its own numbers.
A project list shows what management chose to start. A decision register shows where the company is already exposed.
Why “what are we doing?” returns the wrong answer
The question invites a portfolio answer because a portfolio is what management can count. Sanctioned projects have budgets and owners. Use that grew from the bottom up has neither.
My argument is a view about what to ask, not a finding. Some recent surveys are consistent with it, but they cover different groups and cannot be added together, so I take them one at a time.
Directors. The National Association of Corporate Directors published a data pack from its 2025 public company board practices survey (n=146, July 2025). Asked which AI oversight activities their boards had performed, about 62% reported setting aside full-board agenda time for AI. About 23% reported auditing where AI is currently used in the company, and about 6% reported establishing metrics for management reporting. Respondents could pick several activities. The page gives no fieldwork dates, and the full report is member-gated. These are boards describing their own activity, not a measure of how well they oversee AI.
Employees. Microsoft commissioned Censuswide to survey 2,003 UK employees in October 2025. Of them, 71% said they had used unapproved consumer AI tools at work, and 51% said they still did so every week. It is vendor-commissioned, covers stated behavior in one country, and gives no question wording. It concerns consumer tools in general, not consequential decisions, and does not show harm.
A signal, not a trend. A Stanford GSB article published in November 2026 summarizes a working paper based on 20 interviews with directors of public and private companies in regulated industries. One of its authors says most boards lack a clear picture of which AI solutions are being rolled out, and the article reports that only 54% of respondents had clear guidance from management on what they can and cannot do with AI. Twenty interviews say little about boards in general. Treat it as a prompt to test in your own boardroom.
The samples do not show that boards are blind to employee AI use. At most they suggest that the project question is easy to answer and the others are not yet routine. That is the only inference I draw.
The better question: where is AI already shaping decisions?
My proposal is to put this question to management instead:
Where does AI output already influence a consequential decision that we believe a person is making, and how would we find out if it were wrong?
Each part does work.
Already. It covers the present, including tools nobody procured.
Consequential. It filters out drafting a meeting summary and keeps credit, pricing, hiring, claims, safety, customer communications and financial reporting.
That we believe a person is making. A person may sign the decision while a model wrote the recommendation, and may have neither the time nor the information to disagree. Human sign-off and human judgment are different things. The board needs to know which it has.
How would we find out. This is the oversight part. Systems whose errors would surface only through a customer complaint or a regulator are common.
The NIST AI Risk Management Framework, version 1.0, released in January 2023, already points in this direction. Its GOVERN function includes having mechanisms in place to inventory AI systems, and it states that executive leadership takes responsibility for decisions about AI risk. NIST describes it as voluntary guidance, and NIST has said a revision is in progress, so check the current text. It is a reasonable reference for what to ask for, not a requirement.
Replace the project list with a decision register
As a starting design, I would ask management for one page, refreshed on a fixed cadence, with a short list of rows. Each row is a decision, not a tool or a project:
| Column | The question it answers |
|---|---|
| Decision | What is being decided, and about whom? |
| AI involvement | Does AI recommend, draft, rank, decide or act? |
| Human role | Who reviews it, with what time and what information? |
| Error signal | How would a wrong outcome become visible, and how fast? |
| Owner | Which executive answers for it, by name? |
| Last evidence | When was it last tested against the previous process? |
Two rows are worth more than ten. The register is not an inventory of every prompt typed at work. It lists the places where an error would matter enough for directors to want to hear about it.
An illustration I made up, not a reported case: a support team uses an assistant to draft refund decisions. Each decision is approved by an agent. The project list says “customer service pilot, on track.” The register would ask how many drafts the agents change, how many they accept in under a minute, and which complaints trace back to a draft. If nobody can answer, that is the finding.
A test for your next board meeting
This is a test I propose, not a validated instrument. Before the next meeting, ask management three things in writing:
- Name the three most consequential decisions in the company that AI output now influences, formally or informally.
- For each, who would notice a wrong output first, and how long would it take?
- What do we know about AI use that the company has not sanctioned, and how do we know it?
If the answers arrive as project names, the question has been answered at the wrong level. If the third answer is “we don't know,” that is honest and useful. The task is then to decide how much it matters.
The counter-argument
There is a serious objection. Boards govern; they do not run operations. A decision register could pull directors into management detail they are not equipped to judge, and invite a compliance exercise that goes stale within a quarter.
That is a fair risk. Keep the register narrow, and let the board read it rather than maintain it. Directors need to see that a few high-consequence decisions have owners, error signals and recent evidence, and how management chose them. If that selection logic is weak, that is the oversight finding.
A second objection: unsanctioned use is often harmless. Often it is. The point is not to hunt it but to know whether any of it sits near a consequential decision. Treating all of it as misconduct pushes it further out of sight.
Directors' legal duties around technology oversight differ by jurisdiction and company type. Nothing here is legal advice; check what applies to your board with counsel.
What changes
The project question is still worth asking, because pilots consume money and attention. But it is the second question. The first is where the company already relies on a system it has not examined.
Ask for the register once. How management builds it, what it leaves out and how long it takes will tell the board more about its AI exposure than another slide of initiatives.
Where this connects
This essay asks what a board can see. Your AI roadmap is probably missing the only column that matters asks what evidence should end a project, and AI is making intelligence cheap, judgment is not getting cheaper covers which decisions deserve human judgment. For the human-review column in the register, see The hidden metric in AI automation is supervision.