Executive Leadership
The AI Page in the Board Pack: What a CEO Should Report Every Quarter
Directors have started asking better questions about AI. Most CEOs are still answering them with a slide of pilots and a vendor logo. One page, built the same way every quarter, fixes most of that.
I have written a lot on this site about what a board should see. The change control piece ends with a section by that name. So do the pieces on retiring agents and shadow agents, and the oversight agenda is nothing but questions directors should be asking.
Somebody has to write the answers down.
That somebody is the CEO, and the format matters more than most CEOs think. I sit on both sides of this. As CEO of Visiting Media I am the one preparing the material, and at ACT | The App Association I am one of the people reading what management sends. From the reading side, the AI section of a board pack is usually the least useful few pages in it, because it changes shape every quarter and nobody can tell whether things got better or just got described differently.
One page, and it does not change shape
The guides that rank for this topic mostly recommend a deck of ten or twelve slides, or a two-page written briefing. I would go shorter. One page in the main pack, with an appendix for anyone who wants the detail, and the rows on that page stay identical from quarter to quarter even when a row is boring.
Boring rows are the point. A director who sees the same table in January and again in July can spot movement without being told where to look, which is the only way a part-time board member with a day job ever catches drift. The moment management starts redesigning the page to tell a better story, the page stops being evidence.
What goes on it
These are the rows I use, in the order I use them. Some take a single number. One gets its own section below, and an eighth line lives at the very bottom in small type.
- The register count. How many agents and scheduled model jobs are running in production, with a named human owner for every one of them. If the count and the owner list disagree, say so in the row.
- What changed this quarter. New agents, plus any agent that moved up a tier on the delegation of authority matrix. Model upgrades that went through change control get listed underneath, one line each.
- What we turned off. Retirements and rollbacks. More on this one shortly.
- Incidents, by severity. Severity first, count second. A raw incident count tells a director nothing, and it punishes the teams honest enough to log small ones.
- The sampling grade. The weekly human-graded sample of agent output, trended. If you do not have one yet, write “none” in the row and leave it there until you do. That empty row will do more work in the boardroom than a paragraph of reassurance.
- Consolidated spend, against what the work produced. The model vendor invoices go in that number, and so does the subscription quietly standing in for a role nobody backfilled.
- The decision we need from you, if there is one.
Last row first, when the board meets. If management needs a decision, the board should hear it before the tour of the table.
Report what you turned off
None of the board reporting guides I read before writing this mention subtraction. They cover cost and risk in careful detail, and every row they recommend measures something that was added.
I think the turned-off row is the most honest one on the page. In my own fleet I retired more agents this year than I launched, and each retirement was a judgment that something had stopped earning its keep (often because a model upgrade made it redundant). A management team that never reports turning anything off is either not looking or not willing to say what it found, and a board cannot tell those two apart from the outside. The offboarding discipline I described earlier produces this row almost for free: every agent marked retired on the register, with the date its credentials were revoked.
Rollbacks belong here too. When a model upgrade went out and got rolled back inside the window, that is change control working, and a board that only ever hears about successful upgrades learns to distrust the word “successful.”
Say which parts an agent wrote
The uncomfortable row goes at the bottom of the page, in small type.
Plenty of board packs in 2026 are partly drafted by AI, and I have no objection to that; I would be a strange person to write about agent fleets and then insist on assembling a pack by hand. My objection is to a director reading a confident number that no human ever re-derived from the source system, because agents are very good at producing artifacts that look like evidence, and a board pack is the single most consequential artifact a company produces in a quarter.
So one line at the bottom: which sections an agent drafted, and the name of the person who checked every figure in them against the system it came from. It takes ten seconds to read. It changes how carefully the checking gets done, because now somebody’s name is on it.
What stays off the page
Demos. The board does not need to watch an agent write an email in real time, and a live demo in a board meeting mostly measures how lucky the presenter got that morning.
Model names and benchmark scores go in the appendix if anywhere. So does the list of pilots, which I would cut entirely unless a pilot is about to ask for budget. And competitor AI announcements, which take up more board time than almost anything else in this area and almost never lead to a decision, belong in a separate strategy discussion where someone owns the response.
Between meetings
A quarterly page assumes nothing urgent happens in between, and with agents that assumption fails eventually. Agree on the escalation threshold with the chair before you need it. The incident playbook I use puts a plain-language note in front of every stakeholder within twenty-four hours of a customer-visible incident, directors included. The quarterly page should show those events again as a row, so the record lives in one place and nobody has to dig through old email to reconstruct what the board knew and when.
A board AI page has one job, which is making a complicated operation legible to a smart person with ten minutes and no background. The CEO should write the first version personally, keep the rows fixed for at least four quarters, and resist every urge to make the page look better than the quarter was.
If you want a test, hand last quarter’s page to a technology board advisor with no other context and ask them what got worse. If they can find it in under a minute, the page is working.
This article is part of the Executive Leadership cluster, focused on board governance and the operating discipline required to run AI systems responsibly at the executive level.