What should we tell the board about our AI usage?

Three things: which specific workflows use AI, named precisely rather than described as a general posture; what those workflows cost in tool spend and human review time; and what the team has deliberately chosen not to automate yet, with the reasoning behind each choice. The refusals are what make the rest of the update credible.

How do we answer what is your AI strategy?

By naming production workflows and their guardrails instead of describing an aspiration. Give each one the task, the human checkpoint, and how long it has been running. Then state the rule the company uses to decide what gets handed to a model next — which categories of work are open to it, which are held back, and who signs off on moving something across that line. A strategy question is asking for that rule, not for a list.

What does a credible AI update contain?

Named workflows with real costs attached, a clear statement of what is and is not yet measured, and a list of what has been deliberately left for humans to do and why. It should also stay vendor-agnostic, reporting on decisions and outcomes rather than on which specific tool or model is doing the work this quarter, since that detail will keep changing. The structure should hold steady from one update to the next as well: a trend is only readable if the lines appear in the same order and mean the same thing each time.

How should we report an AI workflow that failed or was rolled back?

In the same detail as one that worked: what the workflow was doing, what went wrong, whether it was fixed, paused, or taken out of production, and what changed in the review controls as a result. Date the change, so the update covers the period since the last one rather than only the current state. A workflow that appears on one update and is missing from the next without explanation leaves the board to reconstruct what happened from an absence.

How to give the board an honest AI update

Guide · Founder · 5 min read · last verified 2026-08-11

Reviewed before publication Editorial board Independent commercial review

In shortA board update on AI should name specific production workflows and their costs rather than describe a general posture or run a live demo, and it should state outright what the team has chosen not to automate yet. Covers why refusals build…

Structure the update around three things: what the company uses AI for, naming actual workflows rather than describing a general posture; what those workflows cost, in tool spend and in the human review time they consume; and what has been deliberately left out — the tasks the team has decided not to hand to a model yet, and why. That third part is what makes the first two believable. A board that hears only wins has no way to tell a considered program from a lucky one. A live demo does not close that gap either; it answers a different question than the one the meeting exists to ask.

Why a demo answers the wrong question

A demonstration answers a capability question: can the tool do this. A board is convened around a governance question: what is running today, under what limits, reviewed by whom, and what happens when it produces something wrong. A screen share can only answer the first, however well it goes. If a workflow has reached production, it presumably passed through some version of a trial first: a bounded test with success criteria set before the first run, a period of watching before anything executed unsupervised. That discipline is its own subject, laid out in how to run an AI agent pilot in marketing; the board update is the summary of what that process concluded, not a repeat performance of the tool that passed it.

Name the workflows, not the posture

"We are exploring AI across the organization" tells a board nothing it can act on. "An agent drafts first-pass outreach copy that a rep edits before sending, and has for two quarters" is a sentence a director can evaluate — it names the task, the human checkpoint, and how long it has been running. Every workflow on the update should reach that level of specificity, and every one of them should trace back to something the company already decided on purpose: which tools are approved, on which tier, and what data rules apply to each. That decision record is what how to write an AI use policy for marketing describes building; a board update with no policy behind it is really just a list of things people have tried. The list has to work in reverse too: a workflow that appeared on the last update and is absent from this one should be named as withdrawn, with the reason, rather than left to disappear from the list.

What it costs

Two kinds of cost belong on the update. Tool spend is the visible one — subscriptions, usage-based fees, whatever line item shows up on the invoice. Review time has to be counted by hand, because no invoice reports it: how many hours a week a human spends checking, correcting, or rejecting what a workflow produced. A workflow whose review cost is not trending down over time is not yet paying for itself, whatever it saved on the tool line. Report both, even though only one of them shows up on a spreadsheet without being asked for.

What gets measured, and what does not yet

State clearly which parts of this program are instrumented and which are still a judgment call, so the board knows which lines it can audit and which it is taking on the team's word. Usage and review time are the easier parts to report with confidence; what a workflow's output changed further downstream is the harder question, and naming that gap outright is worth more than a number that sounds more precise than it is. One metric worth a line of its own, if the company already tracks it, sits a level above internal AI use entirely: how often the company shows up when its own buyers ask AI assistants questions in its market. Whether your company turns up when buyers ask an assistant about your category is heading toward board agendas regardless of this update, as the market metric your board doesn't track yet argues — it belongs here only as one candidate line, not as a replacement for reporting on the company's own AI use. Both are answers to a board's underlying instinct to ask what a company is exposed to, a pattern examined more broadly in what boards ask about marketing; this update is the narrower, AI-specific version of that same instinct.

The refusals are the credible part

State outright which tasks the team has chosen to keep away from automation, and the reasoning behind each choice. Declining to let a model publish directly to a public page without a named approver. Declining to paste customer data into a tool whose contract terms have not been checked. Declining to widen an agent's authority past what its trial run earned it. None of these refusals are permanent — they are current calls, revisited on a schedule — but naming them tells a board something a list of successes cannot: that someone is actively deciding where the line sits, rather than letting adoption run ahead of judgment. Hearing only what a company has adopted gives a board no way to distinguish a governed program from one that has simply not hit a problem yet.

Keep it vendor-agnostic

Report on decisions, workflows, and what they cost and produce — not on which specific tool or model is doing the work this quarter. Tools get replaced; a board update built around a particular vendor's name ages out the moment that vendor is swapped, and worse, it invites the board to evaluate a purchasing decision rather than a governance posture. The underlying models and providers behind any of these workflows will keep changing, sometimes on a timeline outside the company's control, and an update that depends on today's tool staying in place is an update that will need rewriting for reasons entirely unrelated to whether the program is working.

What the board is checking

The board is not grading how sophisticated the AI program looks. It is checking whether someone is minding the risk while the capability grows, and the update that survives scrutiny is the one built to be checked rather than believed: named workflows, real costs, stated limits, and a clear account of what was deliberately left undone.