What a board should ask its CPO about AI
Twelve questions, the reason each one matters, and what a controlled answer sounds like next to a reassuring one.
Most board conversations about AI are answered before they are asked.
The pattern is familiar. A slide lists the AI capabilities shipped this quarter. Adoption is up. A customer quote sits in the corner. Nobody in the room can tell from any of it whether the company has control of what it has deployed, what it costs, or what happens when it is wrong.
That is not a competence problem. It is a question problem. The questions boards have practiced for twenty years were built for deterministic software and for programs whose failures are loud. AI systems fail quietly. They complete. The output looks plausible. The number gets trusted.
What follows is the list I would hand a director or an operating partner before a product review. Twelve questions across five themes. Each one is written so that the comfortable answer and the controlled answer sound different out loud. None of them requires technical fluency to ask, and none of them can be answered well by a leader who has not done the work.
Strategy
Most AI strategy presented to boards is a list of features with the word AI in front of them. The questions below separate a thesis from an inventory.
Question 1
Which customer outcome gets materially better because we use AI, and which competitor cannot copy it next quarter?
Why it matters. Almost every AI feature shipped in the last two years is a wrapper around a capability any competitor can license on the same terms. If the advantage rests on the model, it is rented. Durable advantage comes from proprietary workflow, proprietary data, or operating knowledge a competitor cannot observe from the outside.
What a good answer sounds like. A good answer names one outcome, names the asset that makes it defensible, and concedes which parts of the stack are commodity. An answer that lists six AI features across four products is an inventory, not a strategy.
Question 2
What did we decide not to build because of AI, and what did we stop doing to fund this?
Why it matters. AI work is usually additive in the plan and subtractive in reality. Capacity comes from somewhere. When nothing was cut, the AI roadmap is being funded by slippage on everything else, and the board will learn that two quarters late through missed commitments rather than through a decision.
What a good answer sounds like. A good answer names the trade. "We deferred the reporting rebuild and moved two of eight squads." An answer of "we absorbed it" means the cost is hidden in the delivery plan.
Unit economics
This is the set of questions most boards skip and most CFOs raise too late. AI features are typically priced per seat and cost per token, which means adoption success can compress gross margin.
Question 3
What does this feature cost us per active user per month, and what happens to that number if usage triples?
Why it matters. Inference cost scales with usage while seat pricing does not. A feature that is comfortably profitable at ten percent adoption can be margin-negative at sixty. The risk is not that the feature fails. The risk is that it works.
What a good answer sounds like. A good answer has a number, a sensitivity, and a threshold at which the pricing posture changes. An answer that the cost is immaterial today is only reassuring if someone is watching the point where it stops being immaterial.
Question 4
Which pricing posture are we running on each AI capability: absorb, seat uplift, consumption, or outcome?
Why it matters. Each posture makes a different bet. Absorbing cost is a retention play. Seat uplift is a positioning play. Consumption transfers variance to the customer and creates adoption friction. Outcome pricing requires measurement the company usually does not have yet. Running all four by accident across a product line is common and indefensible.
What a good answer sounds like. A good answer states the posture per capability and the rule behind it. The strongest version of that rule: absorb inference cost only for a feature that increases retention, never for one that only increases usage.
Question 5
What is our exposure if model prices, rate limits, or terms change on ninety days notice?
Why it matters. A meaningful share of gross margin now sits with two or three vendors whose pricing has moved repeatedly in both directions. This is a supplier concentration question, and boards already know how to ask those in every other category.
What a good answer sounds like. A good answer covers the share of the workload on each provider, whether the abstraction layer allows a swap, and what the swap would cost in evaluation and re-qualification time. "We could switch in a week" is almost never true if evals exist.
Operating model
AI changes what product teams do before it changes what they ship. These questions test whether the organization has adapted or is running the old process faster.
Question 6
Who owns each agent or model-backed workflow the way a product manager owns a product?
Why it matters. Internal agents accumulate the way SaaS subscriptions did, except an unowned agent in an operating position makes decisions rather than sitting idle. Ownership means someone knows what it optimizes for, where human judgment overrides it, and what its failure modes look like.
What a good answer sounds like. A good answer produces a list with names against it. No list is the answer. It means agent sprawl is already underway and nobody is accountable for the aggregate.
Question 7
What changed in how we hire and evaluate product managers this year?
Why it matters. If agents genuinely absorb spec production, the hiring rubric written for spec production is now selecting for a contested skill. Rubrics that have not moved are a reliable signal that the AI narrative in the board deck has not reached the operating floor.
What a good answer sounds like. A good answer describes a concrete change to the rubric, the ladder, or the ratio of product managers to engineers in at least one team, plus what has been learned from it.
Question 8
Where has the operating knowledge that makes our best operators effective been written down?
Why it matters. Agents cannot inherit tacit knowledge. The binding constraint on useful deployment is rarely model capability. It is that the rules the best operator follows, especially the exceptions, have never left their head. Companies that treat this as a documentation chore stay stuck at pilot stage.
What a good answer sounds like. A good answer describes an explicit, versioned body of operating doctrine, including escalation rules written before automation rules. A pilot that succeeded and then changed nothing usually failed here.
Governance and blast radius
The failure mode of production AI is not an outage. It is silent degradation: the system completes, the output looks plausible, and a decision downstream is quietly wrong.
Question 9
Who signs off when the model version changes?
Why it matters. This is the single highest-yield question on the list. "Which model are you on?" invites a vendor name and tells the board nothing. Asking who holds the gate reveals whether production behavior is controlled internally or set by a vendor release schedule.
What a good answer sounds like. A good answer names a person or a role, a staged rollout, and a comparison run against a held-out set before the switch. "It updates automatically" means an unowned production dependency sits upstream of customer outcomes.
Question 10
Which model calls sit upstream of a decision that would embarrass us if it were wrong, and what is instrumented on those?
Why it matters. Risk is not uniform. A drafting assistant carries near-zero blast radius. A model summarizing operational data for a leadership dashboard or classifying a regulated record carries a large one. Most teams log inputs and very few log outputs in a form that detects drift.
What a good answer sounds like. A good answer tiers the calls by consequence, names the tolerances set in advance for the top tier, and admits which tiers are not yet instrumented.
Question 11
What is our reversibility posture when the model is confidently wrong in front of a customer?
Why it matters. Boards are used to asking about prevention. The more useful question in probabilistic systems is recovery: can the action be undone, is there a record of what the system did and why, and how fast can the capability be turned off without taking the product down with it.
What a good answer sounds like. A good answer covers the kill switch, the audit trail, the customer remediation path, and who is authorized to pull each one at two in the morning without waiting for a meeting.
Talent and evidence
The last two questions are about whether the board is being shown reality or a narrative.
Question 12
Which three people would take the AI program down with them if they left, and what is in place because of that?
Why it matters. AI capability inside most mid-market companies is concentrated in a very small number of people, often one. That concentration is rarely on the risk register even though it would be for any other critical dependency.
What a good answer sounds like. A good answer names the concentration honestly and describes what is being done about it: documented doctrine, paired ownership, or a deliberate decision to accept the risk for a stated period.
Question 13
Show me the metric that would tell us this AI investment is not working, and where it stands today.
Why it matters. AI progress is usually reported with adoption and volume numbers that only go up. A leadership team that cannot state its own disconfirming evidence is not running an investment. It is running a narrative, and the board will be the last to find out.
What a good answer sounds like. A good answer offers a falsifiable measure tied to retention, margin, cycle time, or quality, states the current value including when it is unflattering, and names the date at which a decision gets made.
How to use the list
Do not ask all twelve. A board session that turns into an interrogation produces defensiveness and better-rehearsed slides next quarter. Pick three: one from strategy, one from unit economics, and the model sign-off question from governance. Those three between them reveal whether there is a thesis, whether anyone owns the margin, and whether production behavior is controlled internally.
Ask them the same way each quarter. The value is not in the first answer. It is in watching the answer get more specific over three quarters, or watching it stay exactly as general as it was in March.
A capable product leader will welcome most of this. These are the questions they are already unable to get funded internally, and a board asking them out loud is the fastest way to put a governance gate on a roadmap that does not have one. The leaders who resist the list are usually resisting the second question, not the first.
If the answers in your last product review were reassuring and you could not say why, that is the finding. Fractional product leadership exists for the gap between a plausible answer and a controlled one.