The Four Questions a Board Should Ask About Any AI ROI Claim

Two 2026 numbers sit uncomfortably next to each other. BCG's AI Radar survey found that roughly 90% of CEOs believe AI agents will deliver measurable ROI this year. IBM's own research found that only about 29% of executives can actually measure AI ROI with any confidence, even though 79% report seeing productivity gains. That's not a technology gap. It's a claim outrunning the ability to verify it, at exactly the moment boards are being asked to approve bigger checks on the strength of that same claim.

The instinct is to assume the fix is board members getting more technical. It isn't, and it can't be — a board's job is oversight, not model evaluation. The real gap is closer to what Writer's 2026 enterprise AI survey found: 58% of executives admit their fellow leaders lack the fundamental AI knowledge to make strategic decisions about it. That's not a call for a crash course in transformer architecture. It's a call for a small number of non-technical questions that work precisely because they don't require any.

Four questions do most of the work. First: which kind of value is this actually claiming — lower cost, more revenue, reduced risk, or a strategic option this opens up later? Most business cases quietly argue only the first and let the room assume the other three are implied. Naming which one narrows what should even be measured. Second: measured against what baseline, and who agreed that baseline was fair before the project started? A number with no pre-agreed starting point is not a result, it's a claim. Third: who owns this number six months after go-live, not who owned the pitch deck at approval? A ROI figure with no owner after launch is a number nobody has to defend. Fourth: what does "this isn't working" look like, concretely, and on what date do we find out? If the answer is vague, the project doesn't have a failure mode — it has a slow fade nobody will be first to name.

None of these four questions requires knowing what a model is, how it was trained, or what architecture sits underneath it. They require only the discipline the CLEAR framework already argues for at the system level — a named owner, a documented baseline, continuous evaluation, not a one-time claim — translated for the reader in the room who will never open a model card and shouldn't have to. The board's job was never to evaluate the AI. It's to make sure someone, specifically named, is accountable for proving the claim was true — and to keep asking until that person can.