Six months ago, it was tempting to judge AI progress by what could be shown in a meeting: pilots launched, people enabled, usage rising, new capabilities demonstrated. Those are signs of activity. They do not establish that the investment paid off, that costs are controlled, or that management understands what has moved into production.

In 2027, boards should ask for an answer that is smaller and more demanding: which work changed, what did the AI-delivered portion cost, what evidence supports the claimed outcome, and what will management do differently as a result?

Do not let the meeting reward the wrong unit

A model call is not a business outcome. A pilot is not a return. A budget is not proof that spend is controlled. The board should ask management to name the unit that connects AI activity to the business: a workflow completed, a customer served, a decision supported, or another unit that already matters to the operating model.

The unit will differ by use case. That is not a flaw. The flaw is presenting one broad enterprise number as if it answers every question. Internal AI tools belong in a discussion of operating expense, use, governance, and quantified impact. AI embedded in a product belongs in a discussion of cost to serve, pricing, and gross margin.

A useful follow-up: Which part of this number is reconciled to provider invoices, which part is allocated, and who owns the gap between the two?

From oversight question to management record

Oversight becomes practical when each question has a counterpart in the operating record. Ask who owns the workload, where its cost comes from, what outcome is being assessed, what changed since the prior review, and what decision is now required. A response without an owner is a status update. A response without a boundary is not comparable. A response with no next decision is not oversight.

Boards should resist prescribing a specific dashboard or model. Those choices belong with management. The board's job is to make the evidence standard explicit and ensure that material uncertainty is visible. Ask what management does not yet know, what would make its current conclusion wrong, and when the answer will improve.

Six months later, demand fewer claims and better proof

Compare the AI update you would have accepted in April with the one you need now. If the old update led with adoption, ask what adoption changed. If it led with projected savings, ask how the baseline was set and whether the work actually moved. If it bundled all cost together, ask which pool belongs in product margin and which belongs in operating expense.

The point is not to punish teams for incomplete attribution. It is to make incompleteness governable. A management team that can state its blind spots, assign owners, and return with a better record is more useful to the board than one that presents false certainty.

Make the evidence earn its place on the agenda

For every AI item, identify the decision the board is being asked to make. If there is no decision, consider whether the update belongs in a different forum. If there is a decision, require the evidence and uncertainty necessary to make it responsibly.

The 2027 board question is not whether the company is using AI. It is whether management can explain the cost and consequence of the work that matters, and whether the explanation is strong enough to guide action.