Every AI proposal that lands at a board meeting gets the same five questions. The proposals that survive have answers in writing before the meeting. The ones that get tabled don't.
If you're scoping a workflow, run it through this filter first. If you're presenting it, put the answers on the second slide.
1. Where does the data go?
The board wants to know: what data is the model seeing, who hosts it, where is it stored, and is any of it being used to train someone else's system. "It's the API" isn't an answer. The shape of the answer is a one-page data flow with the vendor's data-handling policy attached, and a yes/no on training opt-out.
2. Who approves the output?
For workflows that touch customers, regulators, or money, the board wants a named human in the loop. Not a queue, not a Slack channel, a role with sign-off authority. If the answer is "the model sends the email automatically," expect the next question to be about liability.
3. What's the off-switch?
How do you turn this off and revert to the manual workflow without losing data, breaking integrations, or stranding open work-items mid-flight? If pulling the plug takes a sprint, the board hasn't been given an off-switch, it's been given a one-way commitment.
4. What's the cost ceiling?
Token costs scale with usage in ways that are easy to model badly. The board wants the worst-case monthly bill, the alerting threshold, and what happens when the threshold is hit. "It depends on volume" without a ceiling is the kind of answer that makes CFOs ask for a 12-month moratorium.
5. What happens when it gets it wrong?
The model will be wrong sometimes. The board wants to know: how often, with what severity, who catches it, and what the cleanup looks like. A workflow without an answer here is a workflow that will be paused the first time it embarrasses someone.
Boards aren't trying to block AI work. They're trying to find out whether you've done the work that lets them say yes.
Answer the five before you ask. The decision usually takes care of itself.