Board AI Question Pack

Start here. A system without a named owner should not move to the next stage, however good the demonstration.

• □Who is the single executive accountable for the outcome of this system, not just its delivery?
• □Who will operate it day to day once the project team has moved on?
• □Which existing process, team, or role changes when this goes live, and has that owner agreed?
• □If the vendor or key engineer left tomorrow, who could explain and run the system?
• □Who decides whether to expand, pause, or retire it?
• □Which committee receives evidence about this system, and how often?

Oversight should follow what the system can do, not what technology it uses.

• □What can the system read, recommend, draft, decide, or change without a person?
• □Which actions are irreversible or affect customers, money, safety, or employment?
• □Which of those actions require human approval, and who approves them?
• □Is the approval enforced by the system, or does it depend on people remembering to check?
• □Who can stop the system, how quickly, and has anyone tested that stop path?
• □What records are kept of each decision, approval, and action, and for how long?
• □Can the system’s authority be widened later without the board knowing?

Separate what has been demonstrated, measured, claimed by a vendor, and assumed.

• □What has been demonstrated, and on whose data?
• □What has been measured, against what baseline, and over what period?
• □Which figures come from the vendor, and has anyone reproduced them on our work?
• □What does the evidence not show? Which conditions were outside the test?
• □How will quality be measured once it is live, and who reviews those measures?
• □What result would make us stop, and have we written it down in advance?

Assume the system will sometimes be wrong. The question is who notices and what happens next.

• □What are the three most likely ways this system produces a wrong or harmful result?
• □How would we find out: monitoring, complaints, audit, or by chance?
• □What happens to a customer, employee, or counterparty when it is wrong?
• □Is there a manual fallback, and could the team still run it after a year of automation?
• □Which data leaves the organisation, where is it processed, and under what terms?
• □What incidents or overrides will be reported to the board between scheduled reviews?
• □Have we rehearsed a failure, for example in a tabletop exercise?

AI business cases often count freed-up capacity as cash. Test which benefits are real savings and which are options.

• □Which benefits are cash savings, which are capacity, and which are risk reduction?
• □If capacity is freed, what will that time be used for, and who decides?
• □What are the running costs after year one: models, infrastructure, people, evaluation, and support?
• □What does the plan depend on that we do not control: vendor pricing, model behaviour, data access, or regulation?
• □Can the commitment be staged, with the next tranche released only when evidence arrives?
• □What is the cost of waiting six months, and what would we learn in that time?

An AI approval is rarely a one-off. Set the conditions for returning to it.

• □What conditions are attached to this approval, and who confirms they have been met?
• □When does this decision come back to the board, and with what evidence?
• □Which changes in authority, data, vendor, or model would trigger an earlier review?
• □How will we record what we learned, including if the initiative is stopped?
• □Who will tell the board if the original assumptions no longer hold?