Management tells the board that AI now supports its decisions. Here is a short list of questions I would put to management at the board's next meeting. None of them is technical. You don't need to know how the models work to ask them.
Why a board needs to ask now
An answer from an AI tool arrives finished. It reads well, and it sounds the same whether it's right or wrong. Sounding right is a style, not a signal.
That's the change a board has to follow. AI no longer just carries out instructions. It drafts the analysis and proposes the conclusion, and it's capable enough to be trusted too early. A board that sees only the final paper can't tell which parts a person checked and which parts nobody opened.
This is not rare. KPMG's 2025 research report covers a survey of more than 48,000 people in 47 countries. Of the employees who use AI at work, 42 per cent said they had relied on AI output without evaluating its accuracy sometimes or more often.
The second finding matters more to a board. A 2024 review by Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone pooled 106 experiments on people working with AI. People with AI did better than people alone. But the pair did worse than the better of the two on its own, and most of all on decision tasks.
So adding AI to a decision doesn't make the decision sound. AI can draft the plan. It cannot decide whether the plan is right. The questions below are about the people who decide with it.
Six questions for management
1. Which of our decisions are now made with AI, and who owns each one?
I would ask management to name the decisions AI now shapes, not only the tools bought. Those decisions might concern credit limits or supplier choices. Ask for the list and for one named person who remains responsible for each decision made with AI.
2. What did the person check, and against what?
"I reviewed it" isn't an answer. Ask what was opened and compared, such as the source file or the customer record. A routine reply needs little; a decision that moves money needs the key claim traced to its source.
3. What may AI decide alone here, and where is that written down?
Some decisions can go to a machine within set limits. Others still need a person's approval before anyone acts. If the line between the two is not written down anywhere, it will change without anyone deciding to change it. Ask to see where it's written down, and who agreed it.
4. How often do our people change or reject what the AI proposes, and does anyone look?
People get this wrong in both directions. They follow AI when they shouldn't, and they overrule it when they shouldn't, because both feel safe. A rate close to zero may mean nobody is checking. A very high rate may mean the tool is the wrong one for the work. Either way, someone should look at the number and ask why.
The Infocomm Media Development Authority (IMDA) gives a related reference point for AI agents. Its Model AI Governance Framework for Agentic AI, version 1.5 (20 May 2026), gives tracking the human override rate as one way to audit human oversight. The override rate is how often humans reject or modify agent actions. IMDA presents it as guidance.
5. Who is told when AI was used, and when the answer was uncertain?
A customer, or a colleague who acts on the work, may need to know that AI shaped it. They also need to know when the answer was uncertain. A doubt that stays with the person who saw it can become a firm decision later in the process. Ask whether people say so without being asked.
6. Do we judge a decision by how it was made, or only by how it turned out?
A sound decision can have a bad outcome, and a poor decision can turn out well by chance. If the organisation looks only at results, people learn that an unchecked AI answer is fine as long as nothing goes wrong. Ask how decisions are reviewed when the outcome was good, and not only when it was bad.
The board's own papers
The board decides too. It decides on papers that AI may have helped to prepare. So the same questions apply to the board's papers. Ask which parts AI drafted, and what the author checked before the paper reached you. A board that asks this of its own papers will find it easier to ask it of management.
My record, and what I don't claim
My own record includes HSBC Global Resourcing from 2005 to 2008: 33,000 front-line employees and 10 million customers. At that size, a board sees very few of the decisions made in its name. It learns about the rest through the questions it asks.
AIR APAC, which I founded, publishes a standard for judgment in AI-enabled work. You can read the published criteria there.
Nothing on this site measures, scores or certifies anyone. The questions above are my own framing, not a test. The views here are mine.
I give keynotes, briefings and workshops on this.