I'll start with an example I made up. The people and the company in it are invented, and so are the figures.
A clean report, two days before the decision
Benjamin Ong is regional head of marketing at a consumer goods distributor. He has to decide whether a new pack design goes to all six markets. A pack change is slow and costly to reverse once it is on the shelf, and the decision is his.
The new pack has been on trial for eight weeks in 24 stores, four in each market. Two days before the decision, Aisyah Rahman brings him the trial report. It is clean and finished. The summary is short and the charts are tidy.
The comparison came from LeafbirdAI, an AI tool the team uses to analyse store sales. Its line in the report reads:
"The new pack raised sales by 18 percent. Based on eight weeks of sales in the 24 trial stores, compared with 96 stores that kept the old pack. I am 91 percent sure."
Nothing on the page looks wrong. The line even gives its basis and says how sure it is.
The question Benjamin asked
Benjamin doesn't ask whether the report is ready. He asks Aisyah how she decided that the new pack raised sales. Then he asks how the 24 trial stores were chosen.
She didn't know. She had checked the sales figures, and they were correct. The result matched what the team expected, so she had trusted it.
She asks the country teams that afternoon. Each team had picked its four strongest stores, with the most shelf space and the most experienced staff. Benjamin asks for the same stores' sales in the eight weeks before the trial. Before the new pack arrived, those 24 stores were already selling 16 per cent more than the other 96.
So most of the 18 per cent was there before the trial began. The comparison showed how the stores had been chosen, not what the new pack had done.
What he decided, and who he told
Benjamin puts the decision on hold for eight weeks. He asks for a second trial, randomly assigning stores within each market to the new or old pack. The team checks the groups' starting sales.
Then he tells the product director and the six country managers what was found, and why the decision waits.
Three months later
The second trial shows a smaller rise in sales, about 4 per cent. Benjamin sends the new pack to all six markets, with a sales forecast built on the smaller figure. Every comparison the team prepares now says how the stores in each group were chosen.
The report looked the same either way
Look at what the example shows. The report Aisyah brought would have looked exactly the same whether or not anyone had asked how the trial stores were chosen. The charts and the summary would have been identical. Only the decision behind them was different.
A finished piece of work hides the decision behind it. Someone still has to explain what they checked before using the output. AI makes this harder, because its output so often looks finished and sure. Sounding right is a style, not a signal.
So when a leader judges the report, they are mostly judging its style. When they ask how it was decided, they see the decision itself. I think leaders should ask "is it ready?" less often, and "how did you decide?" more often.
Five questions to ask instead
These are plain questions. A leader can ask them at the start of a meeting or in a short conversation at someone's desk. They don't need any knowledge of how the AI works.
- What were you deciding? The report may answer a wider question than the one the person actually decided. Ask them to name it.
- What did you check, and what did you leave? Nobody checks everything. What matters is whether the person knows what they left unchecked, and whether leaving it unchecked was reasonable, given how much depended on the decision.
- What did the AI say, and how sure was it? Ask for the AI's own words. Then ask whether the person agreed with how sure it was, and why.
- Who needed to know? Some decisions need another person to be consulted or told before they go further. In the example, that was the product director and the six country managers.
- What would change your mind? If the answer is "nothing", the person may have accepted the output without deciding anything.
In the example, the second question would have been enough on its own. Aisyah would have said she had checked the sales figures, but not how the stores were chosen. That is an honest answer, and a leader can act on it.
People can trust AI too much or too little
These questions are not only for catching an AI mistake. People fail both ways with AI. They follow it when they shouldn't, and they overrule it when they shouldn't, because both feel safe.
Following the tool feels safe because the tool sounds sure. Overruling it feels safe because it looks careful. A finished report hides both. If someone rejected a correct answer from the AI, ask them to explain why. Asking how they decided is how you find out.
Why I care about the decision behind the result
I learned an early version of this myself. At twenty-nine I delivered 241 per cent sales growth at Electrolux. I was dismissed within the year, because the result had broken the team that produced it. A result is not the same as progress. In the same way, a finished report is not the same as a good decision.
That is why my question for leaders is about decisions, not output. Do you know how your people are deciding with AI? Someone still has to explain the decision behind the finished report. A short conversation about how it was decided is a good place to start.
What I won't claim
These five questions are not a test, and they score nobody. Asking them is how a leader sees the decision behind a piece of work, nothing more.
Deciding with AI is what I speak on. I give keynotes, briefings and workshops on this.