An example I made up
The situation
Vikram Shah is the regional sales director of an office equipment distributor operating across several countries. The regional managing director wants his recommendation by Friday on one way of handling orders for every country to use next quarter. If he recommends a change, the order teams in Indonesia and Thailand will spend next quarter learning a new method. Hafiz Osman manages the country team whose order method is under review.
What the AI produced
Vikram asks ShrikeAI, the company's planning assistant, for a first draft of the recommendation. This is its answer:
"Use Hafiz's team's order method in Indonesia and Thailand next quarter. Malaysia's on-time delivery was 99 percent this quarter, the highest in the region. I am 91 percent sure."
The figures are made up, like the company. In the example, the delivery figure is correct. ShrikeAI read it accurately from the order system, and the order system is the only data it was given.
The check
Before he recommends anything, Vikram asks Hafiz one question: how did your team reach 99 per cent? Hafiz tells him. When an order is going to be late, his team calls the customer early and agrees a later date. Then they enter the later date in the order system as the promised date. Nobody asked them to work this way. The team has done it for more than a year, as its own way of keeping customers informed. The order system holds only the latest promised date, so each delivery is counted against the date agreed after the delay. The date the customer was first given is held only in the order confirmation sent when the order was placed. ShrikeAI was never given those. So "on time" in Malaysia's figure can mean late against what the customer was first told. Vikram asks for this quarter's deliveries in all three countries to be checked against the first order confirmations. Against those first dates, 81 per cent of Malaysia's deliveries arrived on time.
The decision, and who was told
Vikram recommends that Indonesia and Thailand keep their own order methods for another quarter. He tells Hafiz that the early calls to customers are worth keeping. In the regional review, he adds a second delivery figure for every country: on time against the date the customer was first given. The new figure measures deliveries. It does not rate any person. He tells the regional managing director why he decided this way. Then he tells the three country managers the same thing.
Three months later
The regional review now shows both delivery figures for every country. Against the dates customers were first given, Malaysia is no longer the highest in the region. Hafiz's team still calls customers early when an order will be late. Nobody can say what that order method would have done in Indonesia and Thailand. Vikram can say what he knew when he decided, and why he decided as he did. That's the fair way to judge a decision: by what was known at the time, not by how it turned out.
Why I read it that way
When I was 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. Results are not the same as progress. That's why I'd have asked Hafiz the same question Vikram did.
What I ask of a good number now
When someone shows me a number that looks good, I ask three questions before I act on it.
What did it cost the people who produced it? A sales figure records what was sold. It doesn't record how the people behind it got there, or what it cost them. I want to know about those costs before I recommend copying the method.
What is not in the data behind it? Every figure comes from a system, and every system leaves something out. ShrikeAI read the order system correctly. The dates customers were first told were held in order confirmations it never received. So I ask where a number came from, and what that source can't see.
Who needs to hear before I act on it? A good number is passed on quickly. It can reach a leadership meeting before anyone has asked the country manager what it means. Before I act, I ask who knows something the number doesn't show, and who will have to carry out the decision if I'm wrong. In the example, that was Hafiz first, then the regional managing director and the other country managers.
Why AI makes this sharper
An AI-generated figure can sound certain before anyone has checked its basis. It comes with a reason and a percentage beside it. ShrikeAI's "91 percent sure" told Vikram how sure the tool was. It told him nothing about what the tool had not seen.
Sounding right is a style, not a signal. The tool did what it was asked, with the data it was given. Vikram still had to ask how the number was made and check the first dates customers received. AI can draft the plan. It cannot decide whether the plan is right. The person making the recommendation still has to check how the number was produced and what evidence was left out.
What I won't claim
I won't claim that these questions guarantee a better result. Vikram's decision may have delayed a better way of working in Indonesia and Thailand, and nobody will ever know. What the questions give you is a decision you can explain afterwards: what you knew when you decided, and who you told.
I also don't always ask them in time. I still move too fast, and I still default to outcomes. The work, for me, is noticing earlier and listening longer.
I give keynotes, briefings and workshops on this.