An example I made up
The situation. Firdaus Rahim is regional sales director at Batu Jaya Supplies Sdn Bhd, a building materials distributor with its head office in Kuala Lumpur. Field sales cost too much, and he wants to bring the cost down. He asks AntwrenAI, the assistant his team uses, how to do it. Nobody has given him a deadline. The answer arrives in minutes anyway, and it looks as if the work is done.
What the AI produced. "Merging the Penang and Ipoh territories under one manager cuts travel cost by 18%. Based on last year's visit logs and travel claims. I am 80 percent sure."
It's a good answer. The logs are real and the arithmetic is simple, so the travel figure is right.
The check. What makes Firdaus slow down is what the merge would do. It moves customers from one person to another, and once customers are told who their new salesperson is, it's very hard to go back. So before he decides, he asks both territory managers one question: which customers buy only after someone visits them? Between them, they name 14 contractors in Ipoh who place their orders on site, while the salesperson is standing there. None of that is in the visit logs. The logs show where people drove. They do not show why customers bought.
The decision, and who was told. Firdaus splits the decision. The merge has a back-office part (orders, reports and claims) that no customer sees, and a part that moves customers from one person to another. He merges the back-office work now, because that part is easy to reverse. He keeps both managers on their own customers for one quarter. He sets a date at the end of that quarter to review the full merge. Then he tells both managers and the finance lead what he has decided, and what he has not decided yet.
Three months later. The back-office saving is real, and it shows in the monthly accounts. The review finds that a competitor's salesperson already visits sites in Ipoh, and that the 14 contractors are the customers most likely to move. The territories stay as they are.
Everything in this example is made up, including the figures.
What the example shows
Nothing in the AI's answer was wrong. The 18% was about travel cost, and travel cost is what the logs held. The risk was in something the logs did not hold: what the merge would do to customers who buy from a person. It was also in how hard the merge would be to reverse.
That's the pattern I now watch for. Most AI answers I use at once, and I value the time they save. But there are three kinds of decision where I've learned to go slower on purpose, because the answer came fast.
Decisions that are hard to undo
Some decisions can be reversed next week with an email. Others can't be reversed, because customers have been told or a contract has been signed. A fast answer feels the same in both cases. The cost of being wrong is very different.
The test I use: if this turns out to be wrong, could I put things back as they were? Who would I have to tell? If the answer is a list of customers or staff, I slow down.
Decisions about people
An AI answer usually arrives as a saving or a gain. A decision about people also moves work, customers, money or standing from one named person to another. Those people often know things that are not in the data. In the example, the two managers knew which contractors would only buy from someone who visited. Nobody had written it down, so it was not in the data the AI used.
The test I use: does this decision change who does what, or who loses what? If it does, I talk to the people it moves before I decide.
Decisions where the AI agrees with me
This is the one I find hardest. When the AI tells me what I already thought, I feel relief, and I stop checking. A fast answer that agrees with me looks like proof, but it may only repeat my own view, faster. Sounding right is a style, not a signal.
The test I use: would I check this answer as hard if it said the opposite? If the honest answer is no, I check it as hard as I would have checked the opposite.
What slowing down looks like
For me, slower means doing a few things before I commit. Waiting is not one of them.
First, I ask who is affected, and I want names. In the example that meant the two managers and, through them, the 14 contractors. A group name such as "the sales team" hides the person who will lose something.
Second, I split the decision. I decide now on the part that can be undone, and I set a date for the rest. Firdaus took the back-office saving at once. He didn't wait a quarter for a saving that was safe to take straight away. The date matters as much as the split. Without a date, the open part tends to stay open until someone else decides it by default.
Third, I tell people what was decided and what was not. The finance lead heard that the back-office saving was coming and that the territory merge was still open. Saying what is still open stops people from acting on a decision nobody has made.
None of this takes long. In the example, the check was one question to two managers, and the split was one sentence in a message to three people.
Once the AI has answered
The answer will keep arriving faster, and for most work I'm glad it does. For these three kinds of decision, the speed is my signal to look again. Once the AI has answered, a person still has to ask what the decision will do to the people it moves. That person also has to ask how the decision could be put back if it turns out to be wrong.
I give keynotes, briefings and workshops on this.