An example I made up.
Jamal Hussin is head of customer service at a mobile network operator in Kuala Lumpur. The company has about 1.2 million customers, prepaid and postpaid. Customers on a monthly bill pay a fee under MYR 15 when they pay late. Phone calls from people complaining about that fee take up much of his team's time.
His team uses an AI tool, WaterthrushAI, to review call records. One morning it gives him a proposal:
"Waive late-payment fees under MYR 15 for every customer who calls to complain about one. Call records from the last three months show this would cut complaint handling time by 18 percent. I am 86 percent sure."
For one phone call, saying yes is easy. A customer is upset about a small fee, the agent waives it, and the conversation ends quickly. But the proposal is a rule that would reach every customer on a monthly bill.
So before he agrees, he asks two questions.
The first goes to finance. What would this rule cost across a full year if every customer who could ask for the waiver did ask? The AI's figure was about time saved on each complaint. Nobody had asked it for the total.
The second goes to his own team. Who never phones? Plenty of customers pay late and never complain. Many of them are older, or less at ease on the phone in English. Under the proposed rule, the customers who complain would stop paying the fee. The customers who stay silent would keep paying it.
Jamal doesn't approve the rule for the whole customer base. He agrees to run it for three months in one region. The end date is written down before the trial starts, and the cost limit is agreed with finance. He writes to his director and to the complaints team to say what he's trying, and why he hasn't gone further.
Three months later the result is clear. The fees waived in the trial came to more than the limit agreed with finance. The company ends the trial. In its place, every customer on a monthly bill gets one late fee waived a year, without having to phone. Its total cost could be stated before it started.
The names, the company and every number in this example are made up.
A choice becomes a rule
A decision applied once is a choice. You look at the person in front of you and decide what's fair. Apply the same decision to every customer and it becomes a rule. A rule behaves differently from a choice, and three things change.
The first is the total. One small fee hardly matters on one bill. Added up across every late payer for a full year, it becomes a figure a finance director should see before the rule starts. The AI in my example gave a figure about handling time, and nothing in the example says that figure was wrong. It answered a question about time. The decision also needed an answer about total cost.
The second is who the rule treats differently. A rule that rewards people for asking will keep charging the people who don't ask. In my example many of those were older customers, or people less at ease on the phone in English. The proposal was about people who complain. It said nothing about the people who don't.
The third is how hard the rule is to undo. A waiver given on one phone call is finished when the conversation ends. A waiver given as a rule can become something customers expect, and taking it away can feel to them like a new charge. That's why the end date in Jamal's trial was written down before the trial began, not after something went wrong.
Why AI makes this matter more
AI makes it easy to apply one decision everywhere at once. A tool that reads every call record can turn a sensible answer for one case into a proposal for the whole customer base. The proposal arrives looking finished, with a percentage attached.
AI can speed up existing work. People still decide what needs to change.
So before a proposal like this starts, a person still has to work out how far it reaches. That means asking what it costs across the whole base for a full year, and who else it will reach. It also means deciding, before the start, how the rule will be stopped if it goes wrong, and who will be told.
None of these questions needs a data team. Each one needs a person who asks it before the rule begins, and who is willing to wait for the answer.
Jamal didn't refuse the AI's idea. He ran it on a small scale, with a limit and an end date, where its cost could be seen. Then he used the result to decide what came next. Starting small was itself a decision, and he told the people who needed to know.
Where I saw scale
HSBC Global Resourcing, 2005 to 2008: 33,000 front-line employees and 10 million customers.
Organisations of that size are the setting this essay has in mind.
What I won't claim
The example is made up, names and numbers included, and it doesn't describe a real company or a real case. Nothing here scores anyone or tells you how good a person is at deciding.
I give keynotes, briefings and workshops on this.