Using AI inside an operation without losing the people

For a while, Klarna was the standard example of AI replacing a customer service operation. A year earlier, the company had said its chatbot did the work of 700 human agents. In May 2025 its chief executive told Bloomberg that Klarna was recruiting people again. "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality," he said. He added that it was critical to be clear to customers "that there will be always a human if you want." A spokesperson told Fortune that the policy of not replacing employees who leave was unchanged, and that the company was "very much still AI-first."

It is easy to read that as a story about AI failing. It reads better as a story about what was being measured. Cost was visible and immediate. Quality, and what the customer experienced when the assistant could not help, showed up later and in other places. The same Fortune report cited an IBM survey of 2,000 chief executives in which only one in four AI projects had delivered the expected return.

"Losing the people" can mean three different things, and an operation can lose all three while the AI part works fine.

Losing them

The first is the plain one: roles cut because a model can produce an acceptable answer more cheaply. Sometimes that is the right call. Often it is made on the number that is easiest to see, and corrected when the others catch up. The correction is expensive. People who leave take the operation's memory with them, and the people who stay draw their own conclusions about what the next tool is for.

Losing what they know

The second is quieter and, we think, more important. Some of the strongest evidence on AI inside a real operation comes from a study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond of 5,172 customer support agents, published in the Quarterly Journal of Economics in 2025. Access to an AI assistant raised the number of issues resolved per hour by 15 percent on average. The gains were uneven. Less experienced and lower-skilled agents improved in both speed and quality. The most experienced and highest-skilled agents saw small gains in speed and small declines in quality. Customers were more polite and less likely to ask for a manager, and the working-paper summary reports better employee retention.

The authors' explanation is the part worth keeping in mind: the AI model disseminates the best practices of more able workers. The assistant helped novices because it had learned from experts. In other words, much of the system's value comes from the people who benefit from it least.

That creates a trap. If an operation treats its experienced people as interchangeable once the tool is in place, it erodes the source the tool learned from. Practices change, products change, customers find new problems. Someone has to keep producing the good answers that the next version will spread. An AI rollout that cuts or sidelines those people may look efficient for a year and then stop improving.

Losing their trust

The third is trust. People on the floor can tell whether a tool is there to help them or to measure them, and whether it is being presented as support or as a replacement in waiting. A tool they distrust gets worked around: answers rechecked by hand, notes kept somewhere else, the official system fed what it wants to hear. None of that shows up as a failure. It shows up as a system that is technically live and practically ignored. We have no study to cite for this one. It is an argument, not a finding.

Reducing decisions, not people

WAOF, the technology, product and creative studio behind this publication, runs a café in Monterrey, Atlético, as a retail lab. Our thesis describes the question we study there as where AI can reduce decisions without adding complexity. That framing is useful well beyond a café, because it puts the unit of change on decisions, not headcount. We have written before about where an agent should stop and ask a person. This piece is about the people on the other side of that line, and a few principles follow from the evidence above. They are our position, not a record of results.

Start with the decisions people would happily give up. Looking up a policy, sorting a queue, drafting a routine reply. Leave the judgment calls that people are proud of, at least until they ask for help with them.

Treat experienced people as authors of the system, not its first casualties. If the tool spreads their practices, they should review what it says, correct it, and be recognized for it. Otherwise the best people have every reason to stop sharing what they know.

Measure what Klarna said it underweighted. Quality, customer outcomes and staff turnover belong next to cost from the first month, not after the complaints arrive.

Keep a visible human route, for customers and for staff. Not as a fallback buried three menus deep, but as something people know exists.

And notice who gains the most. If the main beneficiaries are new or less experienced staff, AI belongs in the early weeks of a job, in operations where people are hired constantly. That is a different investment case than replacing the people who already know the work.

The Brynjolfsson study found that the tool made customers kinder to agents. That result is easy to skip, and it may be the most useful one. An operation is not only a set of tasks to automate. It is also the people doing them, and how they are treated, by customers and by the systems around them, shapes whether the work is worth doing.