Klarna cut 700 support jobs for AI, then hired humans back

In 2024, Klarna told the world its OpenAI-powered assistant was doing the work of 700 customer service agents. Resolution time had fallen from 11 minutes to under 2. The company had already cut headcount from roughly 3,800 to 2,000, and its CEO called it the future.

By 2025 that same CEO was walking it back. Sebastian Siemiatkowski told Bloomberg the company had leaned too hard on cost and efficiency, that quality had dropped, and that customers would now always have the option to reach a human. Klarna started rebuilding the support team it had taken apart.

It is easy to read that as “AI failed at customer service.” That is the wrong lesson, and it is the expensive one.

What actually went wrong

Klarna did not stumble because AI cannot answer support questions. It answers plenty of them well, and fumbles others. It stumbled because it turned the AI loose on everything at once, including disputes, fraud, and high-stakes money questions, with no reliable way for the system to say “I am not sure enough about this one, send it to a person.” One post-mortem of the reversal put it plainly: there was no confidence-based escalation, no governance, no sense of which topics the AI had earned and which it had not.

So the assistant answered the easy refund question and the frightening fraud question with the same confidence. Customers noticed. On money, they always do. A Gartner figure that made the rounds this year found that roughly two thirds of customers would rather companies not use AI for support at all, and four in five would wait to reach a human. Those numbers are not really about model quality. They are about trust, and trust is what Klarna spent.

The fix is not “add a human button”

The common takeaway from Klarna is “keep a human in the loop.” True, but too small. Bolting a human option onto an AI that still answers everything just means a person cleans up the mess after it reaches the customer.

The better model is older than AI. You would never hand a new support hire every account on their first morning. You let them shadow the team, take a few easy tickets, and earn the harder ones as they prove out. Software can work the same way. Let the AI draft in private first, on your real tickets, scored against what your team actually sent. Watch where it is strong and where it is not. Then promote it one topic at a time. Shipping questions might go fully automatic. Refunds might get a human glance before they send. Disputes and fraud stay with people until the drafts are good enough to trust, and for some topics that day never comes, which is a perfectly good answer.

Two things make that work, and they are exactly the two Klarna was missing. The AI has to know its own confidence and hand off when it runs low. And autonomy has to be a dial you turn per topic, not a switch you flip for the whole queue.

The part that never makes the slide

There is a second cost in this story that the efficiency math leaves out. Those were real jobs. When a company automates recklessly, it does not just gamble with customer satisfaction. It puts people out of work for a version of automation that did not even hold up. That is a bad trade twice over.

We build IMCeleste on a plain belief: automation of support work is coming either way, and the companies that do it well will be the ones that let their AI earn trust instead of assuming it. Celeste learns from your own support history, practices where it cannot hurt, and goes live only on the topics you decide she has earned. You can see how it works, or read why we commit the greater of 5% of our revenue or 70% of our profit to the people this technology displaces.

Klarna’s real lesson is not that AI has no place in support. It is that AI has to earn its place, one topic at a time. The companies that learn that before they cut will not have to hire everyone back.