TL;DR: Klarna ran the most public AI rollout in business history. The CEO fired 700 customer service workers, replaced them with an OpenAI chatbot, claimed AI could do every human job, and pocketed roughly $10 million in initial savings. Two years later he was on Bloomberg admitting the AI produced “lower quality” service, and the company started hiring humans back. More on the augmented approach. The valuation fell from $45 billion to under $7 billion during the same period.
This is the cautionary tale every executive in 2026 should already know by heart, and most of them are about to repeat it anyway.
Every executive I’ve talked to about AI wants the Klarna headline without the Klarna cleanup. Firing the people who know how the work gets done is the fastest way I know to break a business.
In December 2022, Sebastian Siemiatkowski took the stage at the World Economic Forum and announced that artificial intelligence could already do every job humans do.
He was the CEO of Klarna, a Swedish buy-now-pay-later fintech that had reached a $45.6 billion valuation the year before. He’d also just signed a deal with OpenAI to deploy ChatGPT across his company. That announcement kicked off the most public AI rollout in business history, and the most public faceplant.
The worst part of Klarna is the swagger of it. A CEO told the world AI could do every human job, and 700 customer service positions went away on the strength of a benchmark that only measured the easy work. Follow the timeline in order, because the order is the lesson.
February 2024: The headline number
Klarna issued a press release announcing that its OpenAI-powered chatbot was doing the work of 700 full-time customer service agents.
The release was full of impressive metrics. The bot handled two-thirds of all customer service chats in its first month. Average resolution time dropped from 11 minutes to under 2. Customer satisfaction scores were “on par” with human agents. Estimated profit improvement: $40 million in 2024. For more, see six places AI will break your work and the AI skills that will actually matter in five years.
Deploy AI to deflect the routine 70 percent of tickets and keep your humans for the hard 30 percent.Share on X
The press release didn’t say the 700 agents had been fired. It used the phrase “the work of 700 agents.” But the company had stopped hiring through 2023 and let attrition do the visible part of the work. The result was the same.
I find that wording slippery. “The work of 700 agents” lets a company take credit for the cut without ever owning it. Those were jobs held by people who needed paychecks, and the release never said what happened to them.
The market loved it. Klarna’s IPO conversation accelerated. Every executive on every earnings call started getting asked when they’d do what Klarna did.
May 2024: The boast
Siemiatkowski went on a press tour. He told Bloomberg AI was “doing the equivalent of 700 full-time agents.” To the Financial Times, he said the company had “stopped hiring entirely a year ago” and that AI was the reason. He told a podcast he believed AI could “do all of the jobs” humans currently do.
Anyone who’s ever shipped a real production system would’ve gotten nervous reading those quotes.
Underneath the press release, something else was going on. The bot was handling the easy 70 percent of customer service tickets. Then there was the hard 30 percent. The angry customers. The billing disputes the bot couldn’t resolve. The calls where a real human relationship had to be repaired. Those were getting routed somewhere with fewer humans available to handle them than the year before.
The angry customers were louder on Twitter, and on Trustpilot, and inside Klarna’s own NPS scores. But the press release was the story the market heard.
I’d have been yelling in those meetings if I’d worked there. A dashboard full of fast resolution times tells you nothing about the customer who got a fast wrong answer and took their business somewhere else.
February 2025: The reversal begins
One year after the original announcement, Bloomberg ran a follow-up. Siemiatkowski admitted that Klarna had “gone too far” with AI in customer service. He said the company was “starting to hire humans again.” He used the phrase “lower quality” to describe what the AI had been producing.
The reversal got less coverage than the original announcement. That’s how these things always go. Every executive read the February 2024 headline. Almost none of them read the February 2025 one.
The company didn’t disclose what the rehiring cost. They didn’t have to. Anybody who’s ever run a customer service operation can do that math in their head, and it made me wince when I read the announcement. You don’t get back the 700 people you let attrition handle. You hire new people, who don’t know your products, your customers, your edge cases. You train them. That takes months. During those months, your service quality stays bad.
By the time the new team is up to speed, you’ve spent the original “savings” several times over. Your brand has also spent a year telling customers there’s no human to talk to.
April 2025: The valuation
Klarna’s IPO finally happened. The valuation came in at $6.7 billion. Down from $45.6 billion in 2021.
The decline wasn’t entirely about the AI story. Rising interest rates hurt the whole BNPL sector. Bad debt rose. Competition got worse. The AI rollout and the year of customer service chaos that followed did real damage to the brand, and the valuation shows it. Investors who had bought the original story were now reading the second one.
What did Klarna get wrong with AI?
Using AI was fine. Customer service is a sensible place to point a tool like that. Klarna’s mistake was confusing two completely different questions.
The first question is: can AI do this task? The answer for customer service is yes, for about 70 percent of tickets. The bot really did handle them as well as or better than a human.
The second question is: can AI do this job? The answer is no, because the job isn’t 70 percent of the tickets. The job is all of them. That includes the 30 percent where the customer is angry, the billing is complicated, and the relationship is on fire. A human has to walk into the room and put it out.
Siemiatkowski answered the first question and acted as if he’d answered the second one. That’s the entire disease, and it’s playing out right now at thousands of companies you haven’t heard of yet because their reversals haven’t made the news.
What pattern breaks every AI rollout?
Companies benchmark AI against the easy part of a job, replace the people, and then have no one left to handle the hard cases when they arrive.
From the podcast
Ben “The Automator” Christensen watches companies do the arithmetic backwards, on Leaders and Their Stories:
A lot of companies are just implementing AI saying, oh man, we’re going to save twenty or thirty percent on overhead, because we won’t have this cost anymore, which is usually employees.He’s automated more than three hundred full-time positions, and that makes his caution about this worth hearing.
I’ve watched this exact pattern across forty years of building automated systems, long before anyone called it AI. The shape never changes.
Someone with the budget gets excited about a new technology. They benchmark the technology against the easy version of a job. The benchmark looks good. They roll the technology out to handle the full job, fire the humans, take the savings, claim the win.
Then the hard cases start arriving. The system that handled the easy 70 percent has no answer for the hard 30. The humans who would have caught the problem aren’t there anymore. The hard cases pile up. Customers notice. Reviews drop. Competitors win the relationships.
The executive who took the savings has usually moved on by the time the bill arrives, with the win still on their record. The company eats it, and the employees and customers eat it first. Then the next company reads the press release, skips the reversal, and lines up to repeat the mistake. I think that’s the ugliest part of the whole playbook.
What does a successful AI rollout look like?
The version that works runs the same play in reverse. You start with the easy 70 percent and you keep the humans. The bot deflects the routine tickets. The humans handle the hard ones. And the humans now spend all their time on the work that was always the actual job.
Customer satisfaction goes up because the angry customer no longer waits in a queue behind someone asking about shipping. The humans aren’t burned out anymore. The savings are real, but they’re smaller than the press release claimed, and they show up in keeping customers and reputation. The layoff column is where the mistake is.
That’s the rollout I’d back every time. It makes a duller press release, and nobody gets invited to the World Economic Forum to brag about it, but the customers stay and so do the people who know how to keep them.
I lay out the full version of this in Augmented Beats Replaced. Every Time. If you’re sitting on a decision about an AI rollout right now, read that one before you sign anything.
The cost of the Klarna playbook
Klarna’s reversal is one company. The cost of every executive who reads the original press release and not the follow-up is going to be paid by their employees, their customers, and eventually their shareholders, over the next decade.
The companies that figure out the difference between “AI can do this task” and “AI can replace this job” will own the next decade. The ones still running the Klarna playbook will spend the same decade explaining why the savings never materialized.
Sebastian Siemiatkowski wrote the cautionary tale in public. The only thing left is whether anyone reads it. This is the failure mode The Death of Thinking is about. What happens to a company, and a culture, when the loudest voices in the room have outsourced the thinking part of the work. Use AI to take the drudgery off good people, and keep the people. Klarna cut first and asked questions later, and it’s hiring back now at a cost it never disclosed. An executive who copies that playbook in 2026, with the whole reversal on the public record, has no excuse, and I hope their board asks them why.
