Klarna Replaced 700 Support Agents With AI, Then Rehired Humans

For about a year, Klarna was the company every AI vendor pointed to. The Swedish fintech had done what everyone else was only talking about: it deployed an AI customer service assistant, froze hiring, and said the bot was doing the work of 700 full-time agents. Its CEO went on to say, in public, that AI could already do essentially all the jobs its people did. It was the clean proof that the future had arrived.

Then, this spring, the same CEO went on Bloomberg and said three words that ought to be studied by every business owner considering the same move: "We went too far."

This is not a story about AI failing. Klarna's bot worked. It's a story about aiming AI at the wrong goal, and it's the most useful case in this series for anyone deciding what to automate, because the mistake is subtle, it looked like success for a year, and it's the one we most often talk clients out of making.

What Klarna did

At the end of 2023, Klarna froze most hiring. The plan was to let attrition shrink the workforce while an OpenAI-powered assistant absorbed the customer service load. By the following year the numbers were striking: the assistant was handling roughly two-thirds of customer service chats and doing work the company equated to 700 agents. Klarna's total headcount fell from around 5,000 to about 3,500. The story wrote itself: AI had cracked customer service, and Klarna was the company brave enough to prove it.

The framing was aggressive and deliberate. Klarna didn't say AI was helping its support team. It said AI was replacing it. That public positioning is exactly what made the reversal so visible when it came.

Where it broke

The bot was genuinely good at the bulk of the work. Routine questions, order status, simple account issues, the high-volume, low-complexity queries that make up most of any support inbox, it handled fast and at scale. On those, customers were fine.

The problem was the other slice: the disputes, the fraud claims, the account escalations, the moments where a customer was already stressed and the stakes were real. Those are a small share of total volume but a huge share of the value at risk, and they're exactly where a scripted, average-case system struggles. Satisfaction on the hard interactions slipped. And because those are the interactions that shape how a customer feels about the whole company, the damage didn't stay contained to a metric. It reached the brand.

That's the word the CEO himself kept returning to. He said the drive for efficiency and cost had produced lower quality, and that it wasn't sustainable. But the line worth pinning up is this one: he said that from a brand perspective, it was critical that customers always know a real person is available if they want one. The head of a company that had bet its reputation on automating its support staff away was now saying, in public, that guaranteeing a human was essential to the brand.

What the reversal actually was

It's worth being precise, because the lesson lives in the precision. Klarna did not rip out its AI and go back to a call center. The assistant stayed on the front line for high-volume routine work, where it was winning. What changed is that Klarna started bringing humans back specifically for the complex, high-stakes cases, moving to a hybrid model where AI does what it does well and a person steps in where the cost of getting it wrong is high.

In other words, the correction wasn't "AI doesn't work." It was "we pointed it at the whole job instead of the right part of the job." The original error was treating AI as a substitute for the support team when it was really a complement to it, powerful on the routine majority, unreliable on the consequential minority.

The trap, and why it's so easy to fall into

Here's what makes this case dangerous rather than just instructive: for a year, the wrong strategy looked exactly like the right one.

The volume metrics were real. The cost savings were real. The bot really was handling two-thirds of chats. Every number on the dashboard said the deployment was a triumph, because the dashboards measured volume and cost, the things AI improved, and not the quality of the interactions that actually determined customer loyalty. The warning signs were in the satisfaction data on complex cases, and early on they were easy to explain away.

This is the shape of the mistake: automate for headcount reduction, measure success by how many humans you removed, and you will hit that target while quietly damaging the part of the business that doesn't show up in the token count. The goal was wrong, so the metrics that looked like success were measuring the wrong thing.

How to aim AI at the right goal

Automate the work, not the people. The right question is never "how many roles can this eliminate." It's "which tasks can this handle well, and which still need a person." Aim at the work and you naturally get the hybrid model Klarna backed into after a year of brand damage. Aim at the headcount and you get Klarna's first year.

Separate the routine majority from the consequential minority. In almost every process, a large share of the volume is low-stakes and a small share carries most of the risk. AI belongs on the first. A human belongs on the second, or at minimum in the loop. The engineering task is drawing that line accurately, not moving it as far as it will go.

Measure the thing that actually matters, not the thing that's easy to count. Volume handled and cost saved are easy to measure and will look great almost immediately. Customer satisfaction on the hard cases is harder to measure and is the number that actually predicts whether the deployment is helping or hurting. If your only dashboards track the easy metrics, you'll declare victory right up until the brand damage surfaces.

Keep the human door open, and say so. Klarna's hardest-won lesson was that customers need to know a real person is reachable. A visible, easy path to a human isn't a failure of your automation. It's the thing that lets customers trust the automation for everything else.

The real lesson

We build AI into customer-facing and operational workflows for a living, and the first conversation with a client is almost always about this exact trap. The pitch that sells AI internally is usually a headcount number, and that number is the wrong target. The businesses that win with AI use it to make their people dramatically more effective, handing the machine the repetitive majority so the humans can focus their judgment where it counts.

Klarna is not a cautionary tale about the limits of AI. The bot did its job. It's a cautionary tale about what happens when you decide, up front, that the point of AI is to remove people rather than to do work. Klarna spent a year and a chunk of its brand learning the difference. The lesson is available to everyone else for free.