Klarna Rehired Its Humans. That Is the Agentic AI Test

At its peak, Klarna's AI chatbot handled 75% of all customer service chats. That's roughly 2.3 million conversations a month. The company said it did the work of 700 human agents. Klarna cut headcount by 22% during that stretch, down to 3,500 employees. It called this a preview of the future.

Then the company's own CEO walked it back, in public, on the record. Sebastian Siemiatkowski told Bloomberg the quality wasn't there. His words: "Really, investing in the quality of human support is the way of the future for us." Klarna is rehiring now. It's recruiting students, people in rural areas, and dedicated remote workers to answer the phone again. The cheaper option worked, technically. It just didn't work well enough for a brand that depends on customers trusting it with their money.

I keep coming back to something else Siemiatkowski said, because I think it's the sharper insight in the whole story. He wants customers to always know a human is available if they want one. That's a business decision about which parts of the customer relationship can tolerate an agent's mistakes, and which parts can't. It has nothing to do with nostalgia for the old way of doing things. Klarna handles money. A customer who feels brushed off by a bot during a billing dispute doesn't file a polite complaint. They close the account.

Klarna isn't an outlier here. MIT's NANDA research found that 95% of enterprise generative AI pilots show no real financial payoff at all. Most of that spending shows up on a balance sheet as a cost, and nothing else. The tools that actually deliver stay narrow on purpose. IT ticket triage. Procurement work. Coding agents that a person still reviews before anything ships. None of those touch a customer directly at the moment something can go wrong in front of them.

Gartner's broader forecast for agentic AI tracks the same pattern from a different angle. The firm expects more than 40% of agentic AI projects to be canceled by the end of 2027. The reasons it names are cost, unclear ROI, and weak risk controls, not weak models. Projects fail because nobody built an exit ramp before scope grew past what the team could manage. That's the exact trap Klarna walked into, and then walked back out of, at real cost, in full public view.

Developers, of all people, have already made peace with this tradeoff. I think their instinct is the right one to copy. Stack Overflow's most recent survey found 84% of developers now use AI tools regularly. In that same survey, trust in the accuracy of AI-generated code fell to about a third, the lowest level on record. Read those two numbers side by side and they don't contradict each other. They describe the human-plus-agents stack working as it should. Constant use, paired with a standing habit of checking the work before it ships.

That's the real difference between Klarna's first attempt and what developers do every day without thinking twice about it. Developers never handed the agent the whole job and walked away. They kept a human in the loop at the one point where a mistake actually costs something. They let the agent do the rest. Klarna's original rollout removed the human from that same point: the live conversation with a paying customer. The company only found out how much that cost once the quality numbers came in.

I think about this the way I think about every technology that got adopted faster than anyone could test where it actually belonged. The printing press didn't replace every scribe overnight, and it didn't need to. It replaced the copying, not the judgment about what deserved to be copied in the first place. Agentic AI is running into that same boundary now, one industry at a time. Klarna just hit it first, loudly, in front of everyone.

Before your company deploys an agent anywhere a customer can feel the difference, ask the question Klarna learned the hard way. Can we switch this off by noon if the quality isn't there, and will anyone notice fast enough to make that call?

Sources: Entrepreneur/Bloomberg, June 2025 (Klarna reversal, Siemiatkowski quotes) · Gartner, June 2025 (40% project cancellation forecast) · MIT NANDA, August 2025 (95% no-P&L-impact finding) · Stack Overflow Developer Survey, December 2025 (84% AI tool usage, trust decline).

Alex advises enterprise leaders on which agentic AI use cases are ready now and which still need a human in the loop.

book a conversation →