Klarna’s AI Journey: Navigating the Productivity J-Curve (Mycroft)

Every "AI replaces jobs" headline reads the same on launch day and looks completely different a year later. Tanmay Kulkarni pulls one company's actual numbers to show why: Klarna's AI rollout, the dip that followed its early success, and the framework that predicts the whole shape of that story before it even plays out.

The launch: numbers that looked like an answer

In February 2024, Klarna's AI assistant handled 2.3 million chats in its first month, equivalent to the workload of 700 full-time agents. It resolved 67% of all chats on its own, cut resolution time from 11 minutes down to under two, and was on track for $40 million in savings. On paper, the build-versus-buy question looked already answered in AI's favor.

The fine print behind the headline numbers

The early numbers came with real qualifications. Analysts pointed out that this was the easiest slice of support work: authenticated users, structured data, and common questions, not the harder judgment calls that make up the rest of a support queue. By May 2025, Klarna's own CEO admitted the tradeoff directly, saying the company had focused too much on efficiency and cost, and that the result was lower quality. That admission is the pivot point of the whole story, the moment where the early win started to look more complicated.

The productivity J-curve

There's a name for this shape: the productivity J-curve. Economists Brynjolfsson, Rock, and Syverson documented that when a new technology arrives, the savings tend to show up before the real supporting work, retraining people, redesigning processes, and handling the edge cases, actually gets done. The dip that follows isn't the AI failing. It's the bill coming due for homework that got skipped in the rush to capture early gains.

The hybrid fix

Klarna course-corrected by rehiring humans, reportedly for disputes, refunds, and hardship cases, the kinds of calls that need judgment rather than just speed. Routine, high-volume, well-defined questions stayed with the AI. The complicated cases went back to people. That boundary is worth stating plainly: deciding when a case needs a human, because it's a dispute, a hardship, or a genuine judgment call, is not something that gets automated away. It isn't a bug in the design. It's the design itself.

The results after the correction

By the third quarter of 2025, the hybrid model was producing numbers that held up: the equivalent of 853 agents' worth of work, $60 million in annual savings, response times 82% faster, and a customer satisfaction score of 73. Those figures suggest the dip wasn't a dead end, it was a stage the rollout had to pass through before the combined human-and-AI system reached a more durable state.

Build versus buy was never the real question

Klarna didn't build its own model from scratch; it bought, partnering with OpenAI, and still hit a dip that reads exactly like the productivity J-curve. That's the point worth sitting with: the J-curve doesn't check whether you built or bought your AI system. It shows up either way if the complementary work, the calibration, the retraining, the process redesign, lags behind the rollout. The real question isn't build versus buy, it's whether you're investing in that calibration work alongside the AI or skipping it and paying for it later.

Key takeaways

• Klarna's AI assistant handled 2.3 million chats and 67% of all chat volume in its first month, matching roughly 700 full-time agents' worth of work.
• Klarna's CEO later acknowledged that an early focus on efficiency and cost led to lower quality, a public admission of the tradeoff.
• The productivity J-curve, documented by economists Brynjolfsson, Rock, and Syverson, describes how new technology's benefits appear before the retraining and redesign work needed to sustain them is finished.
• Klarna's fix was a hybrid model: AI handles high-volume, well-defined questions, while humans handle disputes, refunds, and hardship cases that require judgment.
• By Q3 2025, the hybrid approach reached the equivalent of 853 agents, $60 million in annual savings, 82% faster response times, and a 73 customer satisfaction score.
• Klarna bought its AI capability rather than building it, and still experienced the J-curve dip, showing the pattern isn't about build versus buy but about whether calibration work keeps pace with rollout.

Try it yourself

Pick a company or industry you're curious about and run the same check Tanmay describes: look for whether the efficiency story and the workforce story actually share one primary source, and see if the productivity J-curve shows up there too. This kind of source-checking exercise is part of the ongoing Mycroft Financial AI series from the Humanitarians AI Fellows program.