Seven hundred.
If you have sat in a boardroom in the last eighteen months and watched somebody make the case for automating a service function, you have met that number. Klarna’s AI assistant, launched with OpenAI in February 2024, handled 2.3 million conversations in its first month – about two-thirds of the company’s customer service chats – and the company said that was the workload of roughly 700 full-time agents. It put the profit improvement at about $40 million for the year. Average resolution time fell from eleven minutes to under two.
It was a good number. It traveled beautifully. It is probably the single most cited data point in enterprise AI, and it has been doing duty in slide decks ever since as proof that the thing works.
Then, in May, Sebastian Siemiatkowski told Bloomberg the company had gone too far, that the focus on cost had produced lower quality, and that Klarna was hiring human agents again. And the number went right on traveling, now in the other direction, as proof that the thing doesn’t work.
Both readings are wrong, and the second one is doing more damage than the first.
What the number actually said
Klarna never laid off 700 people. This matters more than it might seem, because almost every retelling assumes it did.
What happened is that Klarna stopped hiring in September 2023 and let attrition run. Headcount went from around 5,000 to somewhere near 3,000 over roughly two years, and Siemiatkowski has told CNBC that amounts to a workforce about 40 percent smaller. The AI deployment made the freeze survivable. The freeze did the headcount math. Those are different claims, and the company’s own framing was consistently the more careful of the two – the 700 figure described work the assistant absorbed, not people who were shown the door.
But Klarna also did nothing to slow down the louder version. Siemiatkowski spent 2024 saying things like AI can already do every job humans do, including his own. He told Sam Altman he wanted the company to be OpenAI’s favorite guinea pig, which is a wonderful line and also a commitment you have to keep paying for.
So when the correction came, it landed against the loud version rather than the careful one. That is the cost of letting a good headline run.
The walk-back was narrower than the coverage
Read the May comments closely and Klarna did not retreat from AI in customer service. The assistant is still there, still handling the high-volume routine tier, still working across two dozen markets in more than thirty languages.
What came back was humans on the hard cases. Disputes. Complex refunds. Financial hardship – which, at a buy-now-pay-later company, is not an edge case but a core scenario, and one where a wrong-but-fluent answer costs you a regulator rather than a CSAT point. Siemiatkowski described the goal as a customer always being able to reach a person if they want one, and later at SXSW London called human support something closer to a VIP tier: the expensive thing you offer deliberately, not the cheap thing you couldn’t afford to keep.
The hiring model is its own story. Rather than rebuilding a call center, Klarna went for a distributed, freelance-style arrangement – remote agents working flexible hours, recruited from students and rural areas, which Siemiatkowski compared to Uber. Whether that produces the quality he says he’s buying is a genuinely open question and nobody has data on it yet.
This isn’t “AI failed.” It’s a scope correction, executed loudly by a company that had made the original scope loud.
The part everyone skips
Here is what I think the actual lesson is, and it has nothing to do with humans versus machines.
Klarna instrumented throughput and cost. Conversations handled, minutes to resolution, repeat contacts down 25%, dollars saved. Every one of those metrics came back green, and they came back green fast, which is what made the February 2024 announcement possible one month into the deployment.
Quality on complex, emotionally loaded interactions is a different kind of measurement. It moves slower, it’s noisier, it needs segmenting by ticket type before it tells you anything, and a company that is measuring aggregate CSAT will watch the routine tier’s improvement mask the hard tier’s decline for quite a while.
So the system did exactly what it was measured on. It just wasn’t measured on the thing that eventually forced the reversal. The failure was in the instrumentation, and it happened before the model ever ran.
Which should be uncomfortable reading, because most retailers evaluating service automation right now are building the same dashboard Klarna built. Deflection rate. Cost per contact. Handle time. I have yet to see one that segments satisfaction by complexity tier and tracks it as a leading indicator with the same seriousness.
The IPO clock explains the rest
None of this happened in a vacuum, and the vacuum is usually the missing piece in the retellings.
Klarna was valued at $45.6 billion in 2021 and then spent the next few years living down that number. The AI story arrived in early 2024 in the middle of a long, public repositioning of the company from a buy-now-pay-later lender into an AI-native digital bank, and it was excellent investor communication. A fintech that can show structurally falling operating cost is telling a much better story than one that can only show growth.
The F-1 went in on March 14 of this year. Then April’s tariff turmoil froze the US IPO market and Klarna paused. On July 31, Bloomberg reported the listing was being revived, possibly as soon as September.
So the timing of the May admission is worth sitting with. You do not usually volunteer that you cut too deep in the weeks before you go raise money – unless you have concluded that saying it yourself, early, is cheaper than having it come out of a diligence process or a journalist’s inbox during the roadshow. Getting the correction on the record in May, from the CEO, in an interview he controlled, is not the behavior of a company caught out. It’s a company clearing the deck.
The Q2 numbers are the other half. Revenue of $823 million, up 20% year over year, and a first-half net loss of $152 million against $31 million a year earlier, driven by restructuring and expansion spend. Growth is real. Profitability is not, yet.
Which sets up the interesting thing about the next few weeks: investors have to price a company whose AI story is now unavoidably two stories. The efficiency gain was real and is still running. The overreach was also real and cost something to fix. Whether the market treats the admission as a maturity signal or a credibility problem is, as far as I can tell, the first time anyone will put an actual price on an AI walk-back.
What travels to retail
Klarna is a lender, not a merchant, and the mapping isn’t exact. But the service function is close enough that most of it carries.
Split your contact taxonomy before you automate, not after. Klarna’s assistant was excellent at the top of the distribution and unreliable at the bottom, and the aggregate number hid the difference. If your deflection rate is a single figure, you don’t know what your AI is doing.
Decide which tier is a brand moment. For a grocer that’s a missing delivery on a holiday weekend. For an apparel retailer it’s a returns dispute. For a BNPL lender it turned out to be somebody who couldn’t make a payment. Those conversations are worth staffing at a cost that looks irrational on a per-contact basis, and the moment you file them under the same cost line as password resets you have already made the Klarna mistake.
Be careful with the number you publish. Klarna’s 700 was accurate, defensible, and became a liability anyway, because the version that spread was not the version the company said. If your board deck has a headline efficiency claim in it, assume it will be quoted back to you in eighteen months by someone with a different agenda.
And watch the shape of the correction rather than the fact of it. Klarna kept the AI, added humans where the AI was weakest, and said so out loud on its own timing. Everything about that is more sophisticated than the cautionary tale it’s been flattened into.
The uncomfortable version
There is a reading of all this that’s less flattering to everyone, including me, and it goes roughly: the reversal got so much coverage because a lot of people wanted it to be true.
The AI-takes-the-jobs story had been running unopposed for two years and it made people anxious. Klarna handed the anxiety a data point, and the data point got used well past what it actually supports. One fintech recalibrated the scope of one deployment. That is not evidence that automation doesn’t work in service, and treating it as evidence is exactly the same error as treating the 700 as proof that it does.
Both times, the number did the thinking for people. That’s the part worth watching for in your own organization.
About The Author:
Maxim Kozlov is a specialist in e-commerce with years of experience building the online systems retailers rely on to sell their products and manage customer relationships. As artificial intelligence reshapes the industry at pace, his current focus is the convergence of AI and commerce technology – and the new patterns of buying and selling starting to emerge from it. He writes about where those two worlds meet, and what it means for the businesses caught in between
