← The Signal

Klarna Replaced 700 People With AI. Then It Hired Them Back.

The most famous "AI replaced humans" story quietly became the opposite. The lesson isn't that AI failed — it's where it failed, and why that line matters before you cut.

A couple of years ago, Klarna handed us the cleanest "AI replaced humans" headline anyone could ask for. The Swedish fintech said its OpenAI-powered assistant was doing the work of 700 customer-service agents — handling two-thirds of all chats, in dozens of languages, resolving issues in under two minutes instead of eleven. Exhibit A that the future had arrived. Then, quietly, Klarna started hiring people back. So let's talk about what actually happened, because it's more useful than either headline.

First, the part the AI skeptics get wrong: the technology worked. Klarna's numbers were real. It drove cost-per-conversation down by roughly 40%, banked tens of millions in savings, and handled volume no human team could match. On tier-one, high-volume, repetitive requests, the machine won cleanly. Anyone telling you Klarna "proved AI doesn't work" is selling the opposite hype.

Now the part the maximalists skip. As the company leaned all the way in, quality on the hard cases — the emotional ones, the complex disputes, the situations that need judgment — degraded. Customers noticed. The CEO eventually admitted they'd leaned too far toward efficiency and cost, and that the result wasn't sustainable. Klarna began rebuilding the human capacity it had cut, moving to a hybrid model: AI for the routine volume, people for the escalations.

Here's the expensive part nobody models in the original business case. When you remove 700 people, you don't just remove a cost line — you remove accumulated judgment. The pattern recognition for the weird edge case. The instinct for when a frustrated customer is about to walk. That knowledge wasn't written down; it lived in the people. Cutting them deleted it, and rehiring doesn't restore it instantly — new agents have to rebuild it from scratch. Regular readers will recognize the theme: you can't offload your learning. Klarna offloaded the task and lost the learning that came with it.

So, the consultative version.

For a business: replacement and augmentation look identical in a spreadsheet and couldn't be more different in practice. Before you cut, draw the line honestly: which work is genuinely routine and high-volume (great for AI), and which quietly runs on human judgment and relationships (dangerous to automate away)? Get that line wrong and you pay twice — once to remove the capacity, again to rebuild it. The cost of full replacement always includes the cost of unwinding it if you're wrong.

For an individual worried about exactly this: notice which side of the line your work sits on. If most of your day is high-volume and routine, that's the part to hand to AI — and to deliberately move yourself up the chain, toward the judgment and the hard cases the machine fumbles. The Klarna agents who'll matter most in the rebuilt team aren't the ones who answered the most tickets. They're the ones who can handle what the AI can't.

The takeaway is one question, the one Klarna learned the hard way: what in our work looks routine but quietly runs on human judgment — and are we about to automate that away?

AI didn't fail at Klarna. A replacement mindset did. Augment the routine, protect the judgment, and you get the savings without the expensive walk-back.

That's the signal for this Monday. See you next week. — Adam