Klarna gave support to an AI, then went back for the humans

An Authority Report is where we read what people with real operating responsibility said in public, in their own words, and work out what it means for a company that isn’t them.

This one is about the most-cited AI customer service story in Europe, and the part of it that gets left out. Klarna automated support at a scale nobody else had tried. Then its CEO went on the record to say it had gone too far. Both halves are true, and the second half is the one worth reading closely, because it names a design mistake rather than a technology failure.

February 2024: the number everyone remembers

Two-thirds of all chats in the first month, described as the equivalent work of 700 full-time agents.

Klarna, February 2024. Headcount went from roughly 5,500 to around 3,000 over two years, with AI credited for a large share.

The company put a figure on the saving and the industry repeated it for a year.

Nothing in the reporting since suggests the volume claim was wrong. The assistant did handle the chats. That is what makes the next part interesting: this is not a story about an automation that failed to work.

May 2025: “we went too far”

Speaking to Bloomberg, Sebastian Siemiatkowski said the company had cut human support too aggressively and was hiring again for premium and complex cases. His summary was three words long:

We went too far.

He was more specific about the mechanism:

We focused too much on cost. The result was lower quality.

And on what he wanted customers to be able to count on:

From a brand perspective, a company perspective, I just think it’s so critical that you are clear to your customer that there will always be a human if you want.

Read that last line as a spec rather than a sentiment. It is not “we need more agents.” It is: at every point in the flow, the customer must be able to see the way out, and the way out must be real.

June 2025: what they actually rebuilt

A month later Siemiatkowski described the model they landed on, and it is not a retreat to 2023. Human service became the premium tier, not the default one:

We think offering human customer service is always going to be a VIP thing.

And on why he rejects the framing that the company reversed itself:

So we think that two things can be done at the same time. We can use AI to automatically take away boring jobs, things that are manual work, but we are also going to promise our customers to have a human connection.

The staffing behind it is an on-demand pool, remote agents on flexible hours, many of them students and parents, each working with AI assistance in the conversation rather than against a script. The AI did not move out of support. It moved out of the escalation path.

The read: they removed the exit, not the agents

The common telling of this story is “AI wasn’t good enough, humans came back.” That reading is comfortable and it is wrong. Klarna’s assistant was good enough for two-thirds of contacts, and still is.

What broke was the last third. When a system automates the routine cases and leaves no credible route out of the remaining ones, the customers who reach that route are the angriest ones you have: the disputed charge, the frozen account, the thing the FAQ has no entry for. They arrive at the point where the system is weakest, in the state where they can least tolerate it. The average handling time looks excellent. The relationship does not survive.

Siemiatkowski’s own diagnosis says the same thing in financial language: the optimisation target was cost, and quality was what paid for it. That is a design decision, made once, and it does not announce itself in the metrics for months.

What this means at 40 people instead of 5,000

Most companies reading this are not automating 700 agents. They have four people in support, or nine, and they are wondering whether an agent can take the repetitive half. It can. Klarna’s numbers on the routine tier are not the cautionary part of this story.

Three things carry over regardless of size.

Design the escalation before the automation. The question is not “what can the agent answer” but “what happens to the ones it cannot, and how fast does a person see them.” If the answer involves a queue nobody owns, the agent is not ready to be in front of customers. In the email triage agent we published the build for, the escalation path exists before the classifier does, and it is loud on purpose.

Never hide the exit. A customer who cannot find a human assumes there isn’t one. That assumption is expensive and it outlasts the conversation. Klarna is now selling the human as a feature; the smaller version of that move is simply not making people fight for it.

Watch quality, not deflection rate. Deflection is the metric that made Klarna’s 2024 announcement so quotable, and it is the metric that hid the problem. If the only number on the dashboard is how many contacts never reached a person, the dashboard will look better every month the experience gets worse. Pair it with something the customer feels: resolution on first contact, reopen rate, the share of escalations that arrive already angry.

The company that gets this right is not the one that automates least. It is the one that decides up front which conversations a person owns, and then builds the machine around that line instead of discovering it eighteen months later.

Frequently asked questions