Klarna's AI did the work of 700 agents. Two years on, what held up.

In February 2024, Klarna and OpenAI released a number that circled the world: the fintech's new AI assistant was doing the work of 700 full-time agents. It became the reference slide in every "AI is coming for support" deck for the next year. Two years on, it is worth asking the unglamorous question nobody asked at the time — which parts were true, which parts were framing, and what actually happened next.

The original claim

Klarna's press release, published a month after the assistant went live, was specific. The bot had handled 2.3 million conversations across 23 markets and more than 35 languages. Klarna said satisfaction was "on par" with human agents, that repeat contacts had fallen, and that resolution time had collapsed.

The line that actually sold the story, though, was financial.

What held up

The operational core was real, and it has largely stuck. The assistant still runs a large share of Klarna's frontline chat, and the speed claim — under two minutes, always on — is exactly the kind of thing a competent LLM delivers on routine, high-volume questions. By its Q3 2025 earnings, Klarna said the assistant did the work of 853 agents, up from 700. On the narrow question of whether a model can handle a flood of routine buy-now-pay-later queries quickly, the answer held.

Read that stat twice. The equivalent-agent figure went up while customer-service and operations costs also went up. Scale is not the same thing as savings.

What got walked back

The framing is what broke. "700 agents" was widely read as 700 people made obsolete, and that is not quite what happened. Klarna did shrink hard over the same period — but largely through a hiring freeze and attrition, not a support-desk purge. The AI let the company avoid hiring; it did not cleanly delete 700 named jobs.

Then quality caught up. Customers complained the bot returned generic answers and stumbled on anything nuanced; a Forrester analyst called Klarna "almost the poster child for bad AI deployment." In May 2025, CEO Sebastian Siemiatkowski reversed the narrative in public, telling Bloomberg the company had let cost drive the design.

He committed to a model where "there will always be a human if you want," and began recruiting remote agents — an "Uber-type" flexible workforce — to put judgment back in the loop. Notably, Klarna did not switch the AI off. It kept the bot and rebuilt the human layer around it. The reversal was not AI-out; it was an admission that AI-first had gone too far.

What support teams should learn

Four things.

Separate the operational claim from the financial one. "Handles two-thirds of chats, fast" was true. "And we no longer need the people" was the part that needed an asterisk. Apply the same discipline you would to any deflection number or an AI "70% resolved" claim: ask what it is counting before you believe it.

Cost is a corrosive primary metric. Siemiatkowski's own post-mortem — cost became "too predominant" — is the entire lesson. Optimise a support operation for cheapness and quality quietly walks out the back.

Keep the human door open, and visible. Klarna's fix was not a better model; it was giving a frustrated customer a person to reach. An escape hatch to a human is not a failure of automation. It is what makes automation survivable.

Headline case studies are marketing, not benchmarks. The most-cited AI-support statistic of the decade came from a launch post, got repeated for a year, and was then quietly amended by the company that produced it. If you are planning your own rollout, run it like a pilot you are honest about and measure resolution time and quality yourself. Klarna's two years are not a cautionary tale about AI. They are a cautionary tale about believing the first number.