In February 2024, Klarna told the world it had built an AI assistant that could do the work of 700 full-time customer service agents. It was, at the time, the single most-cited proof point in the entire “AI replaces jobs” argument: a real company, a real number, a real product already live, not a vendor deck projection. Eighteen months later, Klarna is recruiting humans again. What actually happened in between is a more useful case study than either the original hype or the backlash headlines suggest.
The pitch: one chatbot, 700 jobs
Built on OpenAI’s models, Klarna’s assistant launched handling refunds, returns, and payment questions across more than 35 languages. Within its first month it had handled roughly 2.3 million conversations, about 75% of all Klarna customer service chats. Klarna framed the figure as equivalent to the output of 700 full-time agents, and for a while, that was the number every AI vendor pitch deck borrowed.
Behind the launch was a hiring freeze that had already been running since December 2023. Over the following year, Klarna’s headcount fell 22%, to roughly 3,500 employees, mostly through attrition rather than layoffs, but the AI assistant was the story the company told about why it no longer needed to backfill those roles. It was a genuinely disciplined rollout: freeze first, deploy the technology, let the numbers justify the freeze after the fact. Most SMBs do the opposite: adopt a tool first, then wonder six months later why headcount didn’t move.
What the numbers didn’t capture
Handling volume and resolving a customer’s actual problem turned out to be different things. Klarna CEO Sebastian Siemiatkowski later acknowledged publicly that while the AI chatbot was cheaper to run than human staff, it produced consistently “lower quality” outcomes, particularly on anything that didn’t fit the script: disputed charges, edge-case refunds, anything with emotional weight attached.
“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.”
· Sebastian Siemiatkowski, Klarna CEO
That’s the tell. Klarna could see the 75% handled-by-AI number in real time, on a dashboard, the day it happened. The quality problem underneath it took eighteen months to surface publicly, not because it was subtle, but because “chat resolved” and “customer satisfied” were being measured, if at all, as the same metric. A chatbot can close a ticket by giving a technically correct but useless answer. That still counts as resolved. It just doesn’t feel resolved to the person on the other end.
Klarna isn’t alone
The pattern shows up wherever a company drew the automation line by what AI *could* do rather than what it *should* do. Duolingo’s CEO sent an April 2025 memo stating the company would only backfill roles AI genuinely couldn’t replace, and the public reaction was fast and negative enough that the company later had to clarify no full-time layoffs had occurred or were planned. The policy wasn’t necessarily wrong. The framing, internal memo as external signal, was.
Salesforce ran the same substitution in reverse and told the story more carefully. CEO Marc Benioff said the company’s AI agent platform, Agentforce, let it cut its support division from roughly 9,000 employees to about 5,000 since early 2025, but described the remaining headcount as redeployed into sales roles, not simply eliminated. Same underlying number, thousands of support roles absorbed by AI, told as a reallocation story instead of a replacement story. It landed very differently.
The quiet rehire
Klarna is now recruiting customer service staff again, initially through flexible, remote, contract-based roles aimed at students and workers in rural areas. In its own public response to coverage of the shift, the company pushed back on the word “reversal,” pointing out it still works with several thousand outsourced human agents and describing the new hiring as an early-stage pilot, not a wholesale retreat from AI.
That distinction matters more than it sounds. Klarna isn’t saying the chatbot failed. It’s saying the chatbot was never supposed to be the whole answer, and treating it as one is what actually failed.
What Klarna actually got right
It’s worth resisting the easy version of this story: “AI failed, humans win.” The 2.3 million conversations the assistant handled in its first month didn’t evaporate along with the narrative. Refunds, return status, basic payment questions: routine, low-stakes, high-volume work is exactly the category AI still handles well at Klarna today, and probably should. The company’s own pushback on the “reversal” framing is, on this point, fair: it never fully walked away from the assistant, and by its own account still leans on several thousand outsourced human agents alongside it.
The mistake wasn’t building the tool. It was treating “AI can technically do 75% of the volume” as the same thing as “AI should own 75% of the relationship.” Those are different claims, and only one of them held up.
This is an Architect-stage problem more than a technology problem. The Pioneer and Builder stages reward fast, personal judgment calls, you decide in the moment and move on. The Architect stage is where those calls have to become a system: a standing rule for what stays human, written down before the next automation decision shows up, not reinvented case by case. Klarna eventually built that rule. The businesses that get hurt are the ones that never write it down at all.
Where the AI line actually sits for your business
Klarna had the engineering budget of a $10B fintech and it still got the AI/human split wrong on the first attempt, in public, with a customer base large enough that the correction became international news. That’s not a reason for an SMB to avoid AI in customer-facing roles: it’s a reason to draw the line more deliberately than “automate everything we can.” Three questions worth sitting with before you hand any customer-facing process to AI:
Does the customer have an emotional stake in the outcome? Klarna’s AI struggled most exactly where Siemiatkowski said it struggled: disputes, refunds, anything where the customer wants to feel heard, not just resolved. If the interaction carries any weight for the customer, that’s a signal to keep a human in the loop, even if AI handles the first response and the routine 80% of cases underneath it.
Can you measure quality, or only volume? Klarna could see the 75% handled-by-AI number long before it could see the quality problem underneath it. If your only metric is tickets closed or chats resolved, you’re measuring the same blind spot they had. Track resolution satisfaction and repeat-contact rate, how often the same customer comes back on the same issue, not just throughput.
What does it cost you to be wrong quietly? A $10B company could absorb roughly eighteen months of degraded service before the story became public and forced a correction. An SMB usually can’t: word travels faster and forgiveness is thinner in a smaller market, where a bad experience reaches your next ten customers directly instead of getting lost in a review aggregate. Build in a real feedback loop from week one, not a quarterly review.
That’s the actual lesson in Klarna’s reversal: not “AI doesn’t work,” but that the company skipped straight to full automation without deciding, module by module, what specifically should stay human. Business First, AI Second isn’t a slogan against automation. It’s the discipline of making that call on purpose, before your customers make it for you.