What Happened When Klarna Replaced 700 Customer Service Agents With AI

The starting point

Klarna is one of the world's largest buy-now-pay-later providers, operating across 45 markets with over 150 million active users. Founded in Stockholm in 2005, the company grew rapidly through the 2010s, peaking at a $45.6 billion valuation in 2021 before a brutal down-round to $6.7 billion in 2022 slashed 85% of that value overnight.

By 2023, Klarna needed a new story. The growth-at-all-costs era was over. Investors wanted a path to profitability. An IPO was on the horizon. The company had roughly 5,500 employees and a large customer service operation handling millions of interactions across dozens of markets and languages.

CEO Sebastian Siemiatkowski saw an opportunity. He called Sam Altman at OpenAI and told him, "I want Klarna to be your favourite guinea pig." The plan was straightforward: deploy AI to handle the bulk of customer service queries, reduce headcount through natural attrition, and build the cost efficiency narrative that would underpin the IPO.

What happened next became the most-cited case study in corporate AI. And then it became something else entirely.

What they built

The AI assistant

Klarna launched its OpenAI-powered AI assistant in January 2024 across all 23 markets. The system handled customer queries end-to-end: payment issues, refund requests, account questions, order tracking, and disputes. It operated 24/7 in 35 languages.

The results, announced in February 2024, were striking.

Siemiatkowski did not announce layoffs. Instead, Klarna implemented a hiring freeze and let natural attrition do the work. In a company with 15-20% annual staff turnover, the workforce would shrink without anyone being fired. By the end of 2024, headcount had dropped from roughly 5,500 to around 3,500.

The financial narrative was compelling. Revenue grew 108% between 2022 and 2024 while operating costs stayed flat. Average employee compensation rose 60%, from $126,000 to $203,000. The remaining employees were better paid and, according to the company, more productive.

In November 2024, Klarna filed its draft S-1 with the SEC. The AI efficiency story featured prominently in the IPO prospectus. By September 2025, shares began trading at a $14 billion valuation. The AI narrative had delivered exactly what it was designed to deliver.

What the numbers did not show

The February 2024 announcement was built on one month of data. Some of the metrics were carefully worded. Customer satisfaction was "on par with" human agents, not better. The 700 full-time-equivalent figure measured workload volume, not quality of resolution. And the $40 million saving was a projection, not a result.

As the months passed, the picture shifted. Routine queries were handled well: payment dates, basic refund processing, account information. These are transactional interactions where speed matters more than empathy, and a two-minute resolution is genuinely better than an eleven-minute one.

The problems showed up elsewhere. Complex disputes required judgement the AI could not replicate. Financially distressed customers needed empathy and de-escalation, not faster routing. Multi-step problems that required context across several interactions fell through gaps. Edge cases that did not match standard resolution patterns produced answers that were technically correct and practically useless.

Customer satisfaction scores began to drop. Klarna, which had been quick to publish the positive metrics, stopped publishing satisfaction data. The company that had been the loudest advocate for AI-first customer service went quiet on outcomes.

For a buy-now-pay-later provider preparing for a US IPO, this was not a minor issue. BNPL companies were already under regulatory scrutiny from the CFPB in the US and the FCA in the UK. Customer service quality in financial products is not a brand preference. It is a regulatory exposure.

The reversal

In May 2025, Siemiatkowski told Bloomberg that Klarna would start hiring human customer service agents again.

His explanation was direct: "As cost unfortunately seems to have been a too predominant evaluation factor when organising this, what you end up having is lower quality."

He later went further: "We went too far. We focused too much on efficiency and cost."

The new model had three elements.

Human support repositioned as premium. In a world where every company uses AI for basic support, Klarna concluded that access to a real human becomes the scarce, valuable resource. The framing shifted from "humans are expensive" to "humans are the product."

An Uber-style flexible workforce. New customer service agents would choose their own hours, work from anywhere in Sweden, and be paid 400 SEK per hour (roughly GBP 30), well above typical contact centre rates. The flexibility was designed to attract a different calibre of candidate.

AI handles volume, humans handle value. The AI assistant continued processing routine transactions. Human agents took on everything requiring empathy, discretion, or escalation. The split was no longer about cost per interaction. It was about what each interaction is worth to the customer relationship.

Siemiatkowski framed the commitment publicly: "From a brand perspective, a company perspective, I think it's so critical that you are clear to your customer that there will always be a human if you want."

The results

A note on the headcount narrative: the overall workforce reduction was real and significant, driven by attrition and hiring freeze across the entire business, not only customer service. Siemiatkowski has been consistent that no one was fired. The 700-agent-equivalent figure refers specifically to the customer service workload absorbed by AI, not to 700 people who lost their jobs. The distinction matters, though the practical effect on employment was the same.

What makes this case interesting

The most aggressive AI replacement became the most visible correction. Klarna was the company every CEO cited when arguing that AI would replace customer service teams. That same company then became the proof that replacing humans without reinvesting in human capability degrades the thing you are trying to improve. The story has more credibility because of how loudly the initial claims were made.

Cost optimisation and quality are not the same objective. Klarna optimised for cost per interaction and got exactly what they optimised for: cheaper interactions. The problem is that cheaper interactions are not better interactions. When the metric is cost, the AI wins. When the metric is customer lifetime value, trust, or regulatory compliance, the answer is more complicated. The company learned this in public.

The reversal created a better model than the original. The hybrid model Klarna landed on, where AI handles volume and humans handle value, is more sophisticated than either the pre-AI model (humans handle everything) or the AI-first model (AI handles everything). The correction was not a retreat to the old way of working. It was a step forward to something that neither the old model nor the AI-only model could achieve. Humans are no longer doing the work the AI is good at. They are doing the work that builds the relationship.

Financial narratives drive premature deployment. The timing of Klarna's AI push was not a coincidence. The company needed an efficiency story for its IPO. The AI deployment was as much a capital markets narrative as it was an operational improvement. When the pressure to show cost savings outweighs the discipline to measure properly, deployment decisions get made for the wrong reasons.

"On par with" is not "better than." The original announcement carefully described AI satisfaction as "on par with" human agents. That phrase did a lot of work in the press coverage, where it was frequently upgraded to "as good as" or even "better than." In practice, "on par" on aggregate hides significant variance: the AI was better on routine queries and worse on complex ones. The aggregate masked the distribution.

The challenges

No published quality decline data. Klarna released detailed positive metrics in February 2024 and then stopped publishing satisfaction data. The reversal was announced without specific numbers showing how far quality had fallen. The absence of data is itself informative, but the case would be stronger with published before-and-after satisfaction scores.

The headcount story is more complex than either version suggests. Klarna's overall workforce reduction was driven by multiple factors: a company-wide hiring freeze, natural attrition, broader restructuring, and AI automation. Attributing the full reduction to AI overstates the technology's role. Attributing the reversal purely to AI quality issues understates the regulatory and reputational pressures on a company approaching an IPO in financial services.

The IPO complicates the narrative. Every public statement Siemiatkowski made during this period was also a statement to future investors. The initial AI enthusiasm served the efficiency narrative. The correction served the "mature, customer-focused leadership" narrative. Both are genuine, but neither is free from capital markets positioning.

The Uber-style model raises its own questions. Flexible hours and premium pay are attractive, but the gig-economy framing (choose your hours, no guaranteed volume) introduces labour precarity. Swedish trade unions have noted this. The model shifts risk from employer to worker in ways that deserve scrutiny, even if the hourly rate is generous.

Lessons for your programme

Measure what you are optimising for, and check it is the right thing. Klarna optimised for cost per interaction and resolution time. These are legitimate metrics, but they are not the only ones that matter. Before you deploy AI in customer-facing operations, decide whether you are optimising for cost, satisfaction, retention, lifetime value, or compliance. They pull in different directions, and what you measure is what you get. The AI Opportunity Assessment Scorecard (04c) in Section 4: Opportunity Identification forces you to score opportunities against business value, not only cost reduction.

Separate the financial narrative from the operational decision. Klarna's AI deployment served two masters: genuine operational improvement and an IPO efficiency story. When your AI business case is also a story you need to tell investors, a board, or a PE sponsor, the pressure to show results fast overrides the discipline to measure properly. The AI Business Case Template (13b) in Section 13: Business Case Production includes quality and risk metrics alongside financial projections, specifically to prevent the cost line from drowning everything else out.

Plan the human role before you deploy the technology, not after it fails. Klarna deployed the AI, let attrition shrink the team, and then discovered that humans were needed for a different kind of work. The reskilling and repositioning happened reactively. The Workforce Impact Model (15c) in Section 15: Designing for Transformation is designed to model the human role change before deployment, so the transition is planned rather than forced by service failure.

A correction is not a failure. The most useful part of Klarna's story is the correction. They measured, they learned, they changed course. The hybrid model they landed on is stronger than either the original or the AI-only version. If your AI programme needs to adjust, that is evidence of learning, not failure. The Experiment Outcome Canvas (07g) in Section 7: The Experimentation Framework captures both what worked and what did not, so that corrections compound into better decisions.

Sources

• Klarna AI Assistant Handles Two-Thirds of Customer Service Chats in Its First Month (Klarna International Press Release, 2024)
• Klarna's AI Assistant Does the Work of 700 Full-Time Agents (OpenAI, 2024)
• Klarna Plans to Hire Humans Again (Fortune, 2025)
• Klarna Changes Its AI Tune and Again Recruits Humans for Customer Service (CX Dive, 2025)
• Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls (Entrepreneur, 2025)
• Klarna CEO Says AI Helped Company Shrink Workforce by 40% (CNBC, 2025)
• How Klarna Has Cut Staff in Half While Raising Pay By 60% (Entrepreneur, 2025)
• Klarna Turns From AI to Real Person Customer Service (Bloomberg, 2025)
• Klarna CEO Admits AI Job Cuts Went Too Far (MLQ AI, 2025)
• Klarna Tried to Replace Its Workforce With AI (Fast Company, 2025)
• What's Behind Klarna's $14 Billion IPO Valuation? (Morningstar, 2025)