Customer Experience · November 20, 2026
Klarna Customer Service: What It Gets Right and Wrong
Klarna's AI assistant cut resolution times and replaced hundreds of agents, then its CEO admitted the company had gone too far. Here's what that reversal teaches CX leaders.
Klarna spent 2024 telling the world its AI assistant had replaced 700 customer service agents. By 2025, its own chief executive was publicly admitting the company had gone too far — and quietly rebuilding the human team it had just dismantled. That reversal is the real story, and it says more about the limits of automation in customer experience than the launch ever did.
The short answer: Klarna's AI assistant is genuinely good at speed and volume, and genuinely bad at the moments that determine whether a customer trusts the brand again. It compresses resolution time and handles routine queries at a scale no human team could match. But Klarna's own leadership has since conceded that stripping out human agents degraded the quality of service on exactly the interactions — disputes, refunds, anything emotionally loaded — where quality matters most. The lesson isn't "AI doesn't work." It's that Klarna measured the wrong thing for too long.
What did Klarna's AI customer service actually do?
In February 2024, Klarna launched an OpenAI-powered AI assistant across 23 markets, rolling it out globally rather than piloting it in a single region. Within a month, according to Klarna's own announcement, the assistant was handling two-thirds of all customer service chats — doing the equivalent work of roughly 700 full-time agents, cutting average resolution time from around 11 minutes to under two, and reportedly reducing repeat contacts. Those are Klarna's own figures, published in its February 2024 press release, not an independent audit — a distinction worth keeping in mind, because every self-reported metric in that release measures throughput, not satisfaction.
That distinction matters more than it sounds. A support operation can be fast, cheap, and widely used while still eroding the thing customer service exists to protect: the customer's belief that the company will make things right when something goes wrong. Klarna's assistant was built for the former. It took the company roughly a year to discover, publicly, that the latter had quietly slipped.
Why did Klarna's AI assistant look like a success at first?
Because it was optimised against metrics that make automation look brilliant on a slide. Resolution time, cost per contact, and deflection rate are all what a behavioural economist would call System 1 metrics — fast, intuitive, immediately legible numbers that a board can nod at without interrogating. Trust, by contrast, is a System 2 verdict: it forms slowly, gets tested under stress, and only shows up months later in churn, complaint escalation, or public sentiment. Daniel Kahneman's dual-process framework, laid out in his 2011 book Thinking, Fast and Slow, is exactly why a company can genuinely improve its dashboard while quietly damaging the relationship the dashboard is supposed to protect.
Klarna's rollout had three things working in its favour during that first wave of coverage:
- Scale from day one. Launching across 23 markets simultaneously generated a volume of data — and a volume of press — that a phased pilot never would have. The story became "AI replaces 700 agents" before anyone had time to ask what those 700 agents used to catch that the model doesn't.
- A category where speed is genuinely valued. Buy-now-pay-later customers mostly contact support for simple, low-stakes queries — payment dates, order status, account details. An AI assistant answering those instantly is a real improvement, not a gimmick.
- A novelty effect in the coverage itself. Financial and tech media ran the efficiency numbers largely unchallenged, because "AI replaces hundreds of jobs" is a more compelling headline than "AI handles simple queries well, humans still needed for hard ones."
None of that is a criticism of the technology. It's a description of how easy it is to mistake a productivity win for a trust win, especially when the metric you're measuring was chosen because it's the one that improves fastest.
Where did the AI-first approach break down?
Klarna's own leadership answered this question before any outside critic needed to. In public comments reported by Bloomberg in May 2025, co-founder and chief executive Sebastian Siemiatkowski acknowledged that the company's aggressive push to automate customer service — partly driven by cost-cutting and headcount reduction — had produced a decline in service quality, and that Klarna was moving to rehire human agents and build a hybrid model offering customers the option of speaking to a person. That is a rare thing in corporate communications: a CEO publicly conceding that an AI-first bet on customer service didn't hold up.
The behavioural mechanism behind the failure is straightforward, and it's one every CX leader deploying AI support should sit with. A chatbot resolving a billing query is a low-stakes, low-emotion interaction — ideal territory for automation. A customer disputing a charge, contesting a late fee, or trying to resolve a payment plan gone wrong is a high-stakes, high-emotion interaction, and the affect heuristic — the tendency, described by Paul Slovic's research on risk perception, for people to judge a situation based on how it makes them feel rather than on the facts of the case — means the customer's verdict on the whole brand gets set in that one moment. An instant, templated AI response to an emotionally loaded complaint doesn't read as "efficient." It reads as "nobody is listening." Speed, in that moment, is not a virtue; it's evidence the company didn't take the problem seriously enough to involve a person.
This is also a peak-end rule problem. Daniel Kahneman's research on retrospective evaluation, most accessibly summarised in Harvard Business Review's coverage of behavioural economics in service design, established that people judge an entire experience by its most intense moment and its final moment — not by the average of every interaction along the way. A customer who had five fast, flawless AI chats and then one cold, unresolved dispute will remember the dispute. Klarna's throughput numbers counted all six interactions equally. The customer's memory didn't.
Klarna didn't prove that AI can't do customer service. It proved what happens when a productivity metric is allowed to grade an experience question.
What is Klarna doing differently now?
The correction, as described by Siemiatkowski in the reporting around Klarna's 2025 shift, is a hybrid model: AI continues to handle the high-volume, low-stakes contacts it's genuinely suited to, while human agents are reintroduced for the interactions that need judgement, empathy, or discretion — disputes, complex refunds, anything where the customer is upset. That's not a retreat from AI. It's a correction of scope.
The principle underneath it is one Renascence applies in every customer experience engagement that involves automation: AI should absorb friction, not absorb accountability. A chatbot that confirms a delivery date is removing friction. A chatbot that delivers a final word on a disputed charge is removing accountability — and customers can tell the difference instantly, even if the company's dashboard can't.
Three structural choices separate the two:
- Stakes, not volume, should set the automation boundary. The right question isn't "how many of these contacts can AI handle?" It's "how much does the customer stand to lose, financially or emotionally, if this interaction goes wrong?" High-stakes contacts need a visible, fast route to a human — not as a last resort after three failed bot attempts, but as a standing option.
- Deflection rate is a vanity metric unless it's paired with a trust metric. A support operation chasing lower contact time without tracking complaint escalation, repeat contact for the same issue, or churn among customers who used the bot is optimising blind.
- Escalation needs to be a design decision, not a failure state. If handing a conversation to a human only happens when the bot gives up, customers experience it as being passed along — the opposite of care. A well-designed escalation strategy treats the handoff itself as a moment of truth, not an admission of defeat.
What should other CX leaders learn from Klarna's reversal?
The practical lesson isn't "don't automate." It's "automate the transaction, not the relationship." Here is a sequence worth following before expanding any AI assistant into territory Klarna's did:
- Map the journey before you map the bot. Identify which touchpoints are transactional (balance checks, order status, address changes) and which are emotional (disputes, cancellations, complaints, anything involving money the customer believes is wrong). Automate the first category aggressively. Treat the second category as human-first by default.
- Set a trust metric alongside every efficiency metric. For every resolution-time or deflection-rate target, pair it with a measure of whether the issue actually stayed resolved — repeat contact rate within 30 days is a reasonable proxy, and it's a number Klarna's own 2024 figures implied but didn't foreground.
- Build the human escalation path before you need it, not after complaints spike. Siemiatkowski's 2025 comments describe Klarna rehiring agents under pressure. It's considerably cheaper, reputationally and financially, to keep a human layer running in parallel from day one than to rebuild it once trust has already slipped.
- Audit the emotionally loaded contacts separately. Pull a sample of disputes, refund requests, and complaints handled by AI and read them as a customer would, not as a transcript. If the tone reads as efficient-but-cold, the model is winning on time-to-resolution and losing on the only metric that compounds: whether the customer stays.
- Tell customers which they're getting, and let them choose. Klarna's hybrid correction works partly because it removes ambiguity. A customer who knows they can reach a person for a hard problem trusts the bot more for easy ones — the choice architecture itself becomes reassuring, not a hurdle to route around.
That last point sits on a well-established behavioural principle: removing a choice doesn't reduce anxiety, it increases it. Richard Thaler and Cass Sunstein's work on choice architecture, laid out in their 2008 book Nudge, shows that people tolerate automated systems far better when an opt-out to a human exists and is genuinely easy to find — even if most people never use it. The safety net changes how the automated path itself feels.
Where does the line between efficient and hollow actually sit?
Klarna's experience draws that line more clearly than most case studies in the sector, precisely because the company ran the experiment publicly, at scale, and then corrected it publicly too. Efficiency is a System 1 metric: it is fast, legible, and flatters a quarterly report. Trust is a System 2 verdict: slow to build, expensive to rebuild, and invisible on a dashboard until it's already gone. An AI assistant can absolutely improve both — but only if the organisation is deliberate about which interactions it's allowed to own outright, and which ones still need a person who can be held accountable for the outcome.
The companies that get this right don't treat automation and empathy as a trade-off to be balanced. They treat them as two different tools for two different jobs, and they're explicit, internally and with customers, about where one ends and the other begins. That clarity — not the model, not the vendor, not the headline efficiency number — is what separates a support operation customers trust from one they merely tolerate until something goes wrong.
The real test was never speed
Klarna's AI assistant will keep answering routine questions faster than any human team could. That was never in doubt, and it was never really the point. The point is what happens in the minority of contacts that are emotionally loaded, financially consequential, or simply strange enough that no script covers them — because that minority is where customers decide, permanently, what kind of company they're dealing with. Klarna built the fast lane first and discovered, the hard way, that the slow lane is the one that keeps the business.
Organisations rolling out AI support now have the benefit Klarna didn't: a public, well-documented case study of exactly where the line sits. The companies that study it closely will design the human layer in from the start, rather than rebuilding it under pressure once the trust is already gone.
Renascence works with organisations designing exactly this kind of hybrid model — where automation and human judgement are deliberately mapped against the stakes of each touchpoint rather than bolted together after the fact. Our customer experience and behavioural economics teams help CX leaders decide, journey by journey, where AI earns its place and where it doesn't. For a structured view of how your own support model holds up against this kind of stress test, our CX Maturity Assessment is a reasonable place to start.
For related reading on how other well-known brands handle the same tension between automation and human judgement, see our analysis of Canva's customer service model, our breakdown of Wise's customer journey, and our look at the research linking customer experience to profitability.
Further reading
FAQ
Questions we get on this topic
Related reading
Writing on how human behavior shapes the experiences brands deliver — at the intersection of behavioral economics and customer experience.
Stay ahead of CX
Get the Journal in your inbox.
Insights, frameworks and event round-ups from the Renascence team. No spam, ever.






