Klarna’s AI U-Turn: 80% of Customer Service Is Automated, So Why Did Humans Come Back?
by 
Klarna’s AI customer-service experiment delivered millions in savings and automated the majority of support chats. Then the Swedish fintech started putting humans back into the customer journey. The real lesson is not that AI failed. It is that Europe’s next customer-service model may depend on knowing exactly where AI should stop.
In February 2024, Klarna looked like one of generative AI’s clearest customer-service success stories. Its OpenAI-powered AI assistant handled 2.3 million conversations in its first month, equivalent to the work of 700 full-time agents. Klarna said customers were resolving issues in less than two minutes instead of 11 minutes, while repeat inquiries fell 25%.
The economics were compelling. Klarna estimated the assistant could improve 2024 profit by around $40 million, while its later reporting put actual 2024 customer-service savings at approximately $39 million. But the story did not end with AI replacing human support. In 2025, Klarna acknowledged that it had pushed too aggressively toward AI-driven cost reduction and began rebuilding parts of its human-service capability.
Here is where the story becomes much more interesting.
Klarna’s AI did not become less important. According to its 2025 annual report, AI handled 80% of customer-service chats, delivered approximately $59 million in 2025 cost savings, and performed work equivalent to more than 850 full-time agents. Klarna also says AI-handled chats ranked on par with human agents for customer satisfaction and showed higher accuracy in issue resolution. At the same time, customers continued to have the option of interacting with human representatives.
That apparent contradiction is the real Klarna case study. AI adoption increased while human access returned.
For European CX leaders, this is far more valuable than another case study claiming that a chatbot reduced costs. Klarna offers a rare view into what happens when an enterprise moves from aggressive automation toward a more deliberate human-in-the-loop AI customer service model.
Klarna’s AI Customer Service Experiment by the Numbers
Metric | Early 2024 | 2025 |
AI customer-service conversations | 2.3M in first month | 31M since launch |
Share of service chats handled by AI | ~66% | 80% |
Equivalent agent workload | 700 FTEs | 850+ FTEs |
AI resolution time | <2 minutes | ~2 minutes |
Human resolution comparison | 11 minutes | 12 minutes |
Repeat inquiries | Down 25% | 25% reduction reported |
Reported savings | ~$40M projected 2024 profit improvement | ~$59M cost savings |
Human support | Available | Explicitly retained |
The progression tells a very different story from the simple headline that “Klarna reversed its AI strategy.”
Klarna did not reverse AI adoption. It expanded it. What changed was the operating model around AI.
Klarna Went From AI Experiment to AI-First Customer Service
Klarna launched its AI assistant in partnership with OpenAI in February 2024. The assistant was designed to handle multilingual customer service as well as refunds, returns and other shopping and payment-related tasks. In its first month, Klarna reported 2.3 million conversations, representing approximately two-thirds of its customer-service chats.
The early performance metrics were unusually strong. Klarna said the assistant was handling work equivalent to 700 full-time agents, customers were resolving their issues in under two minutes, and repeat inquiries declined by 25%. The company also said customer satisfaction was comparable with human agents.
For a CX executive looking at those numbers, the business case was obvious. If AI can handle routine conversations 24/7, respond in seconds, operate across multiple languages and reduce repeat contacts, then customer service can become significantly more scalable without increasing headcount at the same rate.
That is exactly what made Klarna such an influential AI customer-service case study. But there was a problem with interpreting those numbers too broadly.
A customer-service interaction is not the same thing as a customer relationship.
The Problem Was Not That Klarna’s AI Stopped Working
This is the first distinction CX leaders should understand. Klarna’s own 2025 reporting does not suggest that its AI assistant suddenly became ineffective. Quite the opposite. The company reported that AI handled 80% of customer-service chats in 2025 and delivered approximately $59 million in cost savings. It also said AI-handled chats achieved customer-satisfaction results comparable with human agents.
The company’s customer base also continued to expand. Klarna reported approximately 118 million active consumers, 966,000 merchants and $128 billion in GMV for 2025.
So the “AI failed” narrative does not fit the evidence. The more accurate interpretation is that Klarna discovered the limits of using cost reduction as the primary objective for customer-service automation.
In September 2025, CEO Sebastian Siemiatkowski told Reuters that Klarna had overemphasized AI-driven cost savings and was shifting its focus toward productivity, products and customer service. Reuters also reported that Klarna was hiring again after previously reducing its workforce substantially. That is a much more important lesson.
The $59 Million Lesson: Cost Savings Are Not the Same as Customer Value
Klarna’s reported AI economics are impressive. The company reported approximately $59 million in 2025 cost savings from its AI assistant. It estimated that the assistant performed work equivalent to more than 850 full-time agents. Its annual report also says AI-handled chats averaged about two minutes for resolution compared with 12 minutes for human agents in 2024.
But cost savings answer only one question:
How efficiently can the company deliver customer service?
They do not automatically answer:
How valuable was the resulting customer experience?
That distinction matters because customers do not judge a financial-services company by average handling time alone. They judge whether a payment problem was resolved, whether a dispute was handled correctly, whether they could reach someone when necessary and whether the company took responsibility when something went wrong.
This is where Klarna’s decision to retain human access becomes strategically significant. The company is effectively saying that AI can be the default service layer without becoming the only service layer.
Why 80% AI and Human Support Can Exist at the Same Time
At first glance, these two numbers seem contradictory:
80% of chats are handled by AI. Humans brought back into customer service.
They are not contradictory. They describe different parts of the customer-service workload.
Routine interactions can be automated at enormous scale. Questions about payments, refunds, account information, transactions and basic processes can often be structured around known data and predefined workflows.
Complex interactions are different.
A customer may be disputing a transaction, dealing with financial difficulty, challenging a decision, reporting suspected fraud or simply becoming frustrated after several unsuccessful attempts to solve a problem.
The question is therefore not whether AI can technically generate a response. The question is whether AI should be responsible for the outcome. That is the distinction between conversational AI and mature AI customer service.
The Last 20% of Automation May Be the Hardest
Klarna’s experience also exposes a fundamental problem with using automation percentage as the primary CX metric.
Moving from 20% to 50% automation can be relatively straightforward because the easiest and most repetitive interactions are usually automated first. Moving from 50% to 80% can create another major efficiency gain. But pushing from 80% toward 95% is a different problem.
The remaining interactions are likely to contain more exceptions, ambiguity, emotional intensity, financial consequences and policy complexity. The customer-service workload becomes less predictable as automation increases.
That means the economics of automation are not necessarily linear.
The easiest 80% of conversations may not resemble the hardest 20%.
For a financial-services company, that final segment can also carry disproportionate risk. A simple question about a refund is not equivalent to a disputed payment. A transaction-status question is not equivalent to a suspected fraud case.
The correct automation target therefore cannot simply be:
“Automate as much as possible.”
It should be:
“Automate as much as creates value without compromising the outcome.”
Klarna’s Scale Makes the Lesson More Important
Klarna’s customer-service strategy matters because the company operates at enormous scale.
By the end of 2025, Klarna reported approximately 118 million active consumers across 26 countries, approximately 966,000 merchants, and $128 billion in annual GMV. It also reported approximately 46 million active consumers opening the Klarna app each month on average.
At that scale, customer service is not simply a support function. It is part of the product experience.
Klarna is simultaneously operating across payments, commerce, financial services and consumer credit. A customer interaction can therefore involve much more than a basic information request.
The larger the financial consequence of an interaction, the more important it becomes to distinguish between automating the conversation and automating the decision.
The Real Shift: From Employee Automation to Task Automation
This is perhaps the biggest lesson from Klarna’s AI customer-service journey. The first question many companies asked was:
“How many customer-service agents can AI replace?”
A more mature question is:
“Which tasks performed by customer-service agents should AI automate?”
Those are fundamentally different questions.
An AI system might retrieve transaction information, summarize an account history, identify the relevant policy, classify the customer’s problem and prepare a recommendation. A human could then handle the final decision.
This approach allows AI to eliminate repetitive work without requiring the entire interaction to become machine-controlled. The result is a different operating model:
AI performs repetitive work. Humans perform judgment-intensive work.
What a Mature Human-in-the-Loop CX Model Looks Like
A genuine hybrid AI customer-service model is more sophisticated than simply putting a “Talk to a human” button next to a chatbot.
The AI needs to understand when it should continue and when it should escalate.
For example, a low-risk account-information request could remain entirely within the AI workflow. A refund question could be investigated automatically and escalated only when transaction data conflicts with the expected process.
A suspected fraudulent transaction could trigger a completely different workflow, where AI gathers information and authenticates the customer while a human or controlled system handles the sensitive decision.
The most important part is the handoff. If the customer has to repeat everything to the human agent, the company has not created a truly hybrid experience. It has created two disconnected support systems. A better system transfers the conversation history, customer context, actions already taken and unresolved questions to the human specialist.
The Best AI Customer Service May Be Invisible
Customers do not necessarily care whether AI or a human solved their problem. They care whether the problem was solved efficiently, accurately and with an appropriate level of support.
That means the strongest AI customer-service architecture may not be the one with the most visible AI. It may be the one where AI quietly performs the background work while the customer receives a seamless experience.
Imagine a customer reporting a missing refund. Instead of asking the customer to navigate multiple menus, AI could identify the transaction, check the merchant status, verify the refund timeline and determine whether the case falls within normal parameters.
If everything is normal, the AI resolves the interaction. If something is abnormal, the system transfers the complete case to a human specialist. The customer experiences one journey. Behind the scenes, AI and humans perform different parts of the work.
Why Customer Satisfaction Alone Is Not Enough
Klarna says AI-handled chats ranked on par with human agents in customer satisfaction. That is useful evidence, but it should not be interpreted as proof that AI is universally equivalent to human service. The metric is based on Klarna’s own service data and customer surveys.
CX leaders need to go deeper. An overall CSAT score can hide differences between simple and complex interactions. A better measurement system would segment customer satisfaction by:
- interaction complexity;
- financial impact;
- AI-only resolution;
- AI-to-human escalation;
- human-only resolution;
- repeat contact;
- customer effort;
- resolution time.
This can reveal where AI genuinely performs well and where humans create additional value. The question becomes less about whether AI has a higher average score and more about where each model performs best.
The New KPI for AI Customer Service: Cost per Resolved Outcome
One of the most useful lessons from Klarna is that enterprises should rethink the unit economics of customer service. Traditional contact-center metrics often emphasize cost per contact, average handling time and agent productivity. AI makes those metrics even easier to optimize. But the more meaningful question is:
How much does it cost to successfully resolve the customer’s problem?
Consider an AI interaction that costs almost nothing but fails to solve the problem. The customer then contacts the company again. Then again.
Eventually a human agent spends 20 minutes fixing the issue. The company may have achieved a high AI containment rate, but it has not necessarily achieved low cost per resolved outcome.
That is why repeat contact rate is becoming increasingly important in AI customer service. Klarna’s reported 25% reduction in repeat inquiries is therefore arguably more informative than the raw number of conversations handled by AI.
A Better AI Customer-Service Scorecard
European CX leaders should consider moving beyond a single automation KPI.
Metric | What it tells the business |
AI containment | How much demand AI handles |
Resolution rate | Whether the customer actually got an answer or outcome |
Repeat-contact rate | Whether the first interaction really worked |
Escalation rate | How often AI needs human intervention |
Escalation accuracy | Whether difficult cases reach the right team |
Human-transfer friction | Whether customers must repeat information |
CSAT by complexity | Where AI and humans perform differently |
Cost per resolved case | True operational economics |
Customer effort | How difficult the journey was |
Human override rate | How often human judgment changes the AI outcome |
This creates a much more mature definition of AI customer-service performance. A company with 90% automation but poor resolution quality should not automatically be considered more advanced than one with 75% automation and significantly better outcomes. Automation is an input. Customer resolution is the output.
Klarna’s Experience Points to a New AI CX Architecture
The next generation of AI customer service will increasingly move beyond chatbots. The AI agent will not simply answer a question. It will interact with enterprise systems and complete multi-step workflows.
A customer could say:
“My refund has not arrived.”
A basic chatbot might explain the company’s refund policy. An AI agent could instead check the transaction, identify the merchant, verify whether the refund was initiated, compare the transaction against the expected timeline, determine whether an exception exists and either resolve the issue or escalate the complete case.
That is a much bigger opportunity. The value is no longer just conversational. It is workflow automation. And this is where Klarna’s experience becomes relevant to the broader evolution from conversational AI toward agentic customer experience.
Klarna Is Also Showing Why Enterprise AI Needs Better Infrastructure
Klarna’s customer-service assistant did not operate in isolation. The company has extensive transaction, merchant and consumer data infrastructure behind its customer experience. Its 2025 annual report says Klarna processed an average of approximately 3.4 million transactions per day during the year and had more than 6.4 billion transactions conducted through its network since founding.
That scale changes what AI can do. An AI system with access to reliable transaction data can potentially move from answering questions to taking actions. But this also raises the importance of governance, permissions, authentication and auditability.
The future enterprise AI agent will need more than a powerful language model. It will need trusted data, system access, workflow controls, identity management and clear escalation rules. That is why the competitive advantage in AI customer experience is increasingly moving from the model itself to the infrastructure surrounding the model.
Europe’s Human-in-the-Loop Opportunity
Klarna’s experience arrives at a particularly important moment for European enterprises. The European AI environment places substantial emphasis on transparency, accountability and human oversight for applicable AI systems. That does not mean every customer-service chatbot is classified as high risk under European regulation, but it does reinforce a broader direction for enterprise AI governance. For CX leaders, human oversight should not be viewed only as a compliance requirement. It can become part of the customer proposition.
AI gives customers speed. Human experts provide accountability when the situation requires judgment.
That combination could become particularly valuable in financial services, healthcare, insurance, travel and other sectors where customer interactions can carry significant financial or personal consequences. Klarna’s experience provides a real-world example of why that balance matters.
What Klarna’s AI U-Turn Teaches European CX Leaders
The first lesson is to stop treating automation rate as the definition of AI maturity. Klarna can automate 80% of customer-service chats while still maintaining human access because those two strategies address different objectives.
The second lesson is to separate routine interactions from relationship-critical interactions. AI is particularly valuable when the request is structured, repetitive and supported by reliable data. Human involvement becomes more valuable as ambiguity, emotional intensity, financial consequences or exceptions increase.
The third lesson is to build escalation into the AI architecture from the beginning. A human handoff should not be a failure state. It should be a deliberate part of the workflow.
The fourth lesson is to measure the economics of the entire customer journey. If AI reduces the initial interaction but increases repeat contacts, escalations or customer churn, the headline automation figure can conceal the real cost.
The fifth lesson is to treat human agents as an intelligence layer rather than simply a fallback workforce. Human escalations can reveal gaps in AI knowledge, product design, policy clarity and workflow automation.
The Future Is Not 100% AI Customer Service
The most interesting part of Klarna’s story is that the company’s AI adoption continued to expand even after it decided that human customer service still mattered. That tells us something important about the future. The likely destination is not a binary choice between human agents and AI agents. It is a dynamic allocation of work.
AI handles the interactions where automation creates speed and scale. Humans handle the interactions where judgment, empathy, accountability or exception management creates greater value.
As AI agents become more capable, that boundary will continue to move. But it does not necessarily move toward zero humans. It moves toward better allocation of human expertise.
What Klarna Really Proved
Klarna’s original AI launch proved that generative AI could handle customer-service volume at a scale that would have been difficult to achieve with traditional staffing alone. The 2.3 million first-month conversations, 700-agent equivalent workload and reported reduction in resolution time demonstrated the operational potential.
Its 2025 data proved that this was not a short-lived experiment. AI handled 80% of customer-service chats, while Klarna reported approximately $59 million in annual cost savings and more than 850 full-time-agent-equivalent work.
But the company’s subsequent emphasis on human access changed the meaning of the case study. Klarna did not discover that AI was useless. It discovered that customer service is more valuable when automation and human expertise are allocated deliberately.
That is a much more sophisticated lesson than “AI replaced 700 agents” or “Klarna hired humans back.”
The Klarna AI Case Study in One Sentence
AI can automate the majority of customer-service work without eliminating the need for human customer service. That is the paradox CX leaders should pay attention to. The objective is not to maximize the percentage of conversations handled by machines.
The objective is to maximize the percentage of customer problems that are resolved quickly, accurately and appropriately.
Klarna’s numbers show how much AI can contribute to that equation. Its human-service strategy shows why AI should not be the only variable.
Conclusion
Klarna’s AI customer-service experiment began with a spectacular efficiency story. In its first month, the assistant handled 2.3 million conversations and work equivalent to 700 full-time agents. Two years later, the system was handling 80% of customer-service chats and generating approximately $59 million in reported annual savings.
Yet Klarna still chose to preserve human access. That is not a contradiction. It is the more mature version of AI customer service.
The future of CX is not about choosing AI over humans. It is about deciding which interactions deserve automation, which require human judgment, and how intelligently the two can work together.
For European CX leaders, Klarna offers a valuable blueprint: automate routine work aggressively, measure actual customer outcomes, design human escalation into the workflow and use human expertise where it creates disproportionate value.
The companies that master that balance will not necessarily have the highest AI automation rate. They may have something more valuable: the right automation rate for the customer they are serving.
The Next Chapter of AI Customer Experience Is Being Built Now The conversation has moved beyond whether AI will transform customer service. The real question is how enterprises can combine conversational AI, AI agents, automation and human expertise without sacrificing trust or customer experience.
Join the Conversational AI & Customer Experience Summit Europe and discover how leading businesses are building the next generation of AI-powered customer experience.
FAQ’s
No. Klarna continued expanding AI customer service. Its 2025 annual report says AI handled 80% of customer-service chats and generated approximately $59 million in cost savings. The company also continued offering customers access to human representatives.
Klarna’s CEO said in 2025 that the company had placed too much emphasis on AI-driven cost reduction and was shifting toward productivity, products and customer service. Reuters reported that Klarna subsequently resumed hiring while maintaining its long-term commitment to AI.
Klarna reported that its AI assistant handled 80% of customer-service chats in 2025. The company also reported that the assistant had handled 31 million conversations since its launch.
Klarna reported approximately $59 million in 2025 cost savings from its AI assistant. The company had earlier estimated approximately $40 million in 2024 profit improvement and later reported approximately $39 million in 2024 customer-service cost savings.
Klarna says AI-handled chats ranked on par with human agents in customer satisfaction and demonstrated higher accuracy in issue resolution. These are Klarna’s own reported measurements, so they should be understood as company-reported performance rather than independent validation.
Human-in-the-loop AI customer service combines automated AI workflows with human intervention for interactions requiring judgment, exception handling, empathy, accountability or additional oversight. The AI can resolve routine requests while transferring complex cases to specialists with the relevant customer context.
Klarna’s experience suggests a more nuanced direction. AI can automate a substantial percentage of customer-service work, but human representatives can remain important for complex or sensitive interactions. The strategic focus is increasingly shifting from replacing entire roles to automating specific tasks and workflows.
Companies should measure more than automation rate. Resolution rate, repeat contacts, customer effort, escalation quality, satisfaction by interaction complexity and cost per successfully resolved customer outcome provide a more complete picture of AI customer-service performance.
