Case Study: Klarna's AI Customer Service Assistant
In early 2024, Klarna announced that its AI assistant was handling two-thirds of its customer service chats, doing work it compared to about 700 full-time agents. A year later, its chief executive was talking about making sure customers could always reach a human. The story is one of the clearest public lessons yet on what AI customer service can and cannot do.
Key takeaways
- Klarna reported that its OpenAI-powered assistant handled about two-thirds of customer service chats in its first month, according to the company's February 2024 announcement.
- The company said resolution times fell sharply and satisfaction was similar to human agents, though those figures were self-reported.
- In 2025, Klarna's CEO said in interviews that the company was again emphasizing human support, citing concerns about quality when cost was the main focus.
- For small businesses, the lesson is to automate routine questions while keeping an easy, well-staffed path to a person.
Klarna is a Swedish financial technology company best known for "buy now, pay later" payments, which let shoppers split purchases into installments. With tens of millions of users across many countries, it handles an enormous volume of customer questions about payments, refunds, orders and accounts. That made it an obvious candidate for AI customer service, and an unusually visible test case for everyone else.
This case study uses only widely reported facts and attributes each number to its source. Klarna's figures were published by the company itself and have not been independently audited, so read them as company claims. The goal is not to judge Klarna, but to draw practical lessons for businesses of any size considering AI in customer service.
Background: why Klarna went big on AI
Klarna was an early and enthusiastic adopter of generative AI. Its leadership spoke publicly about using AI across the business, from marketing to internal tools, and about slowing hiring as AI took on more work. Customer service was a natural focus. Many questions are repetitive (Where is my refund? When is my next payment due? How do I change my plan?), and answering them quickly at any hour in many languages is expensive with human teams alone.
Klarna worked with OpenAI to build an assistant inside its app and website. Rather than a simple scripted chatbot, it was designed to understand natural questions, look up relevant account information and take some actions, such as handling refunds and returns, while handing harder cases to human agents.
What they did: the February 2024 launch
In February 2024, Klarna published results from the assistant's first month of global operation. According to the company's announcement:
- The assistant had handled about 2.3 million conversations, roughly two-thirds of Klarna's customer service chats.
- Klarna said this was doing the equivalent work of about 700 full-time agents.
- The company said customer satisfaction scores were on par with human agents.
- It reported that average resolution time fell from about 11 minutes to under 2 minutes.
- It said repeat inquiries dropped by about a quarter, suggesting more issues were resolved the first time.
- The assistant was available in many markets and communicated in more than 35 languages, according to Klarna.
- Klarna estimated the assistant would drive a profit improvement of about $40 million in 2024.
The announcement attracted worldwide attention. Many commentators treated it as proof that AI could replace large parts of customer service, while others questioned how satisfaction was measured and whether a drop in repeat contacts might partly reflect customers giving up. Some coverage also reported a Klarna spokesperson explaining that much of its customer service was provided through large outsourcing partners, and outlets noted that the company denied any link between the 700 figure and its earlier 2022 layoffs. In other words, the number described an equivalent workload, not a count of people Klarna let go.
Common mistake
Reading a headline like "AI does the work of 700 agents" and concluding that your business can replace its support team. Klarna had a large engineering team, deep integration with its own systems, huge volumes of simple questions and human agents behind the AI. Most small businesses have none of those conditions.
What happened next: the 2025 shift back toward humans
In 2025, the tone changed. In a May 2025 interview reported by Bloomberg and widely covered elsewhere, Klarna chief executive Sebastian Siemiatkowski said the company wanted to make sure customers would always be able to speak to a human if they wanted to. According to those reports, he acknowledged that focusing heavily on cost had led to lower quality in some cases, and Klarna began recruiting human customer service staff again, reportedly using a flexible, remote working model.
It is important not to overstate this. Klarna did not abandon AI. Reports at the time indicated the company continued to use AI extensively, including in customer service, while adjusting the balance so that human help remained clearly available. The shift is better described as a correction than a reversal: from "AI first, humans as backup" toward "AI for routine work, humans readily available for everything else."
Pro tip
Measure customer outcomes, not just cost and speed. Track whether the issue was actually resolved, whether customers came back with the same problem, how many asked for a human and what they said in follow-up surveys. Faster answers only count as better service if they solve the problem.
Results: what the Klarna story shows
Taken together, the publicly reported facts suggest several things:
- AI can handle very high volumes of routine questions. Even allowing for the figures being company-reported, the scale was large.
- Speed and cost gains can be real. Klarna reported much shorter resolution times and expected significant savings.
- Quality and trust are harder to measure. Satisfaction scores alone may not capture frustration from customers with complex or emotional issues.
- Public perception matters. Bold claims about replacing people attracted scrutiny, and later messaging emphasized human availability.
| Question type | Good fit for AI? | Why |
|---|---|---|
| Order status, payment dates, refund timing | Usually yes | Factual, repetitive, answerable from account data |
| Policy questions (returns, shipping) | Usually yes | Answers come from written policies |
| Billing disputes and fraud concerns | Handoff needed | Money, stress and judgment involved |
| Complaints and emotional situations | Human preferred | Empathy and discretion matter most |
| Hardship or vulnerability | Human required | Needs care and, in finance, may be regulated |
Lessons for a small business
1. Start with the routine questions
Look at your last few hundred support messages and group them. In many businesses, a handful of questions make up a large share of volume. Those are where AI drafts, self-service answers or a chatbot help most.
2. Make the human path easy
Customers get frustrated when they feel trapped in a loop with a bot. Offer a clear "talk to a person" option, explain how long it will take and make sure someone actually responds. Our separate article on AI chatbots covers handoff design in detail.
3. Measure quality, not just cost
Track resolution, repeat contacts, escalations and customer comments, not only cost per contact. Review a sample of AI conversations every week.
4. Keep humans for judgment and empathy
Complaints, disputes, refunds outside policy and anything involving distress deserve a person. These moments shape how customers feel about you more than routine answers do.
5. Be careful with public claims
If you announce AI customer service, focus on what customers gain, such as faster answers at any hour, rather than on staff replaced. Claims about AI performance should also be accurate and supportable under consumer protection rules.
Hypothetical example
Suppose an online pet supply store gets about 600 support emails a month. Grouping them shows that delivery status, return instructions and subscription changes make up most of the volume. The owner sets up AI-drafted replies for those three topics, reviewed by a staff member before sending, and adds a self-service order tracking page. Complaints and anything about sick pets go straight to a person. Response times fall, the team spends more time on the difficult cases and customers can always reply "I want to talk to someone" and get a human.
What worked, as reported
- Very high volume handled by AI
- Much faster resolution times, per Klarna
- Support in many languages at all hours
- Significant expected cost savings
What needed correcting
- Concerns about quality when cost led decisions
- Customers wanting reliable access to people
- Scrutiny of bold replacement claims
- Self-reported metrics that were hard to verify
Your AI support action plan
- Group your last 200 to 500 support messages by topic
- Pick the three most common routine topics for AI help
- Write clear, current policy documents the AI can rely on
- Add a visible, fast route to a human on every channel
- Track resolution, repeat contacts and satisfaction weekly
- Review a sample of AI conversations every week
- Check your tool's data and privacy settings before launch
Frequently asked questions
Did Klarna replace 700 employees with AI?
Klarna said its assistant was doing work equivalent to about 700 full-time agents. Press coverage at the time reported that much of Klarna's customer service was provided through outsourcing partners, and the figure was presented as an equivalent workload rather than a count of people laid off.
Did Klarna stop using AI for customer service?
No. Based on 2025 reporting, Klarna continued to use AI while putting more emphasis on making human support available. Check recent news for the company's current approach, as it may have changed again.
Can a small business get similar results?
On a smaller scale, yes for routine questions. Many helpdesk and e-commerce platforms now offer AI features that draft replies or answer common questions. Results depend on clear policies, good data and a reliable human backup.
How do I know if my AI support is hurting customer experience?
Watch for rising repeat contacts, more requests for a human, lower satisfaction scores, negative reviews mentioning the bot and refunds or cancellations after AI conversations. Read real conversations regularly rather than relying on summary metrics alone.
Should I tell customers they are talking to AI?
Yes. Being upfront builds trust, and in some places it is becoming a legal requirement, for example under the EU AI Act's transparency rules for chatbots. A short line such as "You are chatting with our AI assistant; type 'agent' to reach a person" sets clear expectations from the first message.
Related reading
The bottom line
Klarna's experience shows both the power and the limits of AI customer service. AI can take on a large share of routine questions quickly and cheaply, according to the company's own reports. But customers still want to reach people when it matters, and a strategy led mainly by cost can weaken quality. For a small business, the winning approach is a blend: AI for the repetitive work, people for judgment and care, and honest measurement of what customers actually experience.
