What Klarna Learned About AI Customer Service — And Why They Changed Course

In February 2024, Klarna announced something remarkable: their AI assistant had handled 2.3 million customer service conversations in its first month alone. That's the equivalent of 700 full-time customer service agents.

The headlines were glowing. The stock market took notice. And every business owner started wondering: could AI handle my customer service too?

But here's what most articles don't tell you: Klarna later changed course. They moved away from pure AI automation toward a hybrid model that combines AI efficiency with human judgment. And that pivot contains the real lesson about AI in customer service.

The Problem That Won't Go Away

If you run a small or mid-sized business, you know the customer service math doesn't work.

Customers expect fast responses. Research shows that 60% of consumers define "immediate" as 10 minutes or less when they need support. But staffing for that level of responsiveness means paying people to wait during slow periods — or leaving customers hanging during busy ones.

The traditional solutions all have trade-offs:

• Hire more staff — expensive, and you're still understaffed during peaks
• Outsource to a call center — cheaper per hour, but quality varies wildly
• Use basic chatbots — frustrates customers with scripted dead-ends
• Just accept slower response times — and watch customers leave for competitors who respond faster

This is where AI enters the picture. Not as a magic solution, but as a tool that changes the math.

The Klarna Story: Hype, Results, and Course Correction

Let's look at what actually happened with Klarna, because the full story is more instructive than the headlines.

The Initial Results

Klarna's AI assistant, built on OpenAI's technology, launched in January 2024. The first-month numbers were genuinely impressive:

• 2.3 million conversations handled
• Two-thirds of all customer service chats managed without human involvement
• Average resolution time dropped from 11 minutes to 2 minutes
• Repeat inquiries fell by 25%
• Projected $40 million in annual savings

These weren't inflated marketing numbers. Klarna is a publicly-scrutinized fintech company handling billions in transactions. The AI was genuinely processing millions of conversations about payments, refunds, and account issues.

What They Discovered

Here's where the story gets more honest — and more useful.

As Klarna scaled their AI, they found limits. Complex disputes needed human judgment. Frustrated customers sometimes needed empathy that AI couldn't provide authentically. And some edge cases required institutional knowledge that hadn't been captured in training data.

By 2025, CEO Sebastian Siemiatkowski admitted publicly that "cost was a predominant evaluation factor" in organizing support, which resulted in "lower quality." Customers weren't just complaining about wrong answers — they were complaining about how it felt to be served. The lack of empathy, flexibility, and real connection mattered.

Siemiatkowski even posted on social media: "We just had an epiphany: in a world of AI nothing will be as valuable as humans!"

The Pivot to Hybrid

Klarna began rehiring human agents and bringing work in-house. The company now offers 24/7 live chat with seamless handoffs from AI to human agents when needed. They even introduced a callback option for customers who prefer voice support.

The AI still handles about two-thirds of all customer chats — roughly 1.3 million conversations per month as of mid-2025. But the philosophy changed. As one Klarna spokesperson put it: "AI gives us speed. Talent gives us empathy. Together, we can deliver service that's fast when it should be, and personal when it needs to be."

Siemiatkowski described it this way: human customer service is becoming "a VIP thing" — a premium offering that differentiates companies in a world where AI handles the basics.

The Real Lesson

Klarna's pivot isn't a failure story. It's a maturity story.

The AI handled 2.3 million conversations successfully. It continues to handle millions more. But the company learned that customer service isn't just about resolving tickets — it's about building relationships. And relationships sometimes need humans.

For small businesses, this is actually good news. It means you don't need to choose between "all AI" and "all human." The winning approach combines both.

How AI Customer Service Actually Works

Let's strip away the jargon and explain what happens when AI handles a customer conversation.

Step 1: Understanding the question. Modern AI doesn't just match keywords. When a customer asks "I ordered something last week but it hasn't arrived and I'm leaving for vacation tomorrow," the AI understands there's urgency, there's a delivery issue, and there's time pressure.

Step 2: Finding the answer. The AI checks your business data — order status, shipping information, return policies, previous interactions with this customer. It pulls relevant information together.

Step 3: Generating a response. Instead of selecting from pre-written scripts, the AI writes a response specific to this situation. If the order is stuck in transit, it might offer to expedite a replacement or provide a refund. If the customer has been loyal for years, it might be more generous with the solution.

Step 4: Knowing when to hand off. Good AI systems recognize when they're out of their depth. A billing dispute with unusual circumstances? A customer who's getting more frustrated? A request that requires manager approval? The AI escalates to a human, along with a summary of the conversation so the customer doesn't have to repeat themselves.

Other Companies Getting It Right

Klarna isn't the only data point. Here's what we're seeing elsewhere.

Bank of America's Erica has handled over 3 billion customer interactions since launching in 2018. The virtual assistant serves nearly 50 million users and averages more than 58 million interactions per month. According to Bank of America, more than 98% of users find the information they need. The bank's data scientists have trained Erica to recognize millions of client questions using a library of more than 700 responses, with over 75,000 updates to continuously improve the system. Key insight: Erica succeeds because it's tightly integrated with account data — it knows what it's talking about.

Amtrak's Julie delivers an 8x return on investment and saves over $1 million annually in customer service email costs. The virtual assistant answers over 5 million questions per year. More interesting: customers who interact with Julie generate 30% more revenue per booking than those who don't, with a 25% increase in booking completion. The AI doesn't just answer questions — it helps customers find what they want.

Sephora's chatbot resolves over 75% of daily inquiries without human intervention. Response times dropped from minutes to under 10 seconds. The bot has increased conversion rates by 33% through tailored beauty recommendations based on customer preferences and purchase history. When the bot can answer "will this foundation match my skin tone?" with personalized recommendations, customers buy more confidently.

What Can Go Wrong (And How to Avoid It)

Honest assessment time. AI customer service fails when:

The AI doesn't know enough. If you feed an AI limited information, it gives limited answers. A common mistake is deploying AI without connecting it to your actual business data — order systems, product details, customer history.

Customers can't reach humans when needed. Nothing frustrates people more than being trapped in an AI loop when they have a complex problem. Always provide a clear path to human support. This is exactly what Klarna learned.

The AI makes stuff up. Language models can "hallucinate" — confidently stating incorrect information. This is dangerous in customer service where wrong answers cost money and trust. The solution is grounding the AI in your verified data and having it say "I don't know" when appropriate.

You measure the wrong things. If you optimize purely for "tickets resolved without humans," you might be pushing frustrated customers through automated systems when they need real help. Track customer satisfaction alongside efficiency metrics.

The AI's tone doesn't match your brand. A casual, emoji-heavy AI might work for a trendy D2C brand but feel wrong for a law firm. The AI's personality should match your company's voice.

What This Means for Your Business

Here's the practical takeaway.

AI customer service isn't about eliminating humans. It's about eliminating the mismatch between customer expectations and your capacity to meet them.

The 70% metric in our headline is realistic — that's roughly the portion of routine customer inquiries that well-implemented AI can handle, consistent with what Klarna and others have achieved. But the key word is "routine."

• Password resets? AI.
• Order status checks? AI.
• Basic product questions? AI.
• Return policy explanations? AI.
• Complex disputes with missing context? Human.
• Frustrated customers who need to feel heard? Human.
• Situations requiring judgment calls about exceptions? Human.

The businesses winning with AI customer service aren't the ones automating everything. They're the ones automating the right things, freeing their human team to do what humans do best.

Questions to Ask Yourself

Before considering AI for customer service, work through these:

What percentage of your support tickets are truly repetitive? Check your last 100 tickets. How many could be answered by copying and pasting from an FAQ? That's your automation ceiling.

Do you have the data? AI needs access to order systems, product catalogs, customer histories, and policy documents. If this information is scattered across spreadsheets and people's heads, you need to consolidate it first.

What does failure look like? If the AI gives a wrong answer about a refund policy, what's the cost? If it frustrates a key customer, what's the impact? Understanding downside risk helps you decide where AI should and shouldn't operate.

Who handles the escalations? AI will generate some handoffs. Do you have humans ready to receive them? Is the handoff process smooth?

What would you do with freed-up time? If AI handles 70% of routine inquiries, what does your team do instead? The best answer isn't "less" — it's "more valuable work like solving complex problems and building customer relationships."

The Bottom Line

Klarna's story isn't about AI failing. It's about a company learning that the goal isn't maximum automation — it's optimal automation.

AI can handle millions of routine customer conversations. It can respond instantly, consistently, and around the clock. That's real value.

But the companies getting this right aren't replacing their customer service teams. They're giving those teams superpowers: AI handles the repetitive work, humans handle the nuanced work, and customers get faster, better service overall.

That's not hype. That's just how the technology works when you implement it honestly.