The Timeline Nobody Expected

In late 2023, Klarna announced they were using OpenAI to handle customer support. By early 2024, their AI was handling two-thirds of all customer service chats in its first month. CEO Sebastian Siemiatkowski said it was doing "the equivalent work of 700 full-time agents." The company had already frozen hiring and was letting headcount shrink through attrition, dropping from roughly 5,000 employees to under 4,000.

The stock market loved it. Tech blogs covered it breathlessly. Other companies started asking "should we do what Klarna did?"

Then things went sideways.

What Went Wrong

Klarna's customer satisfaction scores dropped. The AI handled volume well but struggled with anything beyond straightforward queries. Customers couldn't reach humans for complex issues. The bot would loop on problems it couldn't solve, and there was no easy escape hatch.

Refund disputes, account issues requiring manual review, edge cases in buy-now-pay-later terms: the AI stumbled on all of them. And because Klarna had cut so many human agents, there wasn't enough staff to pick up what the AI dropped.

By May 2025, Siemiatkowski publicly admitted they'd gone too far. He acknowledged that quality had dropped and that the AI-only approach wasn't working for complex cases. The company started rehiring, this time with an Uber-style flex model: gig workers (students, parents, rural workers) who can take support shifts on flexible schedules.

The Numbers

Before AI: roughly 3,000 outsourced support agents handling customer queries. Company headcount around 5,000.

Peak AI: AI handling 66% of conversations, equivalent to 700 agents' workload. Total company headcount dropped to around 3,800.

After the U-turn: rehiring with flexible contractors. Customer satisfaction back as a top priority.

Cost savings were real in the short term. But the customer satisfaction hit threatens the long-term brand. In financial services, trust is the product. Klarna learned that the hard way.

Why AI Struggles With Support (Specifically)

A 2025 Qualtrics study found that AI-powered customer service fails at 4x the rate of AI in other tasks. Support is uniquely hard for AI because:

Context matters. The AI needs to know your order history, your account status, your previous conversations. Most chatbots get a blank slate every session.

Emotions matter. An angry customer who's been double-charged needs to feel heard. AI can generate empathetic-sounding text, but customers can tell the difference.

Actions matter. Support often requires doing things: processing a refund, updating an address, escalating to a specialist. Answering questions is easy. Taking the right action on a customer's account is risky.

Stakes are high. When a coding AI gets something wrong, a developer catches it. When a support AI tells a customer the wrong refund policy (like the Air Canada case, where the chatbot made up a bereavement fare policy and a tribunal ordered the airline to pay C$812 in damages), there's real financial damage.

What Klarna Should Have Done

Start with the easy stuff. Automate "where's my order" and "what are your hours" and "how do I update my address." These are high-volume, low-risk, and AI handles them well. That alone could have handled 30-40% of volume without touching the complex stuff.

Keep humans for anything involving money, disputes, or emotions. Refund decisions, payment plan modifications, and complaints about being charged incorrectly should always have a human in the loop.

Build escalation paths that actually work. The biggest complaint from Klarna's customers was being stuck with a bot that couldn't help and couldn't transfer them. Every AI interaction needs a clear, easy path to a human.

Measure satisfaction, not just resolution rate. If the AI "resolves" a conversation but the customer is furious, that's not a resolution. Track CSAT on AI-handled conversations separately.

The Lesson for Everyone Else

Klarna proved that AI can handle massive volumes of support. They also proved that handling volume isn't the same as handling support well.

The companies getting this right in 2026 use AI for classification and simple resolutions while keeping humans for everything that requires judgment. They treat AI as a tool for their support team, not a replacement for it.

A 50-70% automation rate with high customer satisfaction beats a 90% automation rate with angry customers every time.