Klarna: AI Customer Service Assistant Saves $40M, Then Walks Back to Human-Hybrid Model

Title

Klarna: AI Customer Service Assistant Saves $40M, Then Walks Back to Human-Hybrid Model

Content

In February 2024, Klarna's OpenAI-powered customer service assistant handled 2.3 million chats in its first 30 days, automating 67% of customer service conversations and absorbing the workload of roughly 700 full-time agents — an outcome Klarna framed as approximately $40 million in annualised cost avoidance at launch, and which by Q3 2025 had grown to do the work of about 853 agents with annual savings closer to $60 million. The "700 agents" figure was work Klarna avoided hiring, not jobs cut: 2.3M monthly chats multiplied by an 11-minute pre-AI handle time yields roughly 421,000 hours per month, or about 2,630 agent-equivalents, of which the AI absorbed around 1,760. Customer outcomes held: average resolution time compressed from 11 minutes to under two minutes (an 82% improvement), 25% fewer repeat inquiries, and a customer NPS of 73 across 35-plus languages. The deployment mattered because Klarna was scaling into 150 million active users and 500,000 merchant partners across 45 countries, with tier-1 traffic (refund requests, payment reschedules, account lookups) growing faster than its support organisation. The assistant was built in roughly six months: project kickoff in August 2023, model selection and prompt engineering in autumn 2023, internal alpha testing against live Klarna data in December 2023, and a human-in-the-loop beta in January 2024 before the February 2024 public launch. Under the hood the system layered a GPT-4-class OpenAI model, fine-tuned through prompt and system-instruction work for banking and BNPL contexts, on top of direct API integrations into Klarna's account, transaction, payment, and refund systems so the AI could read customer state and take action (rescheduling payments, issuing refunds) without a human handoff. Klarna's help-centre articles, policy documents, and historical ticket resolutions formed a retrieval grounding corpus to suppress hallucination, and a confidence-based escalation layer routed the remaining roughly 30% of conversations to human agents with full context attached. In May 2025, however, Klarna publicly reversed course: CEO Sebastian Siemiatkowski admitted the company had cut too far — "what you end up having is lower quality" — and committed to an "Uber-type" model in which a customer can always reach a human, citing degraded experience on disputes, fraud, and emotionally complex tickets. Klarna began rehiring agents for those cases, tightening the AI's escalation thresholds, and improving handoff context, while keeping the AI doing the work of 853 agents. The lesson embedded in the 2025 walk-back is structural rather than a reversal of the original deployment: AI scales tier-1, humans own the complex 20%, and the strategic question is where the confidence floor and the guaranteed human option sit.