AI & ML Quick Path

Company case

Klarna: Agents / customer support

CommerceIntermediateRebuild effort 30–40 h (estimate)

What the published sources confirm

In February 2024 Klarna announced an OpenAI-powered assistant in its app handling customer-service chats, including refunds and returns, in many languages. In May 2025 the CEO told Bloomberg the company was again investing in human customer support because cost had been too dominant a factor and quality had suffered.

Business problem and AI concept

Problem
High volumes of routine payment, refund and return questions across many markets and languages.
Concept
Tool-using assistant over account APIs, with escalation to people.

Reported results

  • Klarna reported 2.3 million conversations in the first month, two-thirds of its customer-service chats.

    Klarna press release, 27 Feb 2024. Caveat: Launch-month figure published by the company; not a current KPI.

  • In May 2025 Klarna said it was hiring human support staff again to protect service quality.

    Bloomberg interview reported May 2025. Caveat: Shows that automation rate alone is not a sufficient success metric.

Suggested workflow

Written for learning, not the company's system. This flow shows how you could build something similar. It is not a description of Klarna's private architecture.

  1. Step 1: Receive the customer message and detect language and intent

  2. Step 2: Authenticate the customer and fetch their own orders only

  3. Step 3: Answer from policy or prepare an action such as a refund

  4. Step 4: Require approval or strict rules for money-moving actions

  5. Step 5: Hand off to a person on low confidence or on request

  6. Step 6: Measure resolution, satisfaction and re-contact rate together

Suggested technologies and alternatives

LayerSuggestedAlternatives
OrchestrationMicrosoft Agent Framework (C#)LangGraph (Python)
Data accessTyped order and refund APIsGraphQL gateway
StatePostgreSQL + RedisDurable checkpointer
QualityConversation evaluation set + human review sampleVendor evaluation tools

Small example

Input
"I returned the shoes two weeks ago, where is my money?"
Output
Assistant finds the return, sees the refund is scheduled, gives the date, and offers a person if the customer disagrees.

Rebuild it yourself

Build a support assistant with read-only order lookup, a refund tool that requires approval, and human handoff.

Definition of done

  • Never shows another customer's orders in tests
  • Refunds always pause for approval
  • Handoff triggered for 5 scripted angry or ambiguous messages
  • Report resolution rate AND satisfaction proxy, not one alone

Security and privacy risks

  • Optimising automation rate at the cost of quality
  • Prompt injection through customer messages
  • Unauthorised refunds

Sources

Listed by title and publisher. This guide contains no external links; search for the title if you want to read the original.

  • Klarna AI assistant handles two-thirds of customer service chats in its first month. Klarna, 27 Feb 2024. Company press release.
  • Klarna is hiring customer service agents after AI couldn't cut it on calls. Entrepreneur, 9 May 2025. Independent reporting.