An AI arbiter for disputes between customers and merchants

The challenge

When a purchase goes wrong, Klarna sits in the middle: the customer on one side, the merchant on the other. Disputes were handled largely by hand — evidence chased over email, cases bouncing between teams, decisions slow and hard to keep consistent. At Klarna’s scale, that doesn’t hold.

What we built

• A new disputes experience in React Native and React: customers raise a dispute in the app and upload photos, receipts and messages at the moment it happens — so every case starts complete.
• AI evidence review: models read the images and text a customer submits and check them against the disputes rules and guidelines before a case moves on.
• A full case workflow behind it — escalation, merchant responses, back to the customer — with every state and transition modelled explicitly.
• The AI arbiter: it weighs the customer’s case and the merchant’s against the evidence, the contract and the policies, and produces a structured, evidence-backed recommendation for a human to approve. Measured continuously with evals as models and prompts change.
• A chargeback system that tests whether a claim stands up before it reaches the card networks.
• An embeddable version of the disputes flow, so a major US partner could run it inside their own platform.
• AI tooling around it: a classifier that cut disputes raised under the wrong reason, a policy assistant over merchant return and refund rules, and an AI investigator teams can question from Slack.
• Security hardening: 15 live services moved off a legacy all-in-one pipeline onto isolated, least-privilege CI/CD steps, so no single step can reach what it shouldn’t.

The outcome

Disputes now arrive with the evidence attached, move through a defined workflow, and get an AI recommendation grounded in the rules — across 1.4 million cases, for a disputes experience used by 15 million customers a month.