Direct pojišťovna: AI agents for windshield claims

Claims arrived as forms, photos, scans and emails in many different formats, and handling them was manual. Work stopped outside office hours, so cases waited over evenings and weekends.¹

With partner BigHub, the insurer built a modular, agent-based system on Azure AI Foundry (now Microsoft Foundry) and Azure Document Intelligence (now listed under Foundry Tools). It started with windshield claims and keeps people in the process. Once the system proved reliable, Direct added the payout step, so the AI-supported workflow now covers the claim from intake to payment, and processing runs around the clock.

Direct plans to reuse the design for car, property, liability and pet insurance by changing the data model, rules and document types. In a June 2026 interview, group head Pavel Řehák said AI is used most in claims, where it gathers documents, checks whether damage is covered and looks for signs of fraud in photos. The interview does not mention Microsoft or BigHub. It also reports that in March 2026 the group said it would need almost a third fewer staff because of AI, and that about a third of those announced departures had happened by June.¹²

In our view, Direct chose a sensible first target: one narrow, high-volume claim type with clear rules, which gives an insurer a measurable starting point before it extends to complex lines such as liability. Keeping people in the loop and adding the payout step only after the system proved reliable is a cautious order worth copying. The figures come from Microsoft's story without a published method, and the 60% figure changes scope across the page. The workforce context belongs next to the efficiency claims: the group has openly tied staff reductions to AI. If most claims already arrive through structured online forms, a rules engine with document capture may be enough without an agent-based design.