Claims processing: from 14 days to under 48 hours

An African financial services group cut the time needed to process claims by more than 85% using an AI system built with Cognis Group and kept under clear human and risk controls. The client's name is confidential.

The problem

A financial services group working across several African markets was taking an average of 14 days to complete a standard claim, while some took more than 30 days. The board, customer service and internal audit agreed that the process had to improve, but there was no agreed solution.

Two earlier trials could sort test claims, but neither reached everyday use. The risk committee had not seen enough proof that the systems would handle real claims safely.

The approach

We spent four weeks understanding the process before writing any code. The biggest problems were scattered claim data across three systems, no clear rules for AI-assisted decisions and no reliable test that the risk committee could review.

What we built

• An AI claims assistant that sorts new claims, reads attachments, checks cover, flags possible fraud and gives a prepared summary to the human claims officer.
• Clear action limits. The system can recommend a settlement below an agreed amount. A person must decide anything above it, and every check is recorded.
• A test based on 1,200 past claims. Senior claims officers graded the correct answers. The system had to meet agreed accuracy levels before it could move forward.
• Regular risk checks. The team reviews quality every week, reports to the risk committee every month and prepares evidence for an external audit every quarter.

Rollout

For three weeks, the system worked in the background without affecting customers while we measured its accuracy. For the next five weeks, a claims officer reviewed every recommendation. It was then allowed to handle only the narrow group of claims approved by the risk committee. All other claims still go to a person, together with a prepared summary that removes much of the manual paperwork.

Results

Customer satisfaction with the claims process improved during the first quarter of supervised use. At the twelve-week review, the risk committee approved a wider rollout.

Further reading

• AI Agent & Automation Engineering — the service behind this build.
• AI Governance Is Not Optional — why the governance design, not the model, is what carried this engagement.
• MarketSage — the reference implementation for the agent patterns used here.