Claims processing: from 14 days to under 48 hours

An African financial services group reduced claims cycle time by over 85% with a governed AI agent system built and deployed with Cognis Group. Client name withheld under confidentiality.

The problem

A financial services group operating across multiple African markets was losing customers at the claims moment. Average cycle time on a standard claim was 14 calendar days, and the variance was worse than the mean — a non-trivial tail stretched past 30 days. Internal audit, customer service, and the board all agreed it had to change; no one agreed on the shape of the fix.

Earlier attempts had stalled at pilot. Two previous proofs-of-concept had demonstrated automated triage on synthetic data and then never reached production, because nobody could satisfy the risk committee that the system would behave safely on real claims that carried real financial exposure.

The approach

We ran a four-week assessment phase before touching any code. The assessment found the real blockers were not technical — they were the data lineage of incoming claims (fragmented across three systems of record), the absence of a governance framework for automated decisioning, and the lack of an evaluation harness the risk committee could audit.

What we built

• A governed agent system that triages incoming claims, extracts structured data from unstructured attachments, checks policy eligibility, flags suspected fraud signals, and routes to a human adjudicator with a pre-built summary.
• Policy and guardrail layer enforcing role-based action limits: the agent can recommend settlement up to a monetary threshold, above which a supervised human decision is mandatory. Every policy check logged.
• Evaluation harness of 1,200 real historical claims, graded by senior adjudicators. Model promotion required 98% agreement with adjudicator ground truth on eligibility and 95% on settlement amount within tolerance.
• Governance operating cadence — weekly drift review, monthly risk committee reporting, quarterly external audit pack — aligned to NDPA 2023 and the client's internal risk framework.

Rollout

The system ran in shadow mode for three weeks (no customer impact, pure accuracy measurement), supervised mode for five weeks (every auto-decision reviewed by an adjudicator before release), and only then moved to autonomous mode for the narrow eligibility band where the risk committee had explicitly approved it. Everything else still routes to a human — but with a pre-built summary that has eliminated the document-assembly work adjudicators used to do by hand.

Results

Customer NPS on the claims journey moved by a reportable margin within the first quarter of supervised production. The risk committee, which had vetoed two prior attempts, approved a broader rollout at the twelve-week review.

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.