Insurance · Disability Insurance

Disability Claims Review: AI PoC with 98% Extraction Accuracy

Starting point

Claims review in Germany's disability insurance sector is highly complex and predominantly manual. Each case involves hundreds of pages of unstructured documents — medical certificates, findings, applications, correspondence, expert opinions — processed by highly qualified claims handlers. The industry-wide average processing time stands at around 190 days. Rising case volumes meet a persistent shortage of skilled professionals — a linear scaling trap that cannot be solved with conventional automation.

What we did

We set up a proof of concept for an AI-powered disability claims review platform and validated it along three proof axes: technical feasibility (extraction quality on real document portfolios), economic viability (per-unit processing costs versus specialist labor costs), and regulatory compliance (EU AI Act, BaFin supervisory practice, GDPR). Scope: architecture, model selection, data flow design, human-in-the-loop concept, compliance mapping. Duration: three months. Setup: makematiq as technical lead, domain advisory board from medical and insurance expertise.

Results

98 %

Extraction accuracy across all classes

187

Disability document classes covered

Proof axes: technical, economic, regulatory

EU AI Act

Compliant through HitL architecture

What we learned

The temptation in AI projects is to throw a large model at everything and hope for scale. Robust results emerge differently: by deploying human and technical intelligence precisely where each excels. Reverse that, and you build demos. Follow through, and you build products.

This is the summary. How we approached it methodologically — which architectural decisions we made, what we discarded and which patterns can be transferred to other contexts — we discuss in a personal conversation.

Not because we want to sell you something. But because this depth is what our clients engage us for — and it does not belong on the open internet.