Insurance claims assessment with deterministic gates

The challenge

Claims assessors must check long estimates and itemised bills line by line for non-footing totals, duplicate billing and over-authorisation. The work is repetitive and error-prone, and an LLM that checks its own arithmetic can simply repeat the mistake it is meant to catch.

What we built

The system ingests a claim PDF, classifies it as motor or medical, and routes it to a domain-specific schema, prompt and arithmetic gate. A lead agent governs tool calling and may escalate to a stronger model on hard cases, but arithmetic reconciliation, schema validation and invariant checks run as deterministic Python gates the agent cannot override and that host code re-runs after the agent finishes. Findings are benchmarked against a rate card and written into the source PDF as anchored margin comments. A per-run budget guard and token metering escalate pathological files to a human, and a FastAPI dashboard adds role-based login, claim search, comments and Word report export.

Key capabilities

• Automatic classification of claims into motor (body-repair estimate) and medical-aid domains
• Deterministic checks for quantity x price, line sums, tariff totals, member liability and duplicate service lines
• Authorise / query / reject recommendation with severity-ranked findings
• Annotated copy of the source PDF with anchored margin comments
• Rate-card benchmarking and reference-document lookup
• Per-run cost metering with a budget guard, and a role-based web dashboard with Word report export

What it delivers

• Arithmetic verdicts come from plain code outside the model's discretion, so a model cannot rubber-stamp its own sums
• Every review ships with an annotated PDF so a human can verify each flagged figure at its source
• Token usage is metered per claim and a spend ceiling hands runaway cases to a person
• The motor and medical domains share one agentic spine, so a new claim type needs a schema, prompt and gate rather than a rewrite