Illustrative engagement — describes the type of work we deliver; client details are withheld.
Business challenge
Handlers spent significant time reading lengthy claim files and searching policy wording to make coverage decisions.
Existing environment
- Claims system with scanned documents attached
- Policy wordings stored as PDFs
- Manual summarisation by handlers
Technical challenges
- Accuracy and traceability of AI output
- Protecting personal data
- Integrating into the existing claims workflow
Our approach
- 1Defined measurable success criteria with the claims team
- 2Built a retrieval pipeline over policy documents with citations
- 3Created an evaluation set and guardrails before rollout
- 4Embedded the assistant in the claims screen with human sign-off
Architecture overview
- DocumentsOCR and chunking of claim files and policy wordings
- RetrievalVector index with metadata filters per policy
- GenerationLLM with prompt templates, guardrails and citations
- ExperienceReact side panel inside the claims application
Technologies used
- Python
- LLMs
- Vector database
- Azure
- React
Implementation
- Phase 1
Frame
Use-case selection and evaluation criteria
- Phase 2
Prototype
Retrieval quality testing on sample files
- Phase 3
Pilot
Controlled rollout with feedback capture
- Phase 4
Scale
Monitoring, cost controls and wider rollout
Results
- Faster review of lengthy claim files
- Every answer linked to its source clause
- Human decision retained for all coverage outcomes
Business impact
Handlers spend more time on judgement and customer contact, with an audit trail for every AI-assisted step.