Declarations, the base policy and endorsements each carry a different part of the coverage answer.
← All case studies
Case study / Enterprise delivery
Commercial insurance policy analysis with verified citations
Primus built a structure-aware policy analysis workflow that connects declarations, base policy language and endorsements while keeping every answer tied to verifiable evidence.Declarations + Base policy + Endorsements→Structure-first retrieval→Cited provisions + Adjuster decision
01 / Operating situation
An enterprise commercial-insurance platform supporting important adjuster review across varied carrier formats.
Commercial coverage questions draw on declarations, the base policy, endorsements that modify it and carrier-specific document conventions. Manual review required repeated cross-referencing across long policy packages. The new approach needed to preserve every relationship required for safe interpretation.
A single question can require locating and reconciling language across a long legal document.
Find the clause, identify modifying endorsements, confirm limits and reconcile the result.
Naming, footer and endorsement conventions vary while the underlying review obligation stays consistent.
02 / Design decisions
The architecture followed the operating problem.
Primus designed a structure-first retrieval system: multiple extraction methods preserve page content; form numbers connect related policy documents; a hierarchy and structure-aware chunks retain context; carrier-specific collections improve retrieval; reranking sharpens the evidence set; and code validates citations before an answer reaches the user. The system removes unresolved references. The adjuster remains responsible for the coverage decision.
- 01
Treat structure as evidence
Rebuild each document as a hierarchy so a provision remains connected to the section and definitions that govern it.
- 02
Validate citations in code
Check every reference against retrieved material as the answer streams and remove unresolved references.
- 03
Configure carrier behavior
Hold schemas, extraction patterns, prompts and thresholds as records so carrier onboarding follows one managed process.
- 04
Keep judgment with the adjuster
The platform assembles what the documents say; confidence and abstention controls stop unsupported certainty.
03 / What Primus built
One structure-preserving path from raw policy package to reviewable answer.
The design carries the policy's own relationships through extraction, hierarchy, retrieval and answer validation, preserving the package as a connected whole.
- 01
Extract
Use multiple extraction paths for digital PDFs, scans and structured fields.
→ - 02
Relate
Use form numbers to connect endorsements with the applicable declaration and policy.
→ - 03
Structure
Reclassify headings, build a hierarchy tree and assign stable identifiers.
→ - 04
Chunk
Cut at the document's branching points so provisions retain governing context.
→ - 05
Retrieve
Search carrier-specific collections in parallel and rerank the strongest evidence.
→ - 06
Verify
Resolve citations in code, strip unsupported references and present the evidence for human review.
Technology & methodsAzure Cosmos DB vector searchAzure OpenAICohere RerankAdobe PDF ServicesAzure Service BusOpenTelemetry
04 / Business results
What changed in the work.
The results below describe the operating value supported by the case materials and the delivery design.
Traceable answers
A reviewer can inspect the governing passage behind each summary.
Unresolved citations shown
Only validated references reach the user.
Carrier onboarding model
A new format is shaped through managed records in the shared application.
Decision authority
The system recommends and assembles; the adjuster reviews, edits and decides.
05 / What travels
The domain is specific. The engineering lessons are reusable.
Document structure is part of meaning
In regulated and contractual work, the relationship among documents can matter as much as the retrieved passage.
Scaling depends on the onboarding model
Configuration creates a sustainable path across carriers and policy types within one codebase.
Responsible AI has a visible stop
Confidence, abstention and citation resolution define where automation ends and accountable review begins.
The document structure was preserved, so the citations held. The adjuster gained a faster review path while retaining coverage authority.
Case-study takeaway
About this caseEnterprise delivery engagement
This case explains the operating problem, the important design decisions and the resulting business value.
Insurance