Skip to main content
INS-01 · Insurance

Claims Adjuster Copilot

Outcome​

Help a claims adjuster reach a defensible coverage and reserve decision faster by unifying the first notice of loss, policy terms, damage evidence, and repair estimates with mandatory human sign-off on payments.

Portfolio attributeValue
Primary personaClaims adjuster
Primary aggregateClaimFile
Cosmos DB containers1
Grounding dataclaims, policies, images, estimates, correspondence
Development estimateUSD 60-220 monthly dev estimate
Runtime optionsPython FastAPI and C# ASP.NET Core

Productivity hypothesis​

These measures reflect the workflow, evidence, decision boundary, and industry risk for this use case.

Impact typeMeasureConservative internal assumptionCustomer measurement
ProductivityClaim review cycle time8-14% reduction(baseline median claim-review minutes - assisted median minutes) / baseline median minutes * 100
Quality or riskReopened or corrected claims3-6% reduction(baseline reopened or corrected claim rate - assisted rate) / baseline rate * 100

Important: Conservative internal pilot assumptions only; not public benchmarks, commitments, or expected customer results. Validate against a comparable baseline or matched control cohort with mandatory human review.

Rationale: Claims, policies, images, estimates, and correspondence support faster evidence synthesis and recommendation preparation.

Pilot guardrail: Adjusters retain coverage, liability, reserve, and settlement authority with complete evidence lineage.

Customer validation plan​

  1. Baseline the current workflow with representative users and tasks.
  2. Run the accelerator on the same task mix with human review retained.
  3. Measure claim review cycle time and reopened or corrected claims with the documented formulas.
  4. Replace internal assumptions with customer-observed results before a business case.

Why Azure Cosmos DB​

This AI solution needs an operational data plane, not only a vector index:

  • Hybrid vector and full-text retrieval combines semantic recall with exact identifiers, codes, names, and domain language.
  • Hierarchical partition keys and metadata predicates keep tenant and domain boundaries in the normal query path.
  • Operational JSON, source chunks, workflow and agent state, feedback, and audit records remain together for low-latency reads and governed writes.
  • Change feed reacts to source and operational updates without polling.
  • Autoscale throughput, configurable consistency, and multi-region distribution support variable demand and enterprise availability.

Domain workflow​

Domain workflow

StageActorInput → outputApproval
1. Intake first notice of lossInsured or call agentLoss report with date, cause, and parties → ClaimFile with structured loss factsPolicy controlled
2. Bind policy terms in force at loss datePolicy servicePolicy number and loss date → PolicySnapshot with limits and exclusionsPolicy controlled
3. Analyze damage evidence and estimatesEvidence analysis agentPhotos, inspection reports, shop estimates → Damage observations and estimate variancePolicy controlled
4. Draft the coverage position and reserveCoverage reasoning agentLoss facts, policy clauses, evidence → CoverageAssessment with cited clausesPolicy controlled
5. Approve payment or denialClaims adjuster and supervisorCoverage assessment and settlement figure → Authorized payment or issued denial letterHuman required

Technical architecture​

Technical architecture

Cosmos DB data model​

ContainerHierarchical partition keyDocumentsSearch paths
claimFiles/tenantId, /claimIdClaimFile, PolicySnapshot, DamageEvidence, RepairEstimate, CorrespondenceThread, CoverageAssessment/endorsementTextVectorEmbedding, /observationVectorEmbedding, /bodyVectorEmbedding, /assessmentVectorEmbedding, /endorsementTextVector, /observationVector, /bodyVector, /assessmentVector

Query contract​

QueryContainerScopePurpose
Claim file assemblyclaimFilessingle-partitionLoad the complete claim context in one partition read for the adjuster desktop.
Estimate variance reviewclaimFilessingle-partitionHighlight estimates that exceed the guideline schedule by more than the review tolerance.
Unanswered insured correspondenceclaimFilescross-partition-boundedBounded cross-partition control-plane query: find inbound messages awaiting an adjuster response inside the service-level window. It cannot carry the second partition key because the service-level breach report covers every open claim in the unit, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by inbound messages received inside the last seven days with no respondedUtc, executed with MaxItemCount paging and continuation tokens.
Aging severe claimsclaimFilescross-partition-boundedBounded cross-partition control-plane query: surface open high-reserve claims that have not moved recently for supervisor review. It cannot carry the second partition key because supervisory oversight is by definition a portfolio view across claims, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by open claims above the severity reserve threshold with no update for ten days, executed with MaxItemCount paging and continuation tokens.

Python and C# services route each document type to its declared container and hierarchical partition key. ETag-based writes protect mutable aggregates, while append-only records and bounded cross-partition queries preserve scale.

Azure deployment​

Azure deployment

Clone the source repository, authenticate with Azure, and run:

Set-Location accelerators/insurance/claims-adjuster-copilot
az login
./Deploy.ps1 -Runtime python -Location eastus2

Change python to csharp to deploy the .NET implementation.

Explore the implementation​

Return to all accelerators