Skip to main content
INS-02 · Insurance

Underwriting Submission Workbench

Outcome​

Turn unstructured broker submissions into a triaged, priced risk file by extracting exposures, normalizing loss runs, and testing appetite and guideline fit before an underwriter binds or declines.

Portfolio attributeValue
Primary personaUnderwriter
Primary aggregateSubmissionFile
Cosmos DB containers2
Grounding dataapplications, loss runs, guidelines, risk data
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
ProductivitySubmission preparation time8-12% reduction(baseline median submission-preparation minutes - assisted median minutes) / baseline median minutes * 100
Quality or riskCorrected or reopened packets3-6% reduction(baseline corrected or reopened packet 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: Applications, loss runs, guidelines, and risk data can be assembled faster without automating underwriting judgment.

Pilot guardrail: Underwriters retain risk selection, pricing, terms, and approval authority.

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 submission preparation time and corrected or reopened packets 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. Ingest broker submission packageSubmission intake serviceBroker email with schedules and loss runs → SubmissionFile with attachment inventoryPolicy controlled
2. Extract and normalize exposures and lossesDocument extraction agentSpreadsheets and carrier loss-run PDFs → ExposureSchedule and LossRunRecord setsPolicy controlled
3. Score appetite fit and triage priorityTriage agentNormalized exposures and guideline clauses → Triage score with referral reasonsPolicy controlled
4. Produce hazard summary and pricing indicationUnderwriting analysis agentBurn cost, hazard data, capacity limits → RiskAssessment with cited rationalePolicy controlled
5. Quote, refer, or declineUnderwriterRisk assessment and pricing indication → Issued quote, referral to senior authority, or declinationHuman required

Technical architecture​

Technical architecture

Cosmos DB data model​

ContainerHierarchical partition keyDocumentsSearch paths
submissionFiles/tenantId, /submissionIdSubmissionFile, ExposureSchedule, LossRunRecord, RiskAssessment/hazardSummaryVectorEmbedding, /hazardSummaryVector
underwritingGuidelineLibrary/tenantId, /guidelineIdUnderwritingGuideline/clauseVectorEmbedding, /clauseVector

Query contract​

QueryContainerScopePurpose
Submission file assemblysubmissionFilessingle-partitionLoad schedules, loss runs, and the assessment for one submission in one partition read.
Bind-by-date work queuesubmissionFilescross-partition-boundedBounded cross-partition control-plane query: prioritize undecided submissions by effective date and triage score. It cannot carry the second partition key because an underwriter's queue spans every submission assigned to them, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by submissions in triage whose target effective date falls inside 30 days, executed with MaxItemCount paging and continuation tokens.
Adverse loss history scansubmissionFilessingle-partitionIdentify submissions with large incurred losses in the recent policy years.
Low-confidence extraction reviewsubmissionFilescross-partition-boundedBounded cross-partition control-plane query: route schedules whose extraction confidence is too low for automated pricing. It cannot carry the second partition key because the extraction quality queue is a pipeline-wide control surface, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by schedules below the confidence floor on submissions still in triage, 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/underwriting-submission-workbench
az login
./Deploy.ps1 -Runtime python -Location eastus2

Change python to csharp to deploy the .NET implementation.

Explore the implementation​

Return to all accelerators