Behind the scenes of unLocked's AI Underwriting: how carrier underwriting manuals are digitized, structured.
TL;DR
unLocked digitizes 100+ carrier underwriting manuals (200–500 page PDFs) into a structured, queryable rule engine. AI extracts condition rules, medication flags, and build charts, then normalizes terminology across carriers. The system achieves 94.2% agreement with actual carrier decisions and is updated quarterly.
Key Data Points
- 94.2% agreement with actual carrier underwriting decisions
- 100+ carrier underwriting manuals digitized into queryable rule engine
- Conservative bias: 5.1% of cases receive more favorable decisions than predicted
The AI Underwriting engine knows the underwriting guidelines of 100+ insurance carriers. Here is how that data is collected, structured, and kept current.
The Challenge: Unstructured Guidelines
Carrier underwriting guidelines are typically published as 200–500 page PDF manuals. Each carrier organizes differently:
- Some list conditions alphabetically
- Some group by body system
- Some have separate guides per product
- Rate-up tables vary in format
- Medication policies are often buried in appendices
There is no industry standard for how underwriting guidelines are published. This makes automated processing extremely difficult — which is why no one has done it before.
How unLocked Digitizes Guidelines
Step 1: Document Ingestion
Carrier underwriting manuals are ingested in their native format (PDF, DOC, web portal). The system processes:
- 100+ carrier guides
- Updated editions as carriers revise (typically quarterly)
- Product-specific supplements
Step 2: Structured Extraction
AI models extract structured data from unstructured documents:
- Condition rules: Diabetes Type 2 → Standard if A1C < 7.0, Table 2 if 7.0–8.5, Decline if > 8.5
- Medication flags: Metformin → acceptable; Insulin pump → substandard
- Build charts: Height/weight tables converted to BMI-based lookups
- Waiting periods: "No coverage for 12 months post-cardiac event"
Step 3: Normalization
Different carriers describe the same condition differently. The system normalizes terminology:
- "NIDDM" = "Type 2 Diabetes" = "Adult-onset diabetes"
- "COPD" = "Chronic bronchitis" = "Emphysema" (different severity weights)
- "Tobacco" definitions vary: some include e-cigarettes, some don't
Step 4: Rule Engine Compilation
Extracted rules are compiled into a queryable rule engine. When an agent enters a health profile, the engine evaluates every carrier's rules simultaneously and returns results in 15 seconds.
Keeping Data Current
Carrier guidelines change. The system maintains accuracy through:
- Quarterly re-ingestion of updated carrier manuals
- Change detection that flags modified rules for human review
- Carrier relationship team that confirms ambiguous changes directly with carriers
- Version history so agents can see when a carrier's guidelines last changed
Accuracy Validation
The system is validated against actual carrier decisions:
- 94.2% agreement with final carrier underwriting decisions
- 5.1% of cases receive a more favorable decision than predicted (conservative bias)
- 0.7% of cases receive a less favorable decision than predicted
The conservative bias is intentional — it is better to surprise the agent with an approval than a decline.