A production AI platform that reads an insurance application, applies the carrier's actual underwriting manual — fifty-plus discount, surcharge and risk rules — and hands the underwriter a fully reasoned recommendation. Days of review, done in three to five minutes — 96.14% accuracy overall, 99.55% precision on discount and surcharge rules.
| Value | What it is |
|---|---|
| 99.55% | Precision, discounts & surcharges |
| 3–5 min | Per policy, was days |
| 50+ | Manual rules automated |
| 18 | Carriers served by the host platform |
A North American mutual insurer priced home-insurance policies the way most of the industry still does: an underwriter sat with the application, the underwriting manual, and every supplemental document — questionnaires, appraisals, photographs — and worked through more than fifty discretionary discount and surcharge rules by hand.
Thorough, but slow. Each policy took days, and the backlog grew with the book.
I was lead engineer on the underwriting-insights product at a US AI consultancy, building on a platform sold through a policy-software vendor serving eighteen carriers. The brief was strict: apply the carrier's actual manual — not an approximation of it — and keep the underwriter as the decision-maker.
The pipeline evaluates the standardised application dataset a carrier already produces against a distilled, version-controlled copy of the underwriting manual, and emits five types of structured insight — discounts, surcharges, underwriter-review flags, unacceptable risks, and AI-flagged risks. Every insight carries an audit trail: the manual section it rests on, the worked calculation, and the recommended percentage.
| Step | What happens |
|---|---|
| Application data | Standardised submission JSON, plus supplemental documents |
| Rules applied | One prompt per insight category, against the versioned manual |
| Reconciled | Deduplication graph merges sources, flags contradictions |
| Underwriter decides | Reasoned recommendations, with citations to the manual |
Supplemental documents get their own treatment: a vision model lifts structure straight off scanned questionnaires and appraisal forms — benchmarked at 98.1% value accuracy across 709 fields on a standard six-page industry form — before Claude analyses the content. A LangGraph deduplication stage then reconciles documents against the application: complementary findings merged, duplicates collapsed, and genuine contradictions surfaced as explicit discrepancy flags rather than quietly resolved.
Applications arrive incomplete. Rather than let the model improvise, the pipeline degrades deliberately: anything the rules cannot establish becomes an "underwriter review required" flag with no percentage attached — an honest gap instead of a confident error.
An application might declare one oil-tank age and the questionnaire another. The deduplication graph prefers the primary source where policy dictates; where it doesn't, it raises a discrepancy flag — validated above 99% on conflict handling.
Batch processing was sequential because of tight AWS Bedrock quotas. I parallelised evaluation across dwellings and prompts, then added a fail-up chain that promotes throttled calls to the next model, layered with ordered multi-region failover.
Prompts evolve constantly; accuracy must not wobble. A golden-sample regression suite replays real production cases on every change, with numeric tolerances and an LLM-as-judge check on the reasoning — 99.55% is a maintained property, not a launch statistic.
| Metric | Outcome |
|---|---|
| Review time | Days per policy → 3–5 minutes (5–8 with supplemental documents) |
| Accuracy | 96.14% overall; 99.55% precision on discounts and surcharges, against 100+ human-graded production cases |
| Disputed calls | Where model and human graders disagreed pre-deployment, the model was right ~4 times in 5 |
| In production | Live with real carriers; every recommendation reviewed by an underwriter, by design |
This work sits under NDA, so the carrier and platform are anonymised. The figures come from the production system's own documentation and evaluation suites — happy to talk the architecture through in as much depth as confidentiality allows.
A designed PDF version of this case study is in this repo: 01-ai-underwriting-platform.pdf.
Freelance AI engineer — Expert-Vetted on Upwork (top 1%), 100% Job Success over 70+ projects, $400K+ earned, 5,750+ hours billed. I build production LLM systems for regulated industries: insurance, finance, law.