AI in insurance underwriting: what actually ships
Triage, evidence extraction and risk signals — the underwriting tasks AI genuinely improves, and the ones it should stay away from.
Underwriting is a judgement job wrapped in an enormous amount of document handling. The document handling is where AI pays; the judgement is where it should assist rather than replace.
Where to start, in order of risk
Intake first. Decisions last, if at all.
- 1
Intake extraction
Pull structured facts from the submission pile and flag what's missing. A human reads everything anyway.
- 2
Complexity triage
Route by difficulty. A misroute is an annoyance, not a mispriced risk.
- 3
Evidence surfacing
Comparable losses, non-standard clauses, unmentioned exposures — for the underwriter to weigh.
- 4
Decision support
Only with full traceability, bias testing and a documented human review path.
Start with intake, not decisions
A submission arrives as a pile of PDFs, spreadsheets and email threads. Extracting the structured facts — insured entity, exposures, loss history, requested limits — and flagging what is missing removes hours per case before any risk judgement happens. It is also the lowest-risk place to start, because a human reads everything anyway.
Triage by complexity
Not every submission needs a senior underwriter. Classifying incoming business by complexity and routing accordingly is a modest model with an outsized effect on cycle time, and it degrades gracefully — a misrouted case is an annoyance, not a mispriced risk.
Evidence, not verdicts
The systems underwriters actually adopt surface evidence and let the human conclude: prior losses that resemble this one, clauses that differ from the standard wording, exposures the submission does not mention. A tool that says 'decline' gets argued with. A tool that says 'here are the three things you would want to know' gets used.
The regulatory constraints are the design
Pricing and acceptance decisions are regulated in most markets, and 'the model said so' will not survive scrutiny.
- Every decision traceable to the evidence that produced it
- Protected characteristics excluded, and proxies for them tested for
- Model versions and prompts retained for the audit period
- A documented human review path for adverse decisions
Measure the business, not the model
Quality
- Tempting metric
- Accuracy against past decisions
- What the business feels
- Whether loss ratios hold
Speed
- Tempting metric
- Model latency
- What the business feels
- Time from submission to quote
Coverage
- Tempting metric
- Cases the model scored
- What the business feels
- Share handled without escalation
Risk
- Tempting metric
- Confusion matrix
- What the business feels
- Adverse decisions that survive review
| Tempting metric | What the business feels | |
|---|---|---|
| Quality | Accuracy against past decisions | Whether loss ratios hold |
| Speed | Model latency | Time from submission to quote |
| Coverage | Cases the model scored | Share handled without escalation |
| Risk | Confusion matrix | Adverse decisions that survive review |
Measure cycle time and leakage, not accuracy
Accuracy against historical decisions rewards copying past mistakes. The metrics that matter to the business are time-to-quote, the proportion of submissions handled without escalation, and whether loss ratios hold. Agree those before the build.