AI Underwriting

A decision on 400+ pages, in minutes.

The Underwriting Workbench reads the whole application bundle — 17 document types — checks every finding against a range, cites it to source, and returns a priced, auditable recommendation. Straight-through where it's clear; the senior underwriter's judgment where it isn't.

UW-2026-01184 · Life + CI

Medical risk assessment

Loaded
01BaseStandard life
02+60Hypertension+60 debits
03+40HbA1c elevated+40 debits
04Band→ LOADED100 debits
05PriceApply +100% · Table 4
Every loading cited to a rule + guideline+100%

0%

Straight-through

0

Doc types classified

0%

Explainable

0+

Pages per case

Inside the workbench

From a bundle of documents to a priced, auditable decision.

Six stages, each a specialized step — and a clear line where the model recommends and the underwriter decides.

01

Ingest the bundle

The full application lands — forms, medical exams, labs, attending-physician statements, hundreds of pages.

02

Classify 17 doc types

Every document sorted and routed; nothing read out of context.

03

Extract + cite the page

Structured findings pulled from mixed-format, multi-language PDFs — each traced back to where it was read.

04

NIGO completeness

Not-in-good-order gaps flagged up front, so cases don't stall late.

05

Risk score — EPM + table

Signals summed into extra mortality per mille and a substandard table rating.

06

Decision + pathway

A recommended action with the full audit trail — or a routed referral.

Recommended actions

Eight outcomes — not just accept or decline.

Issue standard
Apply loading
Apply exclusion
Request documents
Postpone
Decline
Refer to senior
Partial cover offer

Two pricing levers

Priced the way underwriters actually price.

EPM — Extra Mortality Per Mille

Signals sum into a banded outcome — Absorb · Loaded · Decline — so borderline risk is priced, not bounced.

Substandard table rating

Debits convert to a percentage or table (25 debits = +25% = Table 1; +100% = Table 4), plus any flat extra per mille.

Explainable by design

Every loading traces to a rule and a guideline section.

Nothing is a black box. Each finding is checked against a range and cited to the page it was read from; each debit points at the rule and guideline that produced it. High-confidence cases auto-issue — everything else routes to a senior underwriter, whose override is captured with a reason code so the model learns.

Findings checked against a clinical range
Cited to the source page, auditable
Refer to senior — override with a reason code
100% explainable, days → minutes

Underwrite in minutes, stand behind every decision.

See the Workbench run a real application bundle end to end — with the full audit trail.