// Underwriter
We're not touching the part that's actually underwriting.
We are after the assembly, the re-keying, the chasing, and the reconciling of documents that contradict each other. The file should arrive readable, with every value linked to the page it came from. What you still do is read the risk.
// The situation
The clerical share of the day is the part a build can reach.
The canonical split of an underwriter’s day is 40% administrative, 30% negotiation and sales support, and 30% actual underwriting.
Accenture · 2021 survey · dated.
The 40% is where we work. The 30% that is underwriting is yours.
That figure is five years old and the whole industry still quotes it without a date, so week one is observation rather than a business case: we count touches on your real submissions, by stage, before anything is scoped. That count is the first deliverable.
// The stack you actually run
The material fact is in the email body.
A submission is not one document. It is a set that disagrees with itself, arriving over four days, with the fact that decides the risk sitting in the third paragraph of a broker’s email rather than on any form.
64% of underwriters said technology had increased their workload or made no difference, and 96% said pricing technology needs improvement.
hyperexponential · State of Pricing · undated in source · vendor research.
“Technology has made it easier for more workloads. It helps with better decision-making, but it adds time for each submission.”
Underwriter, survey respondent · hyperexponential / Coleman Parkes · n=350.
Fear of replacement collapsed in a single year in the same study — underwriters “not worried” moved from 26% to 52% — while 73% to 74% reported burnout concerns. This is not a role that needs reassurance about its job. It needs the queue to stop arriving broken.
// The build
Four things, and the fourth is the one you'll test us on.
The file arrives assembled
ACORD forms, the SOV, loss runs, supplementals and the material fact buried in the broker's email — extracted, cross-checked and presented as one record, with every value linked to its source document and page. Checking a number costs one click instead of a re-read.
Surface: submission mailbox and ACORD AL3 into the submission record; write-back through Cloud API, versioned.
Contradictions are surfaced, not resolved
When the SOV total and the ACORD TIV disagree, when the driver count on the 127 doesn't match the schedule, when garaging addresses conflict — you see the conflict and both values. The system does not quietly pick one and hand you a clean number.
Surface: a reconciliation layer over the extracted record, both values retained with provenance.
Appetite screening before the queue
Class, limit, geography and authority boundaries checked at intake, with a citation to the governing guideline. Out-of-appetite risks never reach your desk; borderline ones arrive with the reason attached. The screening rules are yours to own.
Surface: the guideline corpus plus intake rules, versioned alongside your underwriting manual.
Referrals that explain themselves
When a rule fires you see which one, why, what the file would need to clear it, and what you are authorized to do about it — instead of a code and a queue. Named human, named authority, recorded rationale.
Surface: the referral queue in your policy system, with the rationale written back to the file.
What we measure
Touch count per submission. Time from submission receipt to first underwriter decision, on the same cohort definition both sides of the build. Rework rate, meaning how often a file goes back for missing data. And correction rate on extracted values, tracked as a first-class metric rather than hidden inside an accuracy claim. If the baseline cannot be defined in week one, that is the first finding.
// We ran one of these
A badly assembled submission is a cost before it is an inconvenience, and it is paid by the underwriter who has to work it.
The operating record →// What's hard about this
Two things we tell you in week two.
The 40% is from 2021 and everyone still quotes it.
The whole industry builds business cases on one survey, undated. So we observe your team for a week before anyone funds a number. If the real administrative share is 25%, the project is smaller than someone told you, and you should know that before you fund it rather than after.
Compressing the clerical day does not make the risk easier to judge.
The assembly, the re-keying and the reconciliation are addressable. Reading a marginal habitational schedule in a hardening market is not, and no extraction pipeline changes it. So the build is scoped against touch count and cycle time, never against a judgment multiple.
// What ships with it
The authority record, not just the answer.
Every screened submission retains the guideline citation that governed it. Every referral retains the rule that fired, the named human who cleared it, and the rationale they recorded. The extraction carries a model inventory entry with its risk tier, data lineage naming the provenance of each document, pre-deployment testing results, and drift thresholds with remediation triggers.
Roughly half the states have adopted the NAIC AI Model Bulletin, and the variations are material: Connecticut requires an annual AI compliance certification attested by a named officer, and Virginia’s bulletin says eliminate the risk rather than mitigate it.
NAIC and state bulletins · current at September 2026.
What the governance file contains →The file should arrive readable.
Nothing on this page automates a judgment call. It moves the assembly, the chasing and the reconciliation off your desk, and it shows you the contradictions instead of resolving them quietly.