Why insurance, and what the measured returns cover
Industry research identifies opportunities to improve insurance operations with AI (McKinsey, BCG; see the references behind the integration phases). Submissions, loss runs, policy wording, broker email, and claims files offer concrete tasks to test: extraction, comparison, and drafting, with a person checking the result.
Reported results belong to particular workflows, baselines, and measurement periods. They do not establish a general time-saving range for all assistive work, or prove that one level of automation always wins. Start with AI drafts, a person decides, then measure the complete workflow before changing the review gate.
Measurement guide
Compare results within a defined workflow
Use the same task, quality threshold, and start/end points before and after a change. The reported examples later on this page are individual cases.
Count drafting time plus human checking, corrections, and rework. Track errors that survive review.
Count normal processing plus exceptions, retries, and escalation. Check that the same quality threshold still holds.
Suggested measurement approach, not a plot of measured savings. See Costs & value for whole-task economics.
A suggested order for evaluating use cases is set out on Costs & value; anything that influences underwriting, pricing, claims or customers needs the controls in Governance before it goes near a decision. Use the examples below to identify a similar bottleneck in your own business; their results do not predict your return.
Start with the preparation before a decision
Personal lines and standardized small-commercial products increasingly run on structured data, rules, and straight-through processing. Specialty and other large-complex accounts stay bespoke, document-heavy, and expert-driven: manuscript wording, layered exposures, broker negotiation, loss runs, engineering reports, and exposure schedules all need account-level judgment (Capco, which also puts 30–40% of underwriter time in administrative work).
The variable that moves is account size, not the specialty label. Accenture's survey of 434 P&C underwriters found that the share of time going to non-core and administrative work falls as accounts get larger, from 47% on books averaging under $10,000 in premium to 35% above $250,000, while the share going to risk analysis and pricing climbs from 22% to 33% (Accenture, 2021 P&C Underwriting Survey). Read against a specialty book of 50 respondents, the specialty and commercial splits are near-identical; account size is the clearer distinction in this survey. The data do not establish that specialty always offers greater AI returns.
Figure
As accounts grow, less of the month goes to administration
Share of an underwriter's time on each kind of work, at the two ends of the account-size range. The administrative share falls; the risk-analysis share rises.
Source: Accenture, 2021 P&C Underwriting Survey: shares of time, not hours per account; only the published endpoints of the account-size range are shown. Only two work categories are plotted, so the bars do not sum to 100%.
Two cautions on that figure. It measures shares of a month, not hours per account, so it says how underwriter time is divided rather than how much of it a single account consumes, and no public source we could find gives the absolute number. It is also five years old, and Accenture's later work reports incremental improvement rather than a different shape.
The survey places roughly a third of time on the largest accounts in non-core or administrative work. It does not measure how much AI can remove or when that work occurs. A submission can wait hours or days while emails and attachments are read, keyed, validated, and made decision-ready. Hiscox described manual extraction in one London Market workflow as taking up to three days (Hiscox); Arch reported a two-to-three-day intake-to-decision cycle before automating submission intake (Ivans/Arch). The capacity cost compounds: Deloitte reports that many specialty insurers process only 20–30% of the submissions they receive, because manual reentry and disconnected workflows cap how much of the flow a team can look at (Deloitte).
That number points to a processing-capacity constraint. It does not establish how many submissions went entirely unread or why each one was not processed.
Figure
Reported processing covers a minority of submissions
Deloitte reports that many specialty insurers process 20–30% of submissions received. It does not give a reason for every unprocessed submission.
Source: Deloitte: many specialty insurers process 20–30% of submissions received. Solid fill reaches 20%; stripes extend to 30%. This is a reported range, not a confidence interval.
| Personal / standardized small commercial | Specialty / large-complex accounts | |
|---|---|---|
| Account shape | High volume, repeatable questions, structured data, filed products | Lower volume, bespoke terms, layered or unusual exposures, unstructured broker packs |
| Human role | Algorithms handle routine cases; people manage exceptions | Experienced underwriters interpret risk, negotiate terms, and retain pricing/bind authority |
| Best automation target | Straight-through decisions on simple risks | Make the account decision-ready: ingest, extract, enrich, check appetite, surface precedents, draft the memo |
| Why agents pay | Reduce transaction cost at scale | Return scarce expert hours on every account and let the same team evaluate more of the submission flow |
A pilot pattern: prepare a submission for review
A triage pilot can flag incomplete or potentially out-of-appetite submissions for review. Where the work is independent, several agents can run in parallel: one worker extracts exposures, another checks guidelines, another enriches external data, another compares pricing and portfolio constraints, and an orchestrator assembles the decision-ready view. Keep a named underwriter responsible for the consequential decision, and measure whether preparation time falls after exceptions and corrections are included.
Illustrative workflow
Prepare in parallel; keep exceptions and decisions visible
The workflow assembles evidence for an underwriter. Missing or conflicting information follows a referral route before it can support a decision.
Inputs
Approved source pack Submission, loss runs, and broker correspondence Current appetite guide and permitted external dataIndependent preparation
Assemble and check
Draft briefSource links, reconciled fields, and explicit exceptionsHuman authority
Underwriter decisionReview the evidence; approve the consequential actionException route
Missing, conflicting, or outside appetite?Hold the affected field or recommendation. Preserve the source and the reason.Worked example: a loss run does not identify its valuation date. The brief marks the date as unresolved and links to the document. The underwriter requests clarification before using it in the risk assessment.
Illustrative operating design, not a reported carrier implementation. Measure the complete path, including referrals and human review.
Reported results illustrate different parts of that work: AIG reports processing more than 370,000 Lexington submissions in 2025; Cytora reports a 113% increase in gross written premium per full-time equivalent at Markel; Arch cut intake-to-decisioning by 70%; and Hiscox demonstrated three days to three minutes for a terrorism-risk quote in its 2023 proof of concept. These are carrier- or vendor-reported results rather than universal benchmarks; the source passages are collected on the evidence page.
Figure
Reported results on specialty-shaped work
Different workflows, metrics, and stages of deployment. These reported results cannot be ranked or averaged together, or attributed to generative AI alone.
Sources: AIG · Markel · Arch (Ivans) · Hiscox; passages on the evidence page.
Check the regulatory position before expanding authority
The NAIC describes surplus lines as generally outside the rate and form rules that apply to admitted business. That flexibility can matter when changing a product or process. It does not establish that all agentic deployments began in E&S or that a particular AI use is exempt from oversight.
Have Legal/Compliance assess the carrier's domicile, licensing status, states of placement, and applicable consumer protections. Use the controls in Governance as a planning framework; confirm which state requirements apply to the specific entity and use case.
Choose a recurring document task with a measurable baseline and a reviewer who can check the answer. Name the business owner, record current handling time and error rates, and agree a quality threshold before testing. In another industry, use the same approach for contract review, service requests, or internal reporting; rebuild the business case and controls for that setting.
Where this goes next
The phased plan built on this case is the integration phases, whose competitor evidence and gated phases assume specialty economics throughout. The cost side, and what thin claims data does to the credibility of automated model search, is on Costs & value.