A commercial insurance submission is a broker email thread. Attached to it: ACORD forms, loss runs, statements of values, spreadsheets, sometimes photographs of the property. It may be incomplete, duplicated across brokers, or outside the carrier’s appetite entirely. Before any underwriting judgement happens, someone has to read all of it, work out whether the account is worth pursuing, and assemble the facts into a shape the rating and pricing work can use.
That front end is where AI is actually landing in commercial underwriting. The production pattern in 2026 is not autonomous quoting. It is a human-in-the-loop copilot that converts broker submissions into structured risk data, ranks them against carrier appetite, and prepares an underwriter-ready package. The judgement layer, appetite exceptions, pricing, terms, wording and authority, stays with people. What changes is how much of the day gets spent reaching the point where judgement can begin.
From inbox to structured submission object
Ingestion comes first: broker emails, ACORDs, attachments, loss runs, statements of values and images pulled from every intake channel and normalised into a single submission record. The unglamorous work here is identity. Which account is this, is this email an update to an existing submission or a new one, and has the same risk already arrived through another broker. Duplicate detection at intake saves more underwriter time than most extraction features do.
Extraction then converts the documents into what is usefully called a structured submission object: named insured, class code, limits, deductibles, locations, exposure units, loss dates and claim history, property characteristics, and completeness indicators. OCR gets the characters. The harder work is interpretation, reading a loss run whose layout differs by TPA, or reconciling a statement of values whose totals disagree with the location schedule.
A validation layer sits between extraction and everything downstream. It checks completeness against what the line of business requires, flags internal inconsistencies, and compares the extracted record against prior-year data and external enrichment. A submission that fails validation generates a specific missing-items list, and that list becomes the draft of the broker follow-up rather than another task on the underwriter’s desk.
Triage and appetite are ranking problems
Triage is the first economic gate. The system classifies each submission as complete, incomplete, duplicate, out-of-appetite, or needing human review, and routes it accordingly. The economics are blunt: a clean early decline of a non-target risk costs almost nothing, while every unsorted submission that reaches an underwriter consumes attention regardless of merit. Carriers rarely lose money on the risks they decline quickly; they lose it on the good accounts that go stale in a queue behind bad ones.
Appetite matching turns the same extracted attributes into an ordering of the queue. The submission is scored against class, geography, size, loss profile and line, using underwriting guidelines, historical bind data and portfolio constraints, and the output is an appetite or fit score that decides what gets worked first. The portfolio-aware part matters. A risk can fit the guidelines and still be unwanted because the book is already concentrated in that class and region, and a score that ignores the current portfolio will keep feeding the concentration.
The workbench is where the value shows up
Extraction and scoring only pay off if the underwriter sees them in one place. The workbench pattern puts the structured submission next to its source documents, with a risk summary, external enrichment, the missing-items list, suggested questions for the broker, and a draft rationale for the quote, decline or hold decision. The underwriter verifies the package against the documents instead of assembling it, and the drafted rationale becomes the record of the decision rather than an email reconstructed later.
AIG’s reported deployment is a useful reference for the operating model: an AI-assisted layer that reads submissions, asks clarifying questions and surfaces narrative context for senior underwriters, rolling out across commercial lines, including its E&S and reinsurance businesses, through 2026. The pattern is decision support for experienced underwriters, not a parallel automated underwriter.
On speed, the directional claims are aggressive: submission-to-quote compressed from days to minutes for routine business, with one market summary citing standard-policy decision times falling from 3–5 days to about 12.4 minutes and vendor-reported extraction accuracy in the 92%–97% range. Treat the specific figures as vendor-flavoured. The consistent finding for complex commercial placements is faster triage and better prioritisation, not instant quotes.
Running intake AI in production
The engineering reality is mostly measurement and provenance. Extraction accuracy is a per-field, per-document-type number, and the accuracy that matters is on bind-critical fields: limits, deductibles, class codes, loss history. Those need confidence thresholds below which a person verifies the value before it feeds rating. Every field in the submission object should carry a pointer to the source page that produced it, because the first time a disputed quote traces back to a misread loss run, provenance stops being optional.
Expect drift. Broker document formats change, TPAs redesign loss runs, and an extraction model that was accurate at launch degrades quietly. A small labelled evaluation set per document type, re-run on a schedule, catches the decay before the queue does. Watch the appetite model with equal suspicion: trained purely on historical bind data it will reproduce yesterday’s book, so the learned score needs to sit beside explicit, editable appetite rules that express current strategy, with the two kept visibly separate.
The vendor market splits into three layers: ingestion and document processing, appetite and decisioning, and workbench orchestration. Most carriers will buy at least one layer rather than build it. Whatever the mix, the operational tests are the same: fields traceable to sources, queues that actually clear, broker follow-ups that go out the same day, and an audit trail that explains every routing decision. Intake is the cheapest place in underwriting to add AI, and the most expensive place to add silent errors.
FAQ
Where should a carrier start if underwriters are drowning in submissions?
Start with triage, not extraction depth. Classifying submissions as complete, incomplete, duplicate or out-of-appetite requires far fewer extracted fields than full risk digitisation, and it removes the largest block of wasted underwriter time immediately. Once the queue is ordered and non-target risks are declined early, deeper extraction and appetite scoring have a clean pipeline to improve.
How accurate does extraction need to be before underwriters trust it?
Accuracy is a per-field, per-document-type question, so a single headline number tells you little. Bind-critical fields such as limits, deductibles, class codes and loss history need a confidence threshold below which a person verifies the value before it feeds rating. Underwriters trust the system when every field links back to the source page it came from and the low-confidence ones arrive visibly flagged rather than silently filled.
Does appetite scoring risk hardcoding yesterday's portfolio?
Yes, if the score is trained purely on historical bind data it will reproduce the book you already have, including its concentrations. The fix is to keep explicit, editable appetite rules that express current strategy alongside the learned score, and to show both to the underwriter separately. When strategy shifts, the rules move immediately and the model catches up on retraining.