A submission is complete when an underwriter can quote it without writing back. That bar is different for every line of business and often for every market, which is why so many submissions bounce: a field missing from the application, a statement of values that disagrees with it, loss runs that are no longer currently valued, operations that only the account executive understands. Every round trip costs time to quote and goodwill with the market. The fix is not a longer checklist. It is a completeness score per line and per market, checked before the submission leaves the building by an AI agent that extracts, cross-checks and flags, while an account-team member reviews and approves. Below are the documents underwriters actually need, why submissions come back, how the score works, the pre-flight checks we automate, the reference architecture and what to measure.
The bar: quotable without a follow-up question
Every broker knows the email. A carefully assembled submission goes to six markets, and within two days three of them write back: can you confirm the roof year on location 4, the loss runs only go to last March, and what share of revenue is the installation work? None of the questions is unreasonable. Each one restarts a clock, and each one tells the underwriter something about how the account was prepared.
We use a simple definition. A submission is complete when the underwriter it is sent to can quote it without asking anything. That sounds obvious, but it has two consequences that most submission processes ignore. First, completeness is not a property of the documents alone; it depends on the line of business and on what that market needs to price it. Second, it can be checked before sending, because nearly every question an underwriter asks is predictable from the documents in the file.
What underwriters need to quote
The core of a commercial P&C submission is the same across most markets. What varies is the depth each line needs and the supplements that go with it.
| Component | What the underwriter uses it for | Where it usually goes wrong |
|---|---|---|
| Application (ACORD and carrier forms) | Named insured, operations, revenue, payroll, limits and coverages requested | Blank fields, outdated revenue, operations described in one line |
| Exposure schedule (SOV, vehicle or driver list, payroll by class) | Pricing the exposure: TIV, construction, occupancy, protection, vehicles, classes | Totals that disagree with the application; missing COPE details |
| Loss runs | Loss history and development by year | Not currently valued, missing years, carrier runs not matched to policy years |
| Prior terms and expiring program | Expiring premium, limits, retentions, layers and carriers | Expiring structure missing or out of date for layered programs |
| Underwriting narrative | Why this risk, what changed, risk management, the story behind the losses | Missing, generic, or silent about the change the underwriter will spot |
| Line-specific supplements | Cyber controls, contractor work mix, products and completed operations, fleet safety, professional services detail | Not attached, or answered for the wrong entity |
Each line adds its own must-haves. A property schedule without year built and roof details for older locations gets questions. A contractor’s general liability submission without a work-mix breakdown and subcontractor practices gets questions. A commercial auto submission without a current driver list gets questions. These requirements are knowable, and that is the point: if they are knowable, they can be encoded.
Why submissions bounce
When we review a broker’s recent follow-up questions with them, they almost always fall into a handful of patterns. None of them is about the broker’s judgment. They are about the documents.
- Missing fields. A required field on the application is blank, or a supplement was never attached.
- Contradictions between documents. The SOV total differs from the TIV on the application. The application says three locations; the schedule lists four. Revenue on the narrative does not match revenue on the form.
- Stale documents. Loss runs valued many months ago, an SOV carried over from last year without the new location, a driver list from before the fleet grew.
- Unclear operations. The account executive knows the client well, but the submission says “manufacturing” where the underwriter needs to know what is made, for whom and with what share of revenue.
- The change nobody explained. An acquisition, a new state, a large loss, a jump in payroll. The underwriter will notice it either way. A submission that explains it gets priced; one that does not gets a question.
Every one of these is detectable before sending. A contradiction between two documents is a comparison. A stale loss run is a date. A missing supplement is a lookup against the line. An unexplained change is a comparison with last year’s submission.
What the back-and-forth costs
The cost of an incomplete submission is not the minutes it takes to answer one email. It compounds:
- Time to quote. Each round trip adds days. Near an effective date, a late quote narrows the choice of markets for the client.
- Market fatigue. Underwriters triage. A broker whose submissions are reliably complete gets read first; one whose submissions need work gets read later.
- Account-team capacity. Time spent chasing missing details for one market is time not spent on the client conversation or the next renewal.
- Inconsistent answers. When different markets ask the same question separately, the answers can drift. That creates confusion at binding and, in the worst case, an E&O exposure.
We do not quote a percentage for any of this, because it varies too much by book, line and market. We measure it with each broker from their own data, which is the first thing the assessment produces (see what to measure).
A completeness score, by line and by market
A checklist tells you what to collect. A score tells you whether a specific submission is ready for a specific market. We build it from rules in three groups, matching the three layers above:
| Rule group | Example rules | Source of the rule |
|---|---|---|
| Core | Named insured and FEIN present; revenue and payroll within the policy year; loss runs cover the required years and are recently valued | Common to every market |
| Line | Property: year built, construction and protection for each location; GL: work-mix percentages that add up to 100; Auto: driver list with dates of birth and license states | Your placement team’s standards per line |
| Market | This market’s supplemental form attached; this market’s known follow-up questions answered | Market requirements, plus questions this market has asked before |
| Consistency | TIV on the application equals the SOV total; location count matches; revenue agrees across documents | Cross-document checks |
| Change | Material changes from last year’s submission are explained in the narrative | Comparison with the prior submission |
Each rule is weighted. A missing loss run year is more serious than a missing fax number. The score is shown per market, with the failing rules listed in plain language: “Location 4 has no roof year; Market B’s cyber supplement is missing; payroll on the application is 12% higher than the class schedule.” The account team fixes what matters, overrides what does not, and sends.
The most valuable rules are the ones that come from history. Every follow-up question a market has sent is a rule that the next submission to that market should pass. Over time the market layer becomes a record of what each underwriter actually needs.
Pre-flight: what an AI agent checks before sending
This is where AI does real work. The documents arrive as PDFs, spreadsheets and emails, often scanned, rarely in the same layout twice. Before any rule can run, the data has to be extracted, classified and validated. That is the job of document extraction; the agent then runs the checks and drafts what is missing.
The checks, concretely
| Check | What the agent does | What the account team sees |
|---|---|---|
| Required fields | Compares extracted fields with the line’s rule set | “Year built missing for 2 of 9 locations” |
| Cross-document consistency | Reconciles TIV, location counts, revenue and payroll across application, SOV and supplements | “Application TIV differs from SOV total; location 7 is on the SOV only” |
| Currency | Reads valuation dates on loss runs and the effective dates on schedules | “Carrier C loss runs valued more than 90 days before the effective date” (threshold set by you) |
| Supplements | Matches attached documents against each target market’s requirements | “Market B cyber supplement not attached” |
| Changes | Compares with last year’s submission and flags material differences | “Revenue up materially; narrative does not mention the acquisition” |
| Drafting | Drafts the request to the client for what is missing, and a narrative paragraph on the change | A draft to edit, never an email sent automatically |
The agent never sends anything to a market or a client. It reads, checks and drafts. Sending is an account-team decision, and the record shows who approved what.
Reference architecture
The flow sits between where submissions arrive today and where they leave for the market. It does not replace the agency management system; it reads from it and writes back to it.
The components
- Extraction and validation: InsightXtract. Agentic extraction for P&C documents: applications and ACORD forms, statements of values, loss runs and broker emails. It classifies each document, extracts the fields, validates them and routes low-confidence fields to human review instead of passing them on as facts. Accuracy is measured against gold datasets, and corrections feed back to improve the agents.
- Completeness rules. The rule sets per line and per market, owned and edited by your placement leaders, not by engineers. Market rules grow from the questions each market has asked.
- Insured intelligence: InsightPulse. Signals about the insured from outside the file, such as ownership changes or company events, that a market may already know about. If something material changed, the narrative should say so.
- The review step. A screen in the account team’s workflow showing the score per market, the issues and the drafts, with approve and override actions recorded.
- Output in an ACORD-aligned format. Structured data travels with the documents, so a market can load it rather than rekey it.
How it reads on the carrier side
We build for carriers too, which shapes how we design this. On the carrier side, InsightUW reads broker emails and ACORD forms into a submission on arrival and runs appetite, clearance and authority checks before the underwriter opens it. A complete, consistent submission moves through those checks cleanly. An incomplete one stops at the first missing field and waits for a question to be written.
That is the practical argument for a shared standard. When the broker side produces structured, validated data aligned to ACORD, and the carrier side ingests it, both sides stop rekeying the same facts, and the first conversation is about the risk instead of the paperwork.
What to measure
Start by measuring the baseline from the last few months of submissions, then the same numbers after the pilot. We measure:
| Measure | Why it matters |
|---|---|
| Follow-up questions per submission, by market and line | The direct measure of completeness as markets see it |
| Time from submission sent to first quote | The client-facing outcome |
| Most common follow-up questions | The backlog of rules to add next |
| Share of submissions passing pre-flight first time | How well intake and client requests are working |
| Overrides and their reasons | Where rules are too strict, or where the team is taking a risk knowingly |
| Extraction fields sent to review | Where document quality or extraction needs work |
The failure modes we plan for
1. One checklist for every line. A generic list either asks for things a line does not need, so people ignore it, or misses what the line does need, so questions continue. Rules have to be per line and per market.
2. Trusting extraction blindly. A misread TIV that passes a consistency check is worse than a blank. Low-confidence fields must go to a person before rules run on them.
3. Automating the send. The moment an agent emails a market without a person approving, the broker has handed its relationship and its E&O exposure to software. We do not build it that way.
4. Rules owned by IT. If placement leaders cannot edit the rules, the rules fall behind what markets ask. Ownership has to sit with the people who hear the questions.
5. Forgetting the market layer. The biggest gains come from encoding what each market has asked before. Teams that skip it get a cleaner core and the same follow-ups.
Readiness checklist
| Before a pilot | Ready when |
|---|---|
| Scope | One practice or line, and the markets it places most with, are chosen |
| Baseline | Follow-up questions and time to quote are measured for recent submissions |
| Rules | Core and line rules are agreed with placement leaders; the top market questions are listed |
| Documents | Sample applications, SOVs, loss runs and supplements are available for extraction testing |
| Workflow | The review step is placed where the account team already works |
| Controls | Who approves sending, and how overrides are recorded, is written down |
The bottom line
Brokers do not get re-asked because they lack judgment. They get re-asked because documents disagree, go stale or leave out what a particular market needs, and nobody can cross-check every page of every submission by hand before the deadline. That cross-checking is exactly the work an AI agent does well: extracting, comparing, flagging and drafting. The judgment stays with the account team, who see a score per market and decide what to send.
Done well, the submission that leaves the building is the one the underwriter can quote. That is better for the client, better for the market relationship and a better use of the account team’s day.