MGA Submission AI Tests Exposure Data Lineage, Not Just Speed

Vertafore's Velocity AI Submission Processing Agent reads unstructured emails, PDFs and spreadsheets and converts them into structured underwriting data at roughly 87% field-extraction accuracy (Vertafore, July 8, 2026).

The consequence that matters is not the speed gain. It is a provenance problem: when a machine assembles the exposure record, the actuary inherits responsibility for data nobody watched being entered.

Key Takeaways

• 87% accuracy leaves a 13% tail of misread, misclassified or dropped fields, and the completeness check the agent runs is a data-presence test rather than a data-accuracy test.
• A $2.4 million payroll misread as $2.04 million understates workers' compensation exposure by roughly 15%, and that flows straight to bound premium absent a downstream reconciliation.
• MGAs wrote more than $125 billion of U.S. premium in 2025, about 12.5% of the domestic P&C market, with Gallagher Re forecasting another 10% of growth in 2026.
• A block written over 18 months may carry several extraction model versions with different error profiles, which is unexplained variance in loss ratio by cohort if nobody records the version.
• Early adopters report a 30% reduction in processing time, the figure MGAs will cite to brokers. The 13% is the figure a pricing actuary tracks.

Where the 13% Lands

Submission intake sits upstream of nearly all MGA volume. Rating variables, from square footage and payroll to fleet counts and prior-loss history, arrive as broker emails, scanned loss runs and free-form spreadsheets long before reaching a policy administration system. MGAs wrote more than $125 billion in 2025, about 12.5% of the domestic property and casualty market, with Gallagher Re forecasting another 10% of growth.

Vertafore's general manager for MGA and Wholesale, Emily McGinn, framed the target: "Every minute underwriters spend gathering and organizing information is a minute they're not evaluating risk." XPT Specialty's Dan Rieden said the tool lets them "process submissions in any format, without additional configuration or dependencies."

Both describe speed. The actuarial reading is different. An 87% accuracy rate means roughly 13% of extracted fields carry some error before human review, and not all fields carry equal rating weight. A transposed digit in a mailing address rarely moves premium. A misread payroll figure, vehicle count or square-footage entry moves a variable that multiplies into the base rate.

Workers' compensation shows it cleanly. A $2.4 million payroll misread as $2.04 million from a scanned loss run understates exposure by roughly 15%, and that understatement reaches bound premium unless reconciliation catches it. Property total insured value carries the same OCR risk, and commercial auto fleet counts extracted from a broker spreadsheet are exposed to row-parsing errors that drop or duplicate a vehicle.

The agent evaluates completeness and flags missing fields, so the visible failure is a submission kicked back for more information. The invisible failure is a submission that reads as complete because every field holds a value, just not the right one. Completeness is a presence test, not an accuracy test, and that gap is the whole exposure in the 13%.

The Chain of Custody Splits in Two

Rate filings rest on an implicit chain: exposure data flows from application to bound policy to premium, and the supporting actuary can describe how it was captured. Machine-assembled intake splits that chain into two segments needing separate documentation, what the agent extracted from the raw submission, and what a human accepted, corrected or overrode before binding.

Losing the distinction loses the answer to a basic question. Was this record's payroll figure the broker's number, or the agent's read of the broker's number?

Reconciliation control	What it tests	Practical trigger
Field-level sampling	Extracted value vs. source document, on a sample of bound policies	Below-benchmark confidence score or high-rating-weight field
Confidence-threshold routing	Whether low-confidence extractions received mandatory human review before binding	Confidence score below insurer-set floor
Cohort variance testing	Loss ratio drift by underwriting-year cohort around the automation rollout date	Deployment date of a new extraction model version
Completeness-flag audit	Whether flagged submissions are resubmitted, declined, or silently dropped at a different rate than unflagged ones	Quarterly mix-shift review

Three elements make that lineage defensible to a filing reviewer or an exam: the extraction agent's confidence score at time of use, a record of what the agent originally produced against what a human changed, and a version identifier tying each bound policy to the platform version that processed it. Vendors update extraction models on a rolling basis, which makes the third element non-optional. A block written over 18 months may carry exposure data assembled by several model versions with different error profiles, and averaging across them unknowingly is a real source of unexplained loss ratio variance by cohort.

Confidence-threshold routing is the lever an MGA controls directly: a minimum score below which a field routes to human review, rather than trusting one blended accuracy statistic across every field and document type. That threshold should not be uniform. A high-leverage field, payroll, total insured value, fleet count, warrants a stricter floor than a mailing address even when the model's confidence in both is identical. Extraction accuracy and rating materiality are separate axes, and a protocol built on the first under-reviews the fields that move the loss ratio.

The Flag That Filters the Book Before Underwriting Sees It

The incomplete-submission flag is marketed as broker-friction relief: the system generates the follow-up request rather than an underwriter chasing it. The actuarial effect runs through selection instead of accuracy.

A submission that is genuinely hard to classify, a mixed-use property, an unusual payroll structure, an atypical fleet mix, is disproportionately likely to trip a completeness flag, because its data does not match the extraction model's trained expectation of a clean submission. If that friction discourages resubmission, or if flagged files sit behind clean ones that move faster to quote, the book shifts toward risks the model reads easily rather than risks that are well priced.

That is an adverse-selection channel distinct from anything the rating algorithm introduces, and it operates entirely upstream of underwriting judgment. A rate filing's exposure base assumes the mix reaching the rating engine represents what the MGA intends to write. If intake is filtering for extraction-friendly submissions, realized mix drifts from that assumption without tripping any model-validation check, because the rating model never sees the submissions that stopped short of it. A submission-automation rollout date deserves the same treatment as a change in underwriting guidelines: a structural break to test for, not assume away.

The governance frame is built to miss this. The NAIC Model Bulletin, adopted December 2023 and in force in roughly 24 states, holds insurers responsible for outcomes even where a vendor's model produced the data, and the Third-Party Data and Models Task Force formed in 2024 is still building the framework for vendor-supplied pipelines.

But an intake agent produces inputs rather than decisions. A program organized around validating scoring models for bias will not capture it if the tool is filed under operations, even though its 13% tail moves the same variables the pricing model then acts on. Colorado's model-compliance reporting regime previews where that lands.

Manual intake was never error-free either. The difference is that a transcription error was in principle visible to the person making it, while an extraction error surfaces only through a reconciliation program someone has to ask for.