MGA Submission AI Tests Exposure Data Lineage, Not Just Speed

Vertafore's new 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 overlooked consequence is not the speed gain MGAs are marketing, but a new provenance problem: when a machine assembles the exposure record, the actuary inherits responsibility for data it never watched being entered.

MGAs wrote more than $125 billion of U.S. premium in 2025, about 12.5% of the domestic property and casualty market, and Gallagher Re forecasts another 10% of growth in 2026 (Gallagher Re, cited in The Insurer, May 2026). Submission intake sits upstream of nearly all of that volume: an underwriter's rating variables, from square footage and payroll to fleet counts and prior-loss history, typically arrive as broker emails, scanned loss runs, and free-form spreadsheets long before they reach a policy administration system. Vertafore's agent, built on its Velocity AI Platform and embedded across the vendor's AgencyOne, MGA, and Sircon products, targets exactly that gap. Emily McGinn, Vertafore's general manager of MGA and Wholesale, framed the pain point directly: "Every minute underwriters spend gathering and organizing information is a minute they're not evaluating risk and moving submissions forward" (Vertafore, July 8, 2026). Dan Rieden, EVP of workers' compensation at XPT Specialty, an early user, added that the tool "enable[s] us to process submissions in any format, without additional configuration or dependencies" (Vertafore, July 8, 2026), a description of exactly the rigid-template problem submission processing has always had.

That framing is a speed story. The actuarial story is different: once exposure data is machine-assembled rather than manually keyed, an actuary can no longer assume a human underwriter caught the transcription error, the missing endorsement, or the misread loss run before the record reached the rating engine. The extraction layer becomes a new, largely invisible link in the chain between what a broker submitted and what a rate filing certifies as the exposure basis for premium.

The 13% Tail: Where Extraction Errors Land in the Rating Variables

An 87% field-extraction accuracy rate means roughly 13% of extracted data fields carry some error, whether a misread number, a misclassified field, or a dropped value, before any human review catches it (Vertafore, July 2026; TechEdgeAI, July 2026). Early adopters reported a 30% reduction in submission processing time alongside that accuracy figure (TechEdgeAI, July 2026), which is the number MGAs will cite to brokers. The number a pricing actuary should track is the 13%, and specifically where in the rating structure those errors concentrate.

Not all fields carry equal rating weight. A transposed digit in a mailing address rarely moves a premium calculation; a misread payroll figure, vehicle count, or square-footage entry can shift a rating variable that multiplies directly into the base rate. Workers' compensation payroll-based rating is a clean illustration: a $2.4 million payroll figure misread as $2.04 million from a scanned loss run understates exposure by roughly 15%, and that understatement flows straight through to bound premium unless a downstream reconciliation step catches it. Property total insured value carries the same risk from OCR misreads of loss-run PDFs, and commercial auto fleet counts extracted from a broker's spreadsheet are exposed to simple row-parsing errors that silently drop or duplicate a vehicle. Because the extraction agent evaluates completeness and flags missing fields (Vertafore, July 8, 2026), the visible failure mode is a submission kicked back for more information; the invisible failure mode is a submission that looks complete because every field has a value, just not the correct one. The completeness check is a data-presence test, not a data-accuracy test, and that distinction is the entire actuarial risk in the 13% tail.

Data Lineage: What a Rate Filing Needs to Show When a Machine Built the Exposure Record

Rate filings have always rested on an implicit chain of custody: exposure data flows from application to bound policy to premium, and an actuary supporting a filing can generally describe, at least in broad terms, how that data was captured. Machine-assembled intake breaks that chain into two segments that now need to be documented separately: what the AI agent extracted from the raw submission, and what a human underwriter accepted, corrected, or overrode before binding. Losing the distinction between those two segments means losing the ability to answer a basic reserving or ratemaking question: was this exposure record's payroll figure the broker's number, or the agent's read of the broker's number?

The NAIC's Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted December 2023 and now in force in roughly 24 states (NAIC, 2026), already reaches this problem indirectly. The bulletin instructs insurers to maintain written standards for third-party AI systems and to remain responsible for outcomes even when a vendor's model, not the insurer's own, produced the data or decision (NAIC, Insurance Topics: Artificial Intelligence). A submission-intake agent is a genuine edge case for that framework, because it does not score risk or set a rate; it produces the exposure inputs that a separate rating model or underwriter judgment will later act on. The 2024-formed Third-Party Data and Models (H) Task Force is still building a regulatory framework for exactly this category of vendor-supplied data pipeline (NAIC, 2026), which means MGAs adopting intake automation today are operating ahead of a settled documentation standard, not behind one.

In practice, a defensible lineage record for machine-assembled exposure data needs three elements a rate filing or a regulatory exam can request: the extraction agent's confidence score or accuracy estimate at time of use, a record of what the agent originally extracted versus what a human underwriter subsequently changed, and a version identifier tying a given bound policy back to the specific model or platform version that processed its submission. Vendors updating extraction models on a rolling basis, which is the norm for any actively developed AI product, make that third element non-optional: an actuary reviewing a block of business written over 18 months may be looking at exposure data assembled by several different model versions with different error profiles, and averaging across them without knowing that is a subtle but real source of unexplained variance in loss ratio by cohort.

How Completeness Flags Quietly Reshape the Book

The agent's incomplete-submission flagging is marketed as a broker-friction reducer: instead of an underwriter manually chasing missing information, the system generates the follow-up request automatically (Vertafore, July 8, 2026). The actuarial effect runs through selection, not accuracy. A submission that is genuinely hard to classify, a mixed-use property, a business with an unusual payroll structure, a fleet with an atypical vehicle mix, is disproportionately likely to trip a completeness flag simply because its data does not fit the extraction model's trained expectations of what a clean submission looks like. If that friction discourages brokers from resubmitting, or if MGAs quietly deprioritize flagged submissions in favor of clean ones that move faster to quote, the book being written shifts toward risks that are easy for the model to read, not necessarily risks that are well-priced or well-understood.

That is an adverse-selection channel distinct from anything a rating algorithm itself introduces, and it operates upstream of underwriting judgment entirely. A rate filing's exposure base assumes the mix of business reaching the rating engine is representative of the risks the MGA intends to write; if the intake layer is systematically filtering for extraction-friendly submissions, the realized mix can drift from that assumption without triggering any of the model-validation checks built around the rating model itself, because the rating model never sees the submissions that got quietly discouraged before reaching it. Reserving actuaries reviewing loss experience by underwriting-year cohort should treat a submission-automation rollout date the same way they would treat a change in underwriting guidelines: as a structural break worth testing for in the data, not assuming away.

Reconciliation: Matching Extracted Data Against the Policy of Record

The control that closes most of this gap is procedural rather than technical: systematic reconciliation between the AI-structured intake data and the bound-policy data of record, sampled at a rate proportional to the extraction model's measured error tail. An 87% accuracy figure is itself a starting benchmark, not a static guarantee; extraction accuracy varies by document type, submission format, and line of business, and a vendor's blended average across early users (Vertafore, July 2026) will understate the error rate on the specific document types and classes an individual MGA underwrites. A specialty program with heavy reliance on scanned loss runs and non-standard broker templates should expect a materially different, likely worse, error profile than the blended figure suggests.

Confidence-threshold routing is the most direct lever an MGA actually controls: setting a minimum confidence score below which an extracted field routes automatically to human review, rather than trusting a single blended accuracy statistic across every field and document type. That threshold should not be static. A field with high rating leverage, payroll, total insured value, fleet count, merits a stricter threshold than a field with low leverage, like a mailing address or contact name, even if the underlying model's confidence score is identical for both. Extraction accuracy and rating materiality are two separate axes, and a review protocol built around only the first axis will systematically under-review the fields that matter most to the loss ratio.

Governance When the Data, Not Just the Model, Is AI-Produced

Most AI governance discussion in insurance, including the NAIC bulletin's own emphasis, has centered on models that score or price risk: a rating algorithm, a fraud model, a claims-triage tool. Submission-intake automation sits a step earlier in the pipeline and does not fit that frame cleanly, because it produces inputs rather than decisions. That distinction matters for how an insurer's AI Systems Program, the documented governance structure the NAIC bulletin requires (NAIC Model Bulletin, adopted December 2023), should treat it. A governance program built only around model validation, testing for bias and unfair discrimination in a scoring algorithm, will miss an intake tool entirely if that tool is filed under "operations" rather than "AI systems," even though its 13% error tail can move the same rating variables a validated pricing model then acts on.

The practical fix is to bring exposure-data extraction explicitly inside the same third-party AI oversight the bulletin already contemplates for vendor scoring models: written standards for acquisition and use, documented accuracy benchmarks specific to the insurer's own document mix rather than the vendor's blended figure, and an audit trail sufficient to answer a regulator's question about how a specific bound policy's exposure data was assembled. Colorado's model-compliance reporting regime, the first state to require insurers to document AI system use in this level of operational detail, is a preview of where that documentation standard is heading (see actuary.info's coverage of Colorado's first AI model compliance reports). An MGA that can produce a confidence-score distribution and a reconciliation-sample error rate for its intake agent is answering a question regulators are increasingly likely to ask before they ask it about the rating model downstream.

Practical Controls for the Actuary Reviewing the Book

An actuary evaluating a book of business written through an AI intake agent has a short list of concrete questions worth asking before treating the exposure data as reliable input to a rate filing or reserve review: what confidence-score distribution did the extraction model produce across this cohort, what sampling program validates extracted fields against source documents, what threshold routes low-confidence fields to mandatory human review, and does that threshold vary by rating leverage rather than applying uniformly. A fifth question follows directly from the completeness-flag mechanism: is there a way to measure whether flagged, hard-to-classify submissions are being resubmitted and bound at a different rate than clean ones, since that is the channel through which intake friction becomes portfolio selection rather than a data-quality problem alone.

None of those controls require rejecting the underlying technology. A 30% reduction in submission processing time (TechEdgeAI, July 2026) is a genuine operational gain for an MGA market growing 10% a year on top of $125 billion in existing premium (Gallagher Re, May 2026), and manual intake was never error-free either; a tired underwriter transcribing a scanned loss run at 11pm is not a more reliable data source than a well-monitored extraction model. The difference is that a manual process's errors were, at least in principle, visible to the person making them in real time. An AI intake agent's errors are invisible by default, and become visible only through a reconciliation program an actuary has to ask for.