When AI Sorts Commercial Submissions, Pricing Models Inherit the Selection Bias

AIG's Underwriter Companion and Sixfold's newly launched AI underwriting agent both report commercial lines capacity gains of 50% or more, with hit ratios rising 15% or better once AI pre-screens submissions by carrier appetite. Those metrics are real. They also both improve when the population feeding the ratemaking model is being narrowed, and neither one moves when that is what is happening.

Key Takeaways

Hit ratios up 15% or more across six carriers writing a combined $270 billion of gross written premium, with processing times improving 50% to 97% and premium per underwriter up to 30%.
Appetite-screened submissions are censored observations. Risks declined before underwriter review never generate outcome data, so they leave the data-generating process without appearing anywhere in the loss triangle.
Hit rate conflates two mechanisms. Quoting accurately and quoting only risks the model expects to convert both raise it, and only the first says the pricing structure improved.
Generative AI in commercial underwriting is projected at 70% within three years against 14% today, on an Accenture survey of 430 senior underwriting executives across eleven countries.
E&S direct written premium reached $98.2 billion in 2024 against $86.6 billion in 2023, a seventh consecutive year of double-digit growth, and it is where appetite-screened risk goes.

What the Deployments Actually Report

Sixfold launched its AI Underwriter in June 2026 as a configurable agent that processes commercial submissions, learns each carrier's appetite from historical underwriting decisions, and can take standard accounts from intake to quote-ready and bind-ready without review of the individual submission. Across six early customers writing $270 billion of gross written premium, processing times improved 50% to 97%, hit ratios rose 15% or more, and premium per underwriter climbed up to 30%. Each deployment runs a walled model; no training data crosses carrier boundaries.

AIG's system is older and wider. The Underwriter Companion ingests submissions in any format, normalizes them against AIG's exposure data model, surfaces the signals a senior underwriter would flag and drafts follow-up questions. Form-based processing that took four to six hours on a complex submission now takes minutes, and AIG reports roughly 50% less underwriter time on ingestion and triage. It is live across Lexington, Glatfelter, AIG Re and core commercial property and casualty, and in March 2026 AIG and McGill and Partners announced a Palantir-backed capacity arrangement covering up to $1.6 billion of specialty premium.

The market is moving the same way. Accenture projects generative AI in commercial underwriting reaching 70% within three years against 14% today, on 430 senior executives across eleven countries. Celent's Q2 2026 survey puts 22% of insurers planning agentic systems in production by year-end, with document analysis live at 29% of respondents, case analysis at 22% and submission ingestion at 20%. Intake is the first layer to move at scale, and it is the layer where the pricing problem starts.

The Model Is Still Trained on a Pool That No Longer Arrives

Every commercial lines GLM rests on an assumption that has rarely been stated because it has rarely been testable: that submissions received during the experience period approximate the available market, filtered by distribution relationships and by human underwriters applying guidelines. Human triage was always informal, individual and slow enough that a genuine range of risk quality reached the quoted population. That spread is what the model learned from.

Appetite scoring removes it at the front. Submissions are scored against the carrier's historical book on class fit, geography, prior loss signals, account size and broker tier, and those below threshold are deprioritized or declined before an underwriter engages. Those risks never produce experience data. They are censored observations, exiting the data-generating process before any outcome is recorded.

That is a model validity problem rather than a data quality one, and it leaves no fingerprint. The frequency and severity estimates stay calibrated to a pool the carrier no longer sees, and appetite pre-screening produces no anomaly in the loss triangle, the premium development pattern or any standard data diagnostic a ratemaking actuary runs.

Hit rate is where it could be caught, and hit rate is being read the wrong way. Sixfold's 15% improvement is genuine and reflects two mechanisms at once: pricing precise enough that brokers accept terms, and selection that quotes only risks the model expects to convert. Both raise the number.

AI Intake Configuration	Effect on Hit Rate	Effect on Ratemaking Data	Actuarial Diagnostic
No AI screening; human triage only	Baseline; reflects historical pricing accuracy	Broad submission population; GLM representative	Standard experience analysis valid
AI appetite scoring; human still reviews all	Modest improvement; partial selection	Mild narrowing; monitor quarterly	Track submission-to-quote ratio by class
AI sorts; STP on high-confidence tiers	Meaningful lift; selection and pricing blended	Moderate censoring; long-tail classes most exposed	Stratify hit rate by AI confidence tier; compare to pre-AI baseline
Full AI-first intake; most submissions auto-declined below threshold	High hit rate; majority is selection effect	Significant censoring; GLM no longer trained on market population	Rebuild frequency/severity models on post-AI cohorts; flag selection assumption in filing

Separating them takes stratification by submission source and AI confidence tier. If hit rates on high-confidence submissions run well above standard human-reviewed submissions, and above the pre-AI baseline on equivalent classes, the selection component is visible, and the residual over baseline is the pricing precision estimate. A carrier that cannot produce that split is treating a selection effect as evidence its rate structure is adequate on risks it is no longer being shown.

Each Cycle Trains on the Last Cycle's Screening

The narrowing compounds, and the mechanism is structural. Sixfold feeds each carrier's underwriting decisions back into that carrier's model as training data, building what it calls institutional memory existing workbench systems were not built to capture. AIG's Companion learns against AIG's own book. The appetite model governing the second renewal cycle is trained on data the first model already screened, so each cycle the training set describes a narrower slice of the commercial market.

The GLM is not recalibrated at each intake cycle. It updates at the next experience review, against losses from a progressively screened book. Three years of that produces a pricing model calibrated to a subpopulation the carrier drifted into, not the market class the rate filing nominally covers.

Tail length decides how long that runs unseen. Property and commercial auto correct fast: a poorly performing account surfaces in the loss run within eighteen months, while the selection pattern is still forming. A commercial general liability account may not develop recognizably until accident year two or three, by which point two further cohorts of similarly selected accounts are bound and the model has been recalibrated against them.

Through years one to three the book looks well selected and well priced. Calendar-year loss ratios improve, development is favorable, premium per underwriter rises, and nothing on the underwriting dashboard reads as a warning.

The external check sits outside the carrier's data. E&S direct written premium reached $98.2 billion in 2024 from $86.6 billion in 2023, up 13.4% and a seventh straight year of double-digit growth, and E&S is where declined admitted risk lands. The sector's run predates AI intake and has legitimate drivers, but a carrier whose admitted general liability book grew 8% in a market where E&S general liability grew 20% is looking at a divergence more consistent with appetite tightening than with superior penetration.

When AI Sorts Commercial Submissions, Pricing Models Inherit the Selection Bias

Key Takeaways

What the Deployments Actually Report

The Model Is Still Trained on a Pool That No Longer Arrives

Each Cycle Trains on the Last Cycle's Screening

Further Reading

Sources