Allstate's AI Is Closing Policies in Three States: Distribution Shift, Adverse Selection, and Rate Filing Questions

Allstate CEO Tom Wilson confirmed on the Q1 2026 earnings call that AI is closing policies in three states outside the exclusive agent channel. "The AI can also just sell directly. And we're live in the market doing that right now on a particular product. It's more of a learning, but it's doing it in 3 states. It's closing policies," Wilson told analysts. The same quarter produced an 82.0% property-liability combined ratio and $2.8 billion in adjusted net income.

Key Takeaways

• Three states, one product, framed as "more of a learning." Management is treating the pilot as data collection, and the underwriting experience it needs does not exist yet in any form.
• The agent share of new auto policies fell from 71% in 2020 to 38% in 2025. An AI that closes without an agent accelerates a channel shift already half a decade old.
• Exclusive agency auto commissions run 10 to 12 points of earned premium. That is the expense gain; whether the loss ratio gives it back is the open question.
• Channel-specific frequency needs about 18 months of policy year data and bodily injury severity takes 36 months or more, so the pricing answer arrives after the expansion decision.
• The filed rates were supported on an agent-placed book. Nothing in the current indication describes how an AI direct channel selects risk.

What Wilson Actually Disclosed

The qualifier carries the information. "More of a learning" says management is not treating this as a near-term national rollout, and the reason is structural rather than modest: the company cannot know whether AI-closed risks are priced correctly until accident year 2026 develops. Wilson named neither the states nor the product line.

The capability sits inside ALLIE, Allstate's Large Language Intelligent Ecosystem, an agentic platform supporting customer engagement, claims handling, and distribution across licensed sales representatives and call centers. ALLIE already drafts more than 50,000 claims messages daily.

The pilot extends that infrastructure from communication support into the transaction itself, and the change of use is what matters. A claims message is a draft an adjuster reviews and amends, which keeps it inside the human-supervised domain. A policy contract closed by the AI is binding at the moment of issuance. Adverse action, rate adequacy, and selection all attach to the second case and not the first.

The Selection the Agent Channel Used to Do

Point-of-sale underwriting has always carried informal observation that never appears in a rate filing. An agent taking a new auto application by phone asks qualifying questions beyond the rating schedule and reads the conversation. An agent at a home for a homeowners renewal notes property condition, the vehicles in the driveway, visible maintenance. None of it is documented, and all of it filters the book.

The exposure is not that ALLIE cannot price. The filed algorithm applies the same factors to an AI-closed policy that an agent would apply. The exposure is that customers who self-select into a frictionless digital channel differ systematically from customers who seek out an agent, and the difference is not actuarially neutral. Shoppers rated up or non-renewed elsewhere can reapply quickly without a conversation. Younger drivers with thin histories and thin credit files favor the same channel for convenience.

There is also a discretionary layer that disappears. An agent can route a risk to a non-standard market or decline one the guidelines technically allow. An AI executing a direct close applies the schedule. Those risks will be rated correctly per the filing and will still develop worse than the agent-placed cohort at the same rating level, because the screening that kept them out was never in the algorithm.

The expense side runs the other way. Exclusive agency auto commissions typically run 10 to 12 points of earned premium, and an AI close removes them on that cohort. Whether the net combined ratio effect is favorable depends on which force is larger, and the pilot exists to find out. It cannot report quickly: frequency needs roughly 18 months of policy year data, and bodily injury severity does not stabilize inside 36 months.

Progressive is the benchmark for how long that calibration actually takes. It has run a direct channel for more than two decades, reached 37.4 million personal lines policies in force in Q1 2026 with direct auto up 14% year over year, and holds Snapshot telematics data covering more than 100 billion driving miles and $2.2 billion of customer discounts since 2009. Better models compress the customer interaction. They do not compress the accident year.

Nobody Filed for This Channel

A personal auto rate filing is supported by loss and expense experience from the book as it stood when the filing was prepared. If that book was entirely agent-placed, the indication is calibrated to agent-placed selection, and the approved rates carry that assumption whether or not anyone wrote it down. Channel-mix assumptions are rarely broken out explicitly, which is where the gap opens when the distribution model changes.

Timing makes it worse rather than better. Rate revisions in most states run on a 12-to-24-month cadence, so AI-closed policies can reach their second year before the first channel-specific indication is available to file. A blended indication averaging both books into one rate level compresses the signal from each, and by the time the divergence is large enough to correct, the blend has been absorbing it retroactively for several cycles.

The second exposure is documentation rather than rate level. If AI closes policies, AI also declines applicants and assigns tiers. Pennsylvania's May 2026 settlement with GEICO set the current floor: an AI-enabled underwriting review tool flagged a West Philadelphia customer's new auto policy, the policy was cancelled without adequate notice, and the customer drove uninsured without knowing coverage had lapsed. The agreement required adherence to state Insurance Department AI governance guidance, a formal bias detection program, and documentation of the role AI played in any consumer-impacting decision.

For a direct-close pilot the same obligation attaches at issuance rather than renewal, and the record has to be contemporaneous. Reconstructing decision logic after the fact in a market conduct examination is materially more expensive than logging it in real time. The regulatory frame is not uniform either, which is likely part of how the three states were chosen: Colorado's Senate Bill 21-169 mandates documentation of model fairness testing, California has required pre-approval of algorithmic rating systems since 2023, and other states have adopted the NAIC bulletin framework, which asks for governance documentation without pre-approval.