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 (Allstate Q1 2026 Earnings Call, April 2026). That quarter, Allstate posted an 82.0% property-liability combined ratio and $2.8 billion in adjusted net income, its strongest underwriting result in years.

Wilson's Disclosure and What the Pilot Architecture Looks Like

The careful qualifier in Wilson's framing carries real information. "More of a learning" signals that Allstate's management treats the three-state pilot as a data-collection exercise rather than a near-term national rollout. That framing is not false modesty; it reflects the structural reality that AI direct sales at this scale generates underwriting experience that does not yet exist. The company cannot know whether AI-closed risks are priced correctly until accident year 2026 develops over the next two to three years, and Wilson's language acknowledges that uncertainty without disclosing the pilot states or the specific product line.

The AI capability sits inside ALLIE, Allstate's Large Language Intelligent Ecosystem, an agentic AI platform built to support customer engagement, claims handling, and distribution across licensed sales representatives and call center operations. ALLIE already drafts more than 50,000 claims messages daily. The direct-close pilot extends the same infrastructure from communication support into the point-of-sale transaction, which is a structurally different use: the AI output is no longer a draft for a human to review and send but a binding policy contract issued directly to the customer. That distinction is what makes the adverse action, rate filing, and selection questions non-trivial. A claims message that an adjuster reviews and amends stays in the human-supervised domain. A policy contract that an AI closes does not.

The Agent Channel's Contracting Share

Allstate's exclusive agent network has been contracting for years before the AI pilot arrived. The approximately 8,400 exclusive agents under contract as of the 2023 SEC filing (Allstate, 2023) drove 71% of Allstate new auto policies in 2020; that share had fallen to 38% by 2025 (Insurance Business, 2025) as Allstate Direct, CONNECT, and telesales channels absorbed an increasing proportion of new business. An AI that closes policies without a licensed agent relationship does not create the bypass problem from scratch. It accelerates a channel shift that has been developing for half a decade.

The expense ratio arithmetic is straightforward and cuts in two directions simultaneously. Exclusive agency auto commission rates typically run 10 to 12 points on earned premium. An AI direct close that eliminates those commissions on the policies it writes creates a structural expense advantage on that cohort. The loss ratio arithmetic runs the opposite direction if AI-closed risks adverse-select against the agent-placed book, and the net effect on the combined ratio depends entirely on which force is larger. That is the central actuarial unknown the three-state pilot is generating data to answer, and it will take years to answer it reliably.

Adverse Selection When AI Replaces Agent Judgment at Point of Sale

Personal auto underwriting at the point of sale has always involved informal observation that does not appear in the rate filing or the actuarial model. An exclusive agent processing a new auto application over the phone observes communication patterns, asks qualifying questions that go beyond the rating schedule, and applies judgment based on signals that are difficult to document but real in their effect on risk selection. An agent visiting a home for a homeowners renewal notes property condition, the vehicles in the driveway, and visible maintenance quality. That informal filter is absent in an AI direct channel where the interaction is text-guided and the AI applies rating factors to structured inputs without the conversational texture of an experienced agent.

The adverse selection risk is not that ALLIE cannot price. Allstate's filed rating algorithm applies the same rating factors to an AI-closed policy that a licensed agent would apply. The risk is that customers who self-select into an AI direct channel are systematically different from customers who seek out an agent, and the direction of that difference is not actuarially neutral. Sophisticated shoppers who have been rated up or non-renewed elsewhere may disproportionately favor a frictionless digital channel where they can shop quickly and reapply without the friction of a conversation. Younger customers with shorter driving histories and thinner credit files may favor the same channel for convenience reasons. Both groups are representable risks, but their loss development patterns differ from Allstate's historical agent-placed book in ways the current rate structure does not capture.

The pricing mechanism also changes when AI replaces the agent at the moment of issuance. A licensed agent has discretion at the margins: whether to send a risk to a non-standard market, whether to accept a risk the underwriting guidelines technically allow but the agent's experience flags as problematic. An AI executing a direct close applies the rate schedule without that discretionary layer. The actuarial implication is that risks that an experienced agent would have declined or non-standardized will appear in the AI direct book. They will be rated correctly per the filed schedule. They will also develop worse than the agent-placed cohort at the same rating level, because the informal screening that kept them out of the agent book does not operate in the direct channel.

Allstate will not have credible channel-specific loss frequency data until about 18 months into the pilot, and severity data meaningful enough to compare against agent-placed development patterns takes 36 months or more for bodily injury lines. Until that data exists, rate indications for the pilot states rest on the implicit assumption that AI-closed and agent-placed risks develop identically within rating cells. Documenting that assumption explicitly, and flagging it for review at each rate revision, is the minimum actuarial standard for the pilot states' rate filings.

Progressive's Two Decades of Direct Channel Calibration

The timeline question has a benchmark. Progressive has operated a direct-to-consumer auto channel for more than two decades and reached 37.4 million personal lines policies in force in Q1 2026, with direct auto growing 14% year over year (Progressive Q1 2026). Progressive's Snapshot telematics program has logged more than 100 billion driving miles and delivered more than $2.2 billion in customer discounts since its 2009 launch (Progressive, 2026). Those figures represent behavioral pricing data Progressive generates across both its direct and agency channels, giving it an underwriting signal that is partly independent of how a policy was acquired. Progressive can price direct and agent-placed risks accurately not because its algorithms are categorically superior but because it has calibrated channel-specific selection effects across two decades of accident years.

Allstate is attempting to compress that timeline using pre-trained language models and ALLIE's agentic infrastructure. The compression is real at the customer-interaction layer: ALLIE can complete a transaction faster and more consistently than most phone interactions, and the user experience in an AI direct channel can be genuinely better than a phone queue. The compression is not available at the actuarial layer. Loss ratios on AI-closed policies develop at the same biological pace as any other policies. The first reliable channel-specific frequency comparison arrives after the first policy year matures. Severity does not stabilize until the bodily injury tail closes. Allstate's timeline is compressed by better models; the calendar is not.

The Rate Filing Channel-Mix Question in the Pilot States

Personal auto rate filings are actuarially supported by loss and expense experience developed from the insurer's book as it was constituted when the filing was prepared. If that book was 100% agent-placed, the actuarial indication is calibrated, implicitly, to agent-placed risk selection. Introducing an AI direct channel that closes policies outside the agent relationship changes the risk profile of the book in ways the filed rates do not capture until experience on the new channel accumulates and is reflected in a subsequent revision.

Personal auto rate filings across multiple jurisdictions rarely break out channel-mix assumptions explicitly, and that is precisely where the disclosure gap appears when a distribution model shifts. The rate revision cycle in most states runs on a 12-to-24-month cadence; Allstate could have AI-closed policies maturing in year two of the pilot before the first channel-specific actuarial indication is available to incorporate into a rate filing. Regulators who approved Allstate's current rates based on an agent-placed book have not reviewed actuarial support for an AI direct channel that selects risks differently. That gap is not a violation -- nothing in state rate filing laws prohibits a new distribution channel without a pre-approval requirement -- but it is an actuarial disclosure obligation that the pilot's launch creates.

The three-state pilot selection almost certainly reflects favorable regulatory posture. The regulatory landscape for AI-driven policy issuance varies significantly: Colorado's AI bias audit requirements under Senate Bill 21-169 mandate documentation of model fairness testing; California's Department of Insurance has required pre-approval of algorithmic rating systems since 2023; other states have adopted the NAIC's AI Model Bulletin framework, which requires governance documentation without mandating pre-approval. The three states where Allstate launched its AI direct pilot are the template for national expansion. The actuarial documentation -- channel-mix experience, adverse action logs, rate adequacy tests -- developed in those three states is the foundation of any subsequent multi-state filing.

For rating actuaries supporting Allstate filings in the pilot states, the right approach is a separately developed actuarial indication for the AI direct channel, even when the early experience is too thin to be credible on its own. A blended indication that averages AI-closed and agent-placed experience into a single rate level compresses the signal from both books rather than surfacing it. When the AI direct book eventually develops a materially different loss ratio than the agent book -- a difference that may not appear in frequency for 18 months and in severity for 36 -- a blended historical indication will have been incorporating the distortion retroactively, and the rate correction will need to be larger than it would have been with separated experience.

Adverse Action Documentation When No Agent Signs the Policy

If AI closes a policy, AI also declines some applicants and adjusts rates for others. Pennsylvania's settlement with GEICO in May 2026 defines the floor for what regulators now expect from carriers whose AI systems touch these decisions. An AI-enabled underwriting review tool GEICO deployed flagged a West Philadelphia customer's new auto policy for further review, leading to cancellation without adequate notice; the customer drove uninsured without knowing her coverage had lapsed (Pennsylvania Attorney General, May 2026). Pennsylvania AG David Sunday and GEICO agreed on a settlement that 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 Allstate's three-state AI direct pilot, the parallel exposure sits at point of sale rather than at renewal. An AI system that declines to issue a policy, assigns a higher rating tier than the applicant's initial inquiry implied, or applies a surcharge based on model output must produce the same adverse action documentation that a licensed agent's underwriting decision requires under state insurance regulations and the federal Fair Credit Reporting Act where credit-based rating factors are involved. That documentation must be contemporaneous, specific enough to explain the basis for the decision, and accessible to the consumer on request. Carriers that have gone through market conduct examinations involving AI underwriting decisions report that retroactive reconstruction of decision chains is materially more expensive than contemporaneous logging -- the three- to seven-figure exercises required to reconstruct AI decision logic after the fact in examination contexts were all avoidable with real-time audit trail infrastructure at deployment. Three-state pilots that do not build this infrastructure create regulatory exposure before they reach the scale that makes gaps consequential.

Three Signals to Watch as the Pilot Develops

First: loss frequency by channel in the pilot states, compared against the agent-placed book in the same rating territories. A material frequency divergence emerging within the first 12 months of policy year data signals a selection effect large enough to be visible in early development. An early frequency divergence does not prove the channel is adversely selected -- newer books can show favorable early frequency on short-tail claims before the longer-tail bodily injury component matures -- but it is the earliest quantitative signal available.

Second: rate revision timing in the pilot states relative to comparable non-pilot states. If Allstate files rate revisions in the three AI direct states sooner than in paired states without the AI direct channel, and the revision justification references channel-mix or experience divergence, that is the clearest external signal that adverse selection is materializing faster than the blended book would indicate.

Third: the scope of adverse action documentation disclosed in any regulatory examination covering the pilot states. The GEICO Pennsylvania settlement establishes what state regulators will now ask to see. How Allstate's documentation holds up in examination is the real test of whether the infrastructure behind the AI direct channel was built to the standard the compliance environment requires. Wilson's "more of a learning" framing is correct. The data that learning generates will be the actuarial record of whether AI-closed personal auto risks are a viable channel expansion or a loss ratio problem accumulating below the surface of a strong combined ratio.