Allstate Is Building Its Own Large Language Model, ALLIE
Allstate is building a proprietary large language model. CEO Tom Wilson told analysts on the second-quarter earnings call that the carrier’s next step is ALLIE, short for Allstate Large Language Intelligent Ecosystem, built on the data platform the company has been assembling for years. The model is not complete or deployed, Wilson said, but it “should help us to reduce our expenses,” with more accurate pricing and claims among the targets. He declined to put numbers on expected savings or growth.
The platform underneath is already at carrier scale. Wilson described 250 analytical models running on 40 petabytes of data, generating 100 million quotes, buying 50 million leads, often at sub-second response times, and managing hundreds of millions of customer interactions. ALLIE, he said, will “leverage agentic AI to improve customer value, lower cost, and increase growth.” Distribution expense is the first visible target: the Property-Liability president is already taking work out of agent offices.
Two details in the call matter more than the announcement itself. The first is architectural: Wilson said Allstate uses only internal LLMs, so the company is “not worried about our data being exfiltrated or scooped up in the knowledge of someone else’s LLM, so one of our competitors can use it.” The second is what he said next. He is “concerned about the large language models that now break out on their own and tell other agents how to do it without getting caught,” adding that this “is an issue that the country really needs to deal with.” A carrier CEO referencing containment failures on an earnings call, in the same week three labs disclosed escapes from the same evaluation vendor, tells you how far that story has traveled.
An in-house model still has to be governed, and the file looks different when there is no vendor on the other side of it. The vendor oversight checklist exists because a carrier answers for models it licenses. When the model is built internally, the third-party diligence questions move inward and lose their enforcement mechanism. There is no contract to write audit rights into, no SOC 2 to request, no vendor to terminate. The evidence has to come from the carrier’s own model monitoring records and its documented assignment of who owns which risk, produced for an examiner who will apply the same standard either way. Wilson’s own framing raises the bar. Exhibit B leaves the tiering to the carrier and asks it to reference the processes by which risk levels get assigned, with a suggested question aimed at autonomy, reversibility, and reporting impact. A carrier that calls itself technology-driven rather than technology-supported, and that points agentic AI at pricing and claims, has to answer that question about itself.
The build-versus-buy question moves from IT strategy to board agenda the moment a top-five carrier concludes that competitive safety requires owning the model rather than licensing it. Allstate’s answer rests on 40 petabytes of proprietary data and years of platform spending that most carriers do not have. The ones that cannot replicate it will license, which means the vendor-oversight burden concentrates on them rather than dispersing.