In May, Liberty Mutual launched a conversational AI auto insurance quoting app that lets customers get a quote through OpenAI's ChatGPT instead of filling out an online form. The launch rests on more than a decade of work to move workloads to the cloud, slim down data centers, and simplify systems-an infrastructure overhaul designed to make the 114-year-old insurer ready to scale AI.
Andrew Palmer, EVP and CIO of global retail markets at Liberty Mutual, said the company is building an abstraction layer on top of its mainframe data so it can work with multiple AI providers without getting locked into one. The company is using MCP (Model Context Protocol), an open standard that connects AI models to where data resides, and is working with mainframe partner IBM on the layer. "That ability to maintain optionality as well as context and memory within our models so we can traverse across a very complex business and not get anchored into one is a strategic bet, but one we're quite confident in," Palmer said.
Cloud and data readiness
Liberty Mutual has moved 85% of its workloads to the cloud, consolidating from 13 global data centers to one primary facility. Palmer said the company runs a hybrid cloud strategy across AWS, Google, and Azure, and applies the same multi-provider logic to AI. "With the frontier labs, it's especially important," he said, referring to the rapid pace of model releases from leading AI companies.
Data quality is central to the plan. Palmer argued that inefficient data access burns tokens and erodes model accuracy. "You can't run AI on bad data," he said. "If you're making it sweat working across all these different stores and applications and trying to figure out what's going on, you're burning a ton of unnecessary tokens. You probably also won't get the right outcomes you want from a context standpoint."
Mainframe modernization and the abstraction layer
The insurer's core platforms still run on the mainframe, and a five-year modernization effort is underway to replatform to Guidewire Software. In the interim, the abstraction layer built with IBM lets front-end cloud applications use mainframe data through the latest AI capabilities. An AI FinOps team tracks token consumption across the company, with the ability to set caps or add capacity for specific use cases. Palmer said he is pushing operations managers to link token spend directly to business results.
Why this matters for insurance professionals
Liberty Mutual's experience shows that AI for insurance cannot be bolted onto weak infrastructure. Palmer issued a blunt warning for companies that skip the foundational work. "If you're trying to just chase the shiny toy and you think you don't have to do the heavy lifting, that's going to catch up with you," he said. For insurance leaders, the takeaway is clear: model-agnostic architectures, rigorous data readiness, and disciplined token oversight are not optional-they are the prerequisite for turning AI experiments into measurable business value.
Your membership also unlocks:
