Tow Claims: A Multi-Agent AI System for Autonomous Claims Processing

Tow away coverage helps customers tow a broken-down vehicle that is not covered under comprehensive or collision. About 10,000 of these reimbursement claims came in every month, and the entire process ran on fax and email. Duplicate and error-filled faxes slowed everything down, turnaround ran anywhere from 2 to 8 days, the only payout method was a paper cheque, and the whole flow depended on a third-party document tool. It was slow, clunky, and a poor experience for everyone in the chain.

The clearest sign of how broken it was came from inside the company. One of Allstate's own senior executives hit this process personally and could not even find a fax machine to complete it. When a leader cannot get through your own process, the case for change makes itself.

I led the product and design for an agentic AI system to replace that process with straight-through processing, working jointly with engineering. The goal was to move from a reactive, manual workflow to intelligent, autonomous decision-making, while keeping a human in the loop for trust and control.

I drove the key design decisions on a multi-agent architecture built on Google's ADK. A supervisor agent orchestrates the entire claim journey and routing. An invoice agent reads and interprets invoices using LLM-based document understanding. A tow agent validates policy coverage and makes the claim decision, backed by policy-info, claim-creation, and claim-validation functions. The system connects to Allie, Allstate's assistant, so the experience can run conversationally rather than through faxes and forms.

The tempting path was full automation. Let the agents process every claim end to end, show the highest possible straight-through rate, and book the biggest savings number. In a regulated claims process, that is the wrong trade.

I made the call to keep a human in the loop, with explicit checkpoints and a full audit trail built in from the start, rather than chase maximum autonomy. That meant accepting that some claims would route to a person instead of clearing automatically. In exchange, every decision the system made was explainable, traceable, and defensible to compliance, with an exception rate under 30 percent and an explanation attached to each one. In financial services you do not get to choose between autonomy and accountability. I designed for both.

The system was built and delivered as a working MVP. These are early-testing and projected figures, with a staged path to production, and I present them as exactly that.

In early testing, the invoice agent reached over 95 percent extraction accuracy and policy validation reached over 99 percent accuracy, with the exception rate held under 30 percent and every exception explainable. The projected business case is $400K to $600K in annual operational savings, an 85 percent reduction in manual effort, and turnaround compressed from 2 to 8 days down to minutes, with the capacity to absorb volume without adding staff and automated validation that removes the duplicates and errors that plagued the fax process.

The roadmap is staged: an AI-driven intake and validation pilot first, then ecosystem integration with Allstate Motor Club for single-channel payment, then payments integration within 90 days of pilot success, then wider deployment. The architecture was deliberately built to extend beyond tow, to FNOL, glass, and roadside, as the foundation for enterprise-wide claims automation.