Instead of replacing the mainframe, we built an agentic workflow layer that sits alongside it: a planner-executor architecture that reads incoming claims, gathers supporting documentation, cross-references policy terms, and drafts a recommended action — stopping short of any action a human hasn’t approved. The mainframe itself was untouched; the agent layer integrates through the same screens a human adjuster would use, via a controlled automation interface.

Every agent action that would touch claim status or payout is routed through a human-in-the-loop approval gate, with adjusters able to see the agent’s full reasoning trail before approving.