When missing a deadline permanently forfeits a client's legal rights, "move fast and iterate" isn't an option. This talk covers the design of an agentic AI system for automating maintenance-fee and spend decisions in enterprise IP operations, built to reduce cost and deadline risk while keeping humans authoritative over the decisions that matter. Drawing on published research and production experience, I'll cover how to scope agent autonomy safely, where policy-controlled guardrails belong in the decision pipeline, and how to design for auditability from day one rather than bolting it on after an incident. Practical patterns for anyone building agentic systems where mistakes are expensive and irreversible.