Enterprise AP Invoice-Approval Agent Spec

The single highest-leverage artifact in the Enterprise AI PM skill set: the spec a senior PM produces before a money-moving agent gets built. The scenario is deliberately ordinary: a 200-person SaaS company processing 5,000 supplier invoices a month, auto-approve the clean ~95%, escalate the risky ~5% to the right human in under 30 seconds. The repo is the full document set: a ~4,000-word PRD with a five-level escalation decision tree and a trust boundary that caps any autonomous action’s blast radius at $5,000; a 14-case eval suite that runs against a stub and is built to fail a naive approve-all agent; a reproducible cost model showing hybrid model-routing at ~$27/month against ~$29,150/month of manual labor; a build-vs-buy memo with a defended recommendation; and a governance mapping to SOC 2 (CC6.1 / CC7.2 / CC8.1) and SR 11-7 model-risk expectations. Grounded in a five-angle deep-research synthesis; the discipline behind it (route the cheap work to the cheap tier, bound what the agent does alone, test it with evals that fail when it’s wrong, log every decision) is the discipline I run on my own autonomous agent fleet.