AI adoption vanity metrics are pervasive: more code generated, more PRs, faster cycle time. But if those gains come with higher defect escape, rollback rates, and incident load, are they truly gains?

In this session, we'll design an AI ROI metric tree for engineering leaders, linking leading indicators (adoption coverage, review load, flaky test rates, context drift signals) to the outcomes your business cares about: stable throughput, lower incident rates, and reduced remediation tax.

We'll also cover "integrity constraints" that prevent metric gaming, like requiring that cycle-time improvements not coincide with rising production defects. You'll leave with a measurement blueprint you can hand to your ops or analytics team to instrument AI-assisted engineering without lying to yourself or your board.