John Z Black
Brunswick, ME • (207) 245-1010 • contact@johnzblack.com
Enterprise AI governance still gets treated like a policy-writing project. It’s not anymore.
Unit 42’s analysis of AI agents flags the familiar risks: broad permissions, weak tool boundaries, incomplete visibility into agent behavior. Cloudflare’s Custom Regions release addresses the other side of the same coin, enforcing where sensitive traffic and processing can happen. Then the ICML reviewer controversy shows what happens once rules are actually enforced: things get messy fast.
The real governance stack has three layers. Agent authority boundaries: what can the model call, what can it change, what needs approval. Data and processing boundaries: where data can be inspected, transformed, or moved. And enforcement behavior: what actually happens when policy is violated, not just what the memo says.
Most companies can produce governance documentation. Far fewer can show auditable agent permission models, enforceable data-region controls, and a repeatable response to policy violations. That’s the gap that determines which programs hold up under real scrutiny.
If a rule can’t be technically enforced and organizationally defended, it isn’t governance. It’s intent.
Before approving a platform or internal AI deployment, ask four questions. Can high-risk agent actions be gated or require human approval? Can you prove where sensitive processing occurred? Are logs good enough to reconstruct model actions and data paths? Do you have a credible process for governance violations? Fuzzy answers mean governance maturity is still low.
Full post covers how to structure permissions, data boundaries, and enforcement into a real program.