The production layer argument assumes the organisation knows what it's governing. That's the harder problem.

Shadow AI, personal tool use, and AI embedded in SaaS platforms are already inside most organisations — not as planned deployments but as the accumulated result of individuals getting things done with what's available. The person who runs a customer database query through ChatGPT to save time, the vendor platform that updated its product with AI-assisted recommendations last quarter, the analyst who built a workflow around a model the IT team has never heard of — none of these appear on the control plane. None of them triggered an architecture review.

The governance frameworks you describe work downstream of a visibility problem that hasn't been solved. You can build a production layer for the AI you know about. The risk exposure from the AI you don't know about has no recovery point. Unlike a failed software deployment, there's no rollback — the decision already ran, the output already circulated, and the person affected by it has no idea which system produced it.

We've seen this pattern before. Cloud deployments without central oversight produced exactly the same failure mode — spend no one could account for, data in platforms nobody owned, shadow infrastructure that only became visible when something broke. The answer was the Cloud Business Office: a formal function with genuine authority, visibility across the estate, and control over what could be provisioned and by whom. Nobody is building the AI equivalent.

An AI Business Office — with mandate, not just advisory status — is the structural prerequisite for everything the production layer requires. Without it, the same reckoning is coming. The genie probably can't be fully controlled. But it can be tracked, and tracking requires a function that owns the problem rather than a governance framework that assumes someone already does.