Posts

Moving AI governance into the execution layer is the right structural move. What gets less attention is what breaks when you do it.

The emerging consensus is to delegate execution to AI and review the output. But review is inspection, and inspection scales with volume. The constraint is the operating model.

A documented failure mode gets reviewed. A system constraint gets enforced. The difference is whether governance scales with headcount or with the system.

Can a model be substituted without triggering a manual risk assessment? If not, the dependency is on absent execution scope, not the model.

Organisations believe they are captive to AI-model providers. More often, they are captive to their own inability to isolate execution scope, reproduce behaviour under explicit contracts, or generate evidence inside the flow of work.

A control plane solves the AI-model routing problem. It does not solve the authority problem.

AI can generate more change than most organisations can safely absorb. The constraint is not code, it is promotion: a change advances only when it has collected the evidence a gate requires.

The hardest part of AI in a regulated enterprise is not invention or industrialisation. It is the seam between them, moving work into repeatable execution without losing intent, control, or trust.

The first time AI solves a migration pattern, it is novel. The second time, a recipe. The tenth time, deterministic. Context maturity is what makes this compound.

AI makes backward context creation feasible at estate scale. What matters is whether attested context exists — the technology is incidental.