Governance is a decision accelerator, or it is theatre
Almost every AI governance framework is written as a set of restrictions. That is why almost every one of them is routed around within two quarters. The frameworks that survive are built the other way up: their purpose is to let a decision be made quickly, by having already answered the question that actually causes delay — who is allowed to say yes to this, and on what evidence.
Where the time actually goes
Watch an AI proposal move through an organization and time each stage. The model work is rarely the long pole. The long pole is a sequence of pauses: waiting for legal to form a view with no precedent to work from, waiting for security to assess a system nobody has described in their terms, waiting for a data owner who has not been identified, and waiting for an executive to accept a risk nobody has quantified.
None of those pauses are caused by the technology. All of them are caused by an unanswered question about authority. And every one of them is answerable in advance, once, rather than repeatedly per project.
The delay is almost never technical. It is an unanswered question about who is allowed to decide.
The artefact that removes the pause
A decision-rights matrix is unglamorous and does more for delivery speed than any policy document. It states, per class of use case, who approves, what evidence they require, and what the escalation is when they decline. It is short. It fits on a page. Its value is entirely in having been agreed before anyone is waiting on it.
The tiering is what makes it workable. A single review process applied to everything guarantees one of two failures: heavy enough to stall low-risk work, or light enough to wave through the consequential cases. Tier by consequence instead — who is affected, whether a human decides, whether the outcome is reversible. An internal drafting assistant and a system that influences an eligibility decision should not travel the same path, and pretending otherwise is what makes teams route around the process.
What this is worth
Do the arithmetic with your own numbers rather than ours. Take the last three AI proposals your organization considered. Count the elapsed weeks from first proposal to a decision either way, then subtract the weeks in which someone was actively working on it. The difference is queue time, and queue time is what a decision-rights matrix removes.
For most organizations that gap is the majority of the elapsed time, which means governance done properly is not a tax on delivery — it is the largest single lever on how fast the organization can adopt anything. The reframing matters commercially too: a governance programme sold as risk reduction competes with every other risk programme for budget. The same programme sold as decision latency competes with nothing, because nobody else is offering it.
The inventory problem, and why it comes first
One precondition. Organizations consistently underestimate how many model-driven systems they already run, because most arrived inside purchased software rather than through a project with a business case. Governance written against a partial inventory governs the wrong half, and the half it misses is usually the half already touching customer data.
Discovery across procured tools, embedded vendor features, and team-level experiments typically returns a list two to three times longer than leadership expects. That list is not an embarrassment; it is the scope. It also tends to be the artefact that converts governance from an abstract programme into an obviously urgent one.
The cost line nobody budgeted for
The inventory has a second effect that tends to surprise finance more than it surprises technology. Estates assembled team by team accumulate duplicate capability: several transcription tools, several document-summarisation subscriptions, several vector stores, each procured separately because no single purchase was large enough to trigger a review.
Consolidation is the obvious saving and the smaller one. The larger saving is in review effort. Where every proposal is assessed from first principles, the same legal, security, and privacy analysis is repeated per project by expensive people. Tiering converts most of that into a reusable pattern — a class of use case is assessed once, and subsequent cases inside the class inherit the assessment. The marginal cost of the second and tenth use case drops toward the administrative floor.
The sequence, then, is inventory, then tiering, then decision rights — and only then policy. Most programmes run it in the opposite order, publish a policy against an unknown estate, and wonder why nothing moves faster or costs less.
