Consulting Desk · AI Op Model Design
Who commissions AI work.
Who can stop it.
Written down, before it matters.
Who commissions, who reviews, what a human must decide, where escalation goes, and how the whole thing is funded. Designed around decision rights and review gates first — because an operating model that cannot say who stops a project is an org chart with new job titles on it.
You hold: the operating model itself, a decision rights matrix against real decisions, not activities and review gates.
The test an operating model has to pass
Name the person who can stop it. If you cannot, you do not have one.
Most AI operating models are drawn as boxes: a centre of excellence, a governance forum, a delivery function. They look organised and they answer nothing, because the question that matters in practice is never “whose box is this in” — it is “who decided, and who could have said no”.
So we design from decision rights outward. For every class of AI work: who may commission it, who reviews it before it goes anywhere, what specifically must be decided by a human rather than a system, where a disagreement escalates to, and who holds the authority to stop something that is already live. Then the boxes, if you still want them.
The funding model is part of the design, not an afterthought. An operating model where the delivery function is funded but the review function is not will produce exactly the behaviour you would expect, no matter what the policy says.
We run this ourselves
This is not a framework we read about. It is the operating model our own governed AI workforce runs on, published in full on our delivery model — review gate, integration gate, human decision, and the standards corpus behind it. You are buying something we are exposed to daily.
How it runs
Decision rights and review gates first.
Five artefacts, in this order, because each one constrains the next.
- EngagementAI Workforce Deployment — 8–16 weeks
- MethodDecision rights and review gates first
- GateEvery AI decision has a named human owner
- Also runs underFractional AI Leadership, where a founder holds the seat
- Ends inWritten standards your teams actually read
The decision rights matrix
Per class of work: who commissions, who reviews, who approves, who can stop. Named roles against real decisions, not a RACI covering activities nobody disputes.
The review gates
What has to be true before work moves. Where a gate exists on paper but nothing is checked, we say so — an unenforced gate is worse than none, because it is relied on.
The human boundary
What a human must decide. Our own line: anything a reasonable person could disagree about, anything irreversible, legal, regulatory or safety-critical. Yours may differ; it has to be explicit.
Escalation routes
Where disagreement goes, and in what form. Escalation should arrive with options and a recommendation attached, never as a bare question — that one rule changes how much senior time a programme consumes.
Funding and standards
How the model is paid for, and the written standards each function works to. Short enough to be read, specific enough to be failed against.
The two lists that matter
What you hold, and what we will not do.
Both are in the engagement letter before you sign it. The second list is the one worth reading twice — it is where most disappointment in this market actually comes from.
What you hold at the end
- The operating model itself — commissioning, review, approval, stop
- A decision rights matrix against real decisions, not activities
- Review gates, with an honest note on which ones are currently unenforced
- Escalation routes, and the required form of an escalation
- The written standards your teams read, and the funding model behind them
What we won’t do
- Operating models that cannot name who stops a project
- Governance forums with no decision rights attached
- Roles defined without a decision behind them
- Autonomous processes without human decision gates
- Standards documents too long for anyone to read, let alone apply
Design, then run it
An operating model nobody has operated is a document.
The design is contracted as an AI Workforce Deployment because the two belong together: we stand up a governed workforce for a defined function — code review, incident triage, compliance checking, research synthesis — and the operating model is what governs it. You end up with a model that has been run, not one that has been reviewed.
What gets stood up →Assurance is the other half
The model says who decides. Governance & assurance produces the evidence that they did — the inventory, the risk classification, the audit trail that still answers a question a year later.
Before you commit
Decision rights, not a policy document.
Is the output a policy?
No. It is decision rights and review gates against real decisions rather than activities. An operating model nobody has operated is a document, so we record which gates are actually enforced and which are currently aspiration.
Who ends up owning each decision?
A named person, or it is recorded as unowned. Unowned is a legitimate output and often the most useful line in the pack, because it is the one a board can act on immediately.
How is this different from governance and assurance?
This decides who decides. Governance and assurance then evidences it against the obligations you are under. Most organisations need the first before the second is worth building.
Who stops it, in your organisation?
Thirty minutes with a founder. If you can answer that question already, you may not need this engagement, and we will say so.
Fixed fee · Quoted in writing before we start · NDA available