What this layer actually measures
It measures the ability to choose between cases.
An organisation mature here can take two proposals competing for the same budget and choose on explicit criteria. An immature one does both halfway, or does the one belonging to the executive with the most political weight.
What it does not measure: how many people you have dedicated to AI. We have seen excellent operating models run by two part-timers, and disasters run by teams of fifteen.
Why mid-sized companies get it wrong
By importing structures that do not belong to them.
The dominant advice — centre of excellence, AI steering committee, transformation office — is calibrated for organisations of thousands where the real problem is coordination between units that do not talk to each other. In a company of three hundred that problem does not exist: people know each other and pass in the corridor. Importing the structure imports the coordination cost without importing the benefit.
A three-person centre of excellence is not a centre of excellence. It is a committee with an ambitious name, and it usually ends up bottlenecking the initiatives it claimed to accelerate.
The opposite mistake is common too: deciding nothing and letting each area do its own thing. It works surprisingly well for the first few months — genuine cases emerge from below — and stops working the moment someone has to choose, integrate or answer for something.
When structure is needed, exactly
Our practical threshold: when three live cases compete for the same budget or the same people.
Before that, structure is overhead. With one case, the right operating model is a sponsor and an owner. With two, the same plus a monthly conversation. With three you start needing a common yardstick to compare them, and that yardstick is what justifies governance.
It is a deliberately low threshold compared with what the market suggests, and even so most companies that set up a committee have not reached it.
The role that does work in mid-market
What we see work is not a department: it is a named function with allocated time. Someone who answers for the AI portfolio, has authority to stop initiatives, and spends a real fraction of their week on it — not "on top of your existing duties".
It can be internal, and that is the better option when the person exists. The usual problem is that in a mid-sized company that person does not: whoever has the technical judgement lacks the political weight, and whoever has the political weight lacks the time and the technical judgement. Hence this function is frequently outsourced for the first twelve to eighteen months, with an explicit transfer commitment. Without that commitment, the dependency becomes structural and the differentiation stays outside the house.
Diagnosis: where you are
Stage 1 · Scattered initiatives. Each area does its own thing. Nobody has the full picture. Observable signal: if you ask how many AI initiatives are running, you have to ask several people and the answer is still incomplete.
Stage 2 · Owned portfolio. There is a single list, someone answers for it, and there is a review cadence. Observable signal: you can say what is being spent and who decides what comes next.
Stage 3 · Governance that can kill. Cases are compared on the same yardstick and budget is pulled from those that do not perform. Observable signal: in the last twelve months an AI project that worked technically but did not pay was stopped.
That last point separates real governance from theatre. A committee that has never said no is a decorative committee.
The first 90 days
Weeks 1-2 · Inventory of initiatives. All of them, including those that never went through IT. There is usually more than expected.
Weeks 3-4 · Name the function. Who answers for the portfolio, with what explicit weekly allocation and what authority to stop. Without all three, the appointment is symbolic.
Weeks 5-8 · Common yardstick. The criteria for comparing cases: measured value, cost, risk, effort. Written before you need them, so they cannot be calibrated to suit the favourite case.
Weeks 9-12 · First portfolio review with a real decision, however small. Governance that decides nothing in its first session rarely decides anything later.
What NOT to do yet
- Do not create an AI department. It concentrates in one place what has to happen everywhere, and turns the rest of the organisation into a customer rather than an owner.
- Do not build a CoE with three people. Call it what it is: a committee, and that is fine.
- Do not appoint an owner without time. "On top of your duties" means this happens on Friday afternoons, if at all.
- Do not outsource the decision. Operations can be outsourced; deciding where your company invests cannot.
What skipping it costs
It costs dispersion, and dispersion is lethal in mid-market because the scarce resource is not money: it is the three or four people capable of carrying this. Spread across five simultaneous initiatives, none reaches production.
The late symptom is recognisable: eighteen months on there are six pilots, none in production, all with some progress and none with a clear owner. At that point the problem is neither technical nor budgetary; it is that nobody has authority to pick three and kill the other three.
Sixth layer of the series on the seven layers of AI maturity.
Frequently asked questions
When does a company need an AI committee?
When three live cases compete for the same budget or the same people. Below that threshold, a sponsor and an owner are enough; a committee adds process without adding judgement. Its real function is not to accelerate, it is to be able to say no.
Do you need an AI centre of excellence?
In a mid-sized company, usually not. A centre of excellence with three people is a committee, and it tends to become a bottleneck. The CoE structure solves a coordination problem between units that does not exist in an organisation of three hundred.
What is an AI Officer and when does it make sense?
It is the function that answers for the AI portfolio: prioritises, stops initiatives and answers for the spend. It makes sense from the moment there are several cases. It can be internal if someone combines technical judgement and political weight; when nobody does, it is outsourced with an explicit transfer plan.
Can the AI function be outsourced?
Operations, yes; prioritisation, no. If the vendor decides which cases happen and which do not, the vendor is running your strategy. Any arrangement like this should state how and when the function comes back in-house.
Jordi García is Tech Lead at onext. He works on bringing AI into governed production across development and product teams —with Spec-Driven Development, context engineering and human verification at every step— and authors onext's technical insights on the method, quality and cost of applied AI.
LinkedIn →