Sergio CastagnaA chief executive's competence on artificial intelligence is not measured by what they understand about the technology. It is measured by the quality of what they refuse to approve without a clear answer.
The useful questions are not the ones that produce information. They are the ones that cannot be answered with reassuring vocabulary. A confident "yes, we're working on it" should disqualify the answer, not close the discussion. What follows are ten such questions — chosen because each is difficult to evade, and because the shape of the evasion tells the executive more than any roadmap.
1.Which line of the P&L does this initiative improve, and by how much?
This is the question that separates a strategic project from a curiosity that has been dressed up as one. If the impact cannot be tied to growth, margin, cash, productivity or quality, the company is funding a research project under another name. The absence of a number is itself the answer.
2.Are we building a defensible advantage, or renting a commodity our competitors rent too?
The frontier models are available to everyone. The moat is never in the technology — it sits in proprietary data, in distribution, in the depth of integration into the business, in the speed of execution. A company that mistakes adopting the same tools as everyone else for a competitive edge has confused access with advantage.
3.Do we have a usable data asset, or a legacy that AI will only amplify into disorder?
Artificial intelligence does not compensate for poor data. It amplifies it. The majority of failures attributed to AI are, on inspection, data failures — incomplete, fragmented, unreliable — that surface only once the investment has been committed. Better to know before than after.
4.What is the organisation's real capacity to absorb this change?
A sound strategy fails if the company cannot digest it. AI does not bolt onto existing processes; it redraws them, and it touches roles, habits, skills and fears as it goes. Move too fast and the organisation rejects the graft; move too slowly and the ground is ceded to others. This pace is the chief executive's to set, and no one else's.
5.On exactly what conditions does this pilot move into production — and who has the power to say no?
Without an explicit exit, the company enters the purgatory of pilots: many demonstrations, much enthusiasm, and a P&L that never moves. Scaling should require proof — of unit economics, of reliability, of security, of compliance, of real adoption by the teams. A pilot that survives because no one has defined the conditions for stopping it is not a pilot. It is a leak.
6.Who owns AI risk in the company, and can they actually stop a deployment?
As long as the answer is "everyone and no one," there is no governance — only an intention that reassures. The decisive test is not the org chart. It is whether anyone holds the power to block a deployment, and the last time that power was used. A governance that cannot say no does not exist.
7.What are our teams already doing with AI, outside any framework?
In most companies, AI is already in use before the leadership has set any rule for it. Shadow AI is often the first real risk: sensitive data pasted into external tools, unvalidated content shipped to customers, decisions quietly assisted by systems no one authorised. Governing does not begin with what comes next. It begins with reclaiming what already exists.
8.How will we know the system is drifting before the customer does?
A model does not break. It degrades. The quality slips, the answers drift from reality, and nothing lights up. Most organisations discover the drift through a complaint, an incident or an unflattering article — which is to say, too late. The question is not technical. It is whether the company insisted on seeing before others would.
9.If this vendor doubles its price or disappears, what happens?
This is the reversibility stress test, and it turns a procurement matter into a strategic one. Dependence on a single model, API or platform becomes a question of margin, of continuity, of negotiating power. Building a critical capability on something that cannot be replaced without major cost is a governance decision — not a technical detail to be settled below board level.
10.Even if AI can do it, should we?
This is the only question on the list that does not concern feasibility. The chief executive is the guardian of trust — the asset that takes longest to rebuild. Not every possible automation is desirable, and some choices that are efficient in the short term are corrosive to the brand, the culture, the relationship with the customer. The question protects what the others cannot.
There is an eleventh question, and it is not addressed to the organisation. It is addressed to the chief executive alone:
Am I able to challenge these answers, or do I approve them on the word of those who produce them?
That question decides the value of the other ten. A leader does not need to understand how a model works. They need to ask the questions whose answers cannot be dressed up — and to know the difference between an answer and an evasion.
These ten questions are not the framework. They are its skeleton.