AI fluency: why board training comes before operational training

In the AI training plans we review there is an inverted pyramid: prompting courses for hundreds of operational staff, workshops for middle managers, and for the top, when things go well, a one-hour inspirational keynote with slides about the future. The implicit logic is that AI is learned by using it, so training goes to those who use it. The forgotten logic is that the decisions determining adoption's success, how much to invest, where, with which acceptable risks, when to stop, are all taken at the top, and they are taken badly when the people taking them lack the instruments to interrogate what is being proposed.

The labour market, meanwhile, runs in the opposite direction to the inverted pyramid: the Artificial Intelligence Observatory of Politecnico di Milano (2025 research, February 2026) records a 93% increase in AI skills requested in job postings and 41% of workers performing, thanks to AI, activities previously out of reach. Widespread skills grow on their own, pushed by demand; governing skills do not, and they are scarcest exactly where they matter most.

An obligation, even before a choice

There is also a regulatory fact many training plans ignore: from Article 4 of the EU AI Act, applicable since February 2025, organisations carry the task of ensuring an adequate level of AI literacy for the staff operating these systems, commensurate with context and role. The phrase "commensurate with role" is the point that interests us: for those who resolve strategies, allocate capital and answer for governance, adequate literacy is not knowing how to write a prompt, it is knowing how to exercise one's own trade on this matter. A board approving an AI investment without the instruments to evaluate it has a compliance problem on top of a decision-quality one.

What a board must know (and what it must not)

The misunderstanding to dissolve concerns the content, because top-level fluency is not a reduced version of technical training. A board does not need to know how to build, it needs to know how to interrogate, and the curriculum that works organises itself around four capabilities.

The first is understanding the probabilistic nature of these systems, just enough to ask the right questions about error: how often it fails, on what, who notices, what it costs. The second is the economics of AI, the five cost families we have written about, the difference between a pilot and a production system, why returns come from process redesign more than from the tool. The third is the regulatory perimeter in its managerial consequences, what having high-risk systems entails, which questions to ask about data and suppliers, when to demand that risk and the DPO be at the table. The fourth, the most precious, is the repertoire of questions: the ten questions a director asks in front of any AI proposal, from "who signed the expected value" to "what is the stopping condition".

Everything else, architectural details, model taxonomies, tool demos, is ballast at this level, and the programmes that include it buy hours of the top's attention only to spend them badly.

The right sequence, and what it unlocks

Training the top before the operations has a sequencing justification, beyond a hierarchical one: it unlocks everything else. A fluent board approves widespread training plans with criteria instead of by catalogue, tells suppliers apart from brochures, sustains the right projects through difficult quarters (the Confidence we discuss in the rating is built this way too) and stops oscillating between keynote enthusiasm and newspaper-article panic. In the journeys we run, a well-executed half day with the board is worth months of friction saved downstream.

The format counts as much as the content: short recurring sessions beat the marathon day, the company's own cases beat textbook examples, and the most effective exercise we know is the resolution simulation, a real AI proposal brought into the room and dismantled with the repertoire of questions. In our AI Academy the tracks are built by level, board, management, operations, with this sequencing principle embedded: calendly.com/fabiolalli/zerofive, or hello@zerofive.ai for the programme.

The test to see whether your top is already fluent, meanwhile, is free and slightly cruel: at the next AI proposal in the boardroom, count the questions asked before approval. If the most technical one concerns the licence cost, the pyramid is still inverted.