The scarce skill isn't prompting

The OECD spent years studying AI and skills and published the summary on 5 June. Fewer than 1% of workers need advanced AI skills. Not ten percent. Less than one. I have read that line about thirty times, because almost every AI training budget I have ever seen is built as if the number ran the other way.

Before I argue with it, the five things I read last week:

• The OECD: fewer than 1% of workers need advanced AI skills. The skills actually in demand in AI-exposed jobs are management and business skills. (OECD)
• Cloudflare's internal AI use jumped more than 600% in a quarter. The productivity gains ran from 2x to 100x, on the same tools. (Fortune)
• Gallup asked 23,717 American workers. 65% feel more productive. One in ten say the work itself has changed. (Gallup)
• A study of more than 12,000 European firms found AI lifts productivity about 4%, and every extra point spent on workforce training adds 5.9 to that, against 2.4 for software and data. (CEPR)
• Singapore has a billion-dollar AI strategy. The thing it built this month was a coach, not another grant. (Singapore AI Association)

So what do I take from all that?

We are skilling the wrong people

The reflex everywhere is the same. Buy the licences, run the prompt courses, get the whole company "AI fluent." The OECD just told us that is solving a problem almost nobody has. Baseline fluency is real, but it is cheap and getting cheaper. The tools teach it themselves now, in the flow of the actual work, the way nobody took a course to learn to use a search engine.

The scarce skill, the one the OECD found in demand, is management: deciding which work actually changes, redrawing the role around it, and reviewing an AI-assisted draft without either waving it through or quietly redoing it by hand. That skill is expensive, rare, and funded almost nowhere. We pour money into the cheap one and starve the costly one, then act surprised when nothing lands.

The European study makes the same point with a multiplier. The technology itself buys you about four percent, but every point spent on people adds nearly six, more than double what software and data return. The money is in the human layer and the budgets sit in the tool layer, and we have it backwards.

The 100x person was managed, or they weren't

This is why the Cloudflare spread reads the way it does to me. 2x to 100x on identical tools is not a prompting contest. The person at 100x had the room, the permission, and probably the manager to rebuild their own work around the model. The person at 2x had the same licence and none of that. Somebody manages the distance between those two people, or nobody does, and in most companies nobody does. There is no role that owns it, so it slips between L&D, who run the courses, and IT, who run the rollouts, and lands on no one.

I think that gap is the most expensive empty chair in the building right now.

Singapore made the bet out loud

Singapore had every excuse to make the other choice. With a billion-dollar national strategy, subsidised tools, and 64% of its businesses already using AI, the obvious move was to spend more on all three. Instead the government looked at stalled progress and decided the missing piece was not more money or more software but a person sitting next to the work. Its new AI Association now runs Discovery Clinics, where a business owner brings one real problem and walks out with a redesigned use case matched to training, coaching rather than a catalogue.

A government with a billion to spend decided the bottleneck was coaching capacity. That should tell you something about your own budget.

Where I would put the money

Moderna did exactly this in May 2025, when it put HR and IT under one person, Tracey Franklin, with a title that gives the game away: Chief People and Digital Technology Officer. Her line was "I've shifted from workforce planning to work planning." That is the management skill the OECD is talking about, sitting at the top of a company instead of in a prompt course at the bottom.

There is a fair objection here. You cannot redesign work with people who cannot use the tool at all, so the prompt course is not wasted. True, up to a point. But baseline fluency is a weekend, not a budget line, and the tool closes most of that gap on its own. What does not happen on its own is a manager learning to look at a process and rebuild it. One critic of the Moderna move, Jim Woods, argued they automated HR's compliance and called it reinvention. He might be right. But the direction is the one that matters: put the scarce skill where the decisions are.

If you run L&D, HR, or transformation

Look at the split in your AI budget this week. How much goes to tools and universal fluency, and how much to the managers who have to redesign a workflow, redefine a role, and review AI-assisted work? If it is the first number that is large, you are funding the under-1% problem.

I am not against training. I am saying it is pointed at the wrong people. So, two moves. Move a real slice of the prompt-training budget to the management layer, and measure that group by workflows redesigned, not courses completed. Then fund the empty chair: one person per team, slightly ahead, whose actual job is to sit next to the work and help the others get better. Not a trainer. A coach. Singapore just spent a billion dollars working out that this role matters. You can copy it for the price of a licence renewal.

The provocation

So, one thing to look at before Friday. If you added up every euro you will spend on AI skills this year, what share of it reaches the handful of people who decide how the work is actually done? If the answer is almost none, you are training the wrong people, and no amount of fluency at the bottom will fix a constraint that lives at the top. Who owns the redesign in your organisation, and have you ever given them a budget?