Picture two people on the same team. One has just finished a well-reviewed AI certification — a dozen modules, a badge for the LinkedIn profile, the lot. The other has never sat a course in their life. Hand them both the same problem and the same tools, and the second one has something useful by the end of the afternoon, while the first is still typing “write me a strategy” into the box and wondering why the output reads like a brochure. I’ve watched a version of this play out more than once. The certificate measured something. It just wasn’t the thing that mattered.

The numbers around all of this are alarming, and worth taking seriously. Research from the University of Birmingham puts the cost of the UK’s digital skills shortage at up to £27.6 billion a year by 2030, with the equivalent of 380,000 jobs at risk. TechRadar ran a piece in April, by Cognizant’s Rohit Gupta, making the case, rightly, that hands-on digital skills will decide how much value AI delivers. I agree with every word of it. But look closely at what people need to do useful work with these tools, and that list has almost nothing in common with what corporate training programmes are busy buying.

So here’s my view. The skills that pay off in AI-augmented work aren’t the ones in any training prospectus, and the reason is simple: the ones that matter can’t be certified, and the training industry is built almost entirely around things it can. That mismatch is why so much training budget produces confident, badge-holding people who still can’t get a decent result out of a model.

When facts and first drafts are effectively free, the premium shifts to what you do with them. I’ve made that case before, in AI doesn’t replace thinking, it replaces knowing; what follows is the practical other half. If thinking is the scarce asset, what does it actually look like on a Tuesday afternoon with a deadline and a half-finished task? In my experience it comes down to four things, and not one of them has a curriculum. I won’t pretend the list is complete; I’m not sure how well some of this holds up in a year, given how fast the tools are moving, but these are the four I’d bet on today.

The four skills that actually pay off

Judgement is the first, and probably the foundation for the rest. A model will give you an answer to almost anything, delivered with the same easy confidence whether it’s right or complete nonsense. The skill is being able to tell the difference — to read a fluent, plausible paragraph and know in your gut that the third claim is wrong. That instinct only comes from having done the work yourself, often enough and badly enough, to smell when something’s off. You can’t download it. It’s the slow residue of experience, which is exactly why no module can hand it to you.

Prompt design sounds technical, but it really isn’t. Most disappointing AI output isn’t the model failing; it’s a vague instruction getting the vague answer it deserved. The people who get extraordinary results are usually just the ones who can say, precisely, what good looks like — the context, the constraints, the audience, the thing they’re actually trying to achieve. That’s not a trick you memorise from a list of “prompt formulas”. It’s the ordinary discipline of clear thinking, learned the way it’s always been learned: by trying, getting it wrong, and adjusting.

Cross-functional translation is the quiet one that’s becoming enormously valuable. It’s the ability to take a messy business problem and reshape it into something a machine can act on. The marketer who can describe a data question clearly enough to get a working dashboard out of an AI — without ever filing a ticket to the analytics team — is doing translation, and it’s a big part of why the boundaries between disciplines are dissolving. Knowing both worlds well enough to carry an idea across the gap between them is rare, valuable, and stubbornly resistant to being taught in a slide deck.

Restraint is the least discussed of the four and possibly the most important. It’s knowing when not to reach for the tool at all. Sometimes the right move is to pick up the phone, or to think a problem through on paper, or to write the awkward email yourself because the relationship matters more than the minutes you’d save. The instinct to automate everything is seductive and often wrong. Knowing when the clever tool is the wrong tool is a genuine skill, and it tends to be the one that separates people who use AI well from people who simply use it constantly.

Why training keeps missing them

Notice what these four have in common. None of them sit in a curriculum. You can’t issue a certificate in judgement, there’s no exam in restraint, and “clear thinking under pressure” doesn’t fit neatly into twelve modules with a quiz at the end. They’re built in practice, through repetition and feedback and the occasional humbling mistake — not in a classroom.

This is the heart of it. Corporate training keeps buying what it can measure, and skipping what it can’t. Look at what most “AI training” actually contains and it’s platform fluency — which buttons to press, where the export option hides — dressed up as something grander. Worse, it tends to treat the whole shift as a rollout: a system to deploy, a box to tick by the end of the quarter, rather than the slow change in how people work that it really is. And almost always, the learning happens miles from the job. People earn the badge in a tidy sandbox exercise, then meet their first real, messy problem and haven’t the faintest idea how the two connect.

A certificate is easy to count, easy to put on a dashboard, easy to defend at budget time. The capability it’s supposed to stand for is none of those things, so the certificate gets funded and the capability gets left to luck.

I’ve watched this same mistake through four skills shifts now — the web, mobile, cloud, and data. When the web arrived, companies booked their people onto HTML courses by the dozen, yet the person who ended up building the firm’s first proper website was almost always the one who’d been tinkering with their own at home on weekends, course or no course. The same thing happened with mobile, then the cloud, then data. Each time, the certifications followed the capability; they never created it. Hands-on always won, and I see no reason AI breaks the rule. If anything, with tools changing this fast, the gap between what you can certify and what you can do has never been wider.

A one-question audit for your training budget

If you run a team, or a training budget, you don’t need a framework to act on this. You need one question, asked of every line of spend: does this produce judgement, or does it produce a certificate?

Run it down the list. The AI-fluency course that ends in a badge but never touches a real piece of your work: that’s a certificate. The afternoon where you hand your team a genuine problem and the tools to solve it, then get out of the way: that’s judgement. You’ll be surprised, and probably a little uncomfortable, at how much of the budget turns out to be buying the former while assuming it’s buying the latter.

None of this means abandoning structured learning. Some basic fluency (what the buttons do, where the data goes, what not to paste into a public model) is worth teaching directly, and quickly. But that’s the floor, not the building. The organisations that come out ahead in this transition won’t be the ones with the most certified staff. They’ll be the ones that gave their people the time, the tools, and the permission to get their hands dirty — and trusted that the skills that matter would grow where they’ve always grown, in the doing.