You Can’t Predict AI. You Can Still Prepare For It.

It’s a fair challenge, and it deserves a straight answer rather than a dodge. The honest response isn’t to pretend we can see where AI is going. It’s to notice that the objection quietly assumes training has to predict the technology. It doesn’t — and the best of it never has.

What dates, and what doesn’t

Any simulation, course, or programme about working with AI is really made of two different materials, and they age at completely different rates.

The first layer is volatile. It’s the specifics: what a given tool can do this quarter, the cost of a particular model, the claim that “AI now handles X% of some task.” This layer is genuinely perishable. Build the core of your training around it and yes, it dates in months — a “chatbot scenario” written in 2024 already reads like a period piece.

The second layer is invariant. It’s the organisational trade-offs that sit underneath whatever the technology happens to be this year: how fast to adopt versus how fast people can actually absorb the change; how much to delegate to the machine versus how much to oversee; the cost of trusting an output weighed against the cost of verifying it; what you owe the people whose roles are being rewritten. None of these tensions are new. They showed up with every major technology transition before this one. AI doesn’t introduce them — it sharpens them.

Good training about AI keeps its mechanics keyed to that second layer and confines the volatile specifics to the surface, where they can be swapped out cheaply. Get that division right and the shelf-life problem largely dissolves. You’re not teaching people what AI will do. You’re teaching them how to decide well when they can’t be sure — which is the one capability that survives every version bump.

The tensions intensify; they don’t expire

Here’s the part that turns the objection on its head. The faster AI advances, the more those invariant tensions matter — not less.

When an AI assistant was a novelty, “should we trust this output?” was a low-stakes question. As the same systems move from drafting emails to taking actions inside real workflows, that question becomes the whole game: what may it do without a human in the loop, who checks it, and who answers when it gets something confidently wrong? The uncertainty everyone is anxious about isn’t a reason to wait for the picture to settle. It’s the precise thing worth rehearsing — because it isn’t going to settle.

So “we can’t predict AI” is not an argument against training. Properly understood, it’s the argument for a particular kind of training: the kind that builds decision-making under irreducible uncertainty, rather than knowledge about a snapshot that’s already moving.

A concrete example

This isn’t abstract. In one simulation we built, teams deliver a project for a demanding client with the help of an AI teammate that is — by design — powerful, fast, and occasionally confidently wrong. They can hand work to it or keep it human. The learning sits entirely in the calibration: when to trust it, when to verify, when to override.

We pressure-tested the finished design by having AI agents play it many times over under different strategies, and scoring each one. The strategy that trusted the AI blindly finished bottom. The strategy that refused to use it at all never won either. The strategy that checked, questioned, and calibrated — a human kept firmly in the loop — topped the board. Nobody wrote that lesson in as a message; it fell out of the mechanics.

Notice what’s volatile and what’s invariant there. The scenario dressing — the client, the deadline, the specific things the AI is good and bad at — can be re-skinned to any organisation and refreshed as the technology moves. The thing being trained — the judgement to direct a fallible machine well — doesn’t date. That’s the two-layer discipline in a single build.

The trap this sets for off-the-shelf

There’s a competitive edge hiding in here, and it runs the opposite way to the usual assumption. On most topics an off-the-shelf product has the advantage: it’s cheaper because the cost of building it is spread across many buyers. On AI specifically, that model works against it. A productised AI simulation is frozen at the moment it was made, and its volatile layer starts decaying from day one. Refreshing it for every buyer is exactly the economy of scale it can’t afford to surrender.

Bespoke inverts that. When each build rebuilds the volatile layer from the client’s present reality — in days rather than the months a traditional build takes — the perishable content is always current, and the invariant mechanics underneath are where the real design rigour goes, behind more than 25 quality gates. On this one theme, bespoke isn’t just a better fit. It’s the only version that stays true.

What to actually prepare for

So the practical answer to “how do we train for an AI future we can’t predict” is to stop trying to predict it. Prepare your people for the decisions the future will keep asking of them, whatever the tools turn out to be: when to trust the machine and when to check it; how to keep human judgement sharp enough to catch a confident error; how to weigh speed against oversight when both genuinely matter. Those questions won’t be answered once. Your people will face versions of them for the rest of their careers.

That capability isn’t built by a briefing on this month’s tools. It’s built the way judgement is always built — by making consequential decisions, living with what follows, and having to account for them. Safely, and repeatedly, before the stakes are real.

The durable bet

You can’t forecast AI, and you don’t need to. The bet worth making is that the value will keep moving toward the people who can decide well with a machine in the loop — and that this is a skill, not a briefing. Train the invariant, re-skin the volatile, and the ground stops shifting under you.

The organisations that work this out will waste less energy chasing the technology’s tail, and spend it instead on the thing that compounds: people who can meet whatever AI becomes with judgement that’s already been practised.