Forecasts break when the rules change. Foresight is built for that.

A forecast extends the present. It takes the trends visible today and carries them forward, on the quiet assumption that prices, institutions and rules stay roughly where they are. For stable processes that works well. It fails at the moment a decision matters most, which is when the rules themselves change.

Tariffs show the pattern clearly. Their effects on global supply chains and currency tend to last well beyond the headline rate, because firms respond by re-routing sourcing, renegotiating contracts and repricing, and each of those adjustments has its own inertia. A forecast built on the previous tariff regime is wrong twice: once when the tariff arrives, and again when it is withdrawn and the supply chain does not snap back.

AI planning has the same structure. Model capability, compute cost, regulation and the labour market are all moving, and most of them move independently of your organisation. A plan that rests on one number for adoption or savings three years out is a guess presented as a projection. Strategic foresight does something different. It lists the drivers, separates the ones that are predictable from the ones that are genuinely uncertain, and builds a small set of scenarios that differ on the uncertain drivers only. The question then changes from what will happen to what we would do in each case.

The useful output is short. A driver list ranked by impact and uncertainty. Three or four scenarios, each described in a paragraph. And a set of signposts, each with an owner and a trigger, that tell you which scenario is unfolding. From these a leadership team can sort its AI decisions into three groups: moves that hold up in every scenario and should start now, options worth keeping open at a small cost, and bets that only pay in one scenario and deserve a harder look.

The same logic applies to institutions. Governments and development bodies planning AI for national development are committing to long-lived assets, such as skills, data infrastructure and policy, under external conditions they do not control. A foresight exercise that tests those commitments against several plausible futures is cheaper than discovering, five years in, that the plan only worked in one of them.

It matters for governance too. Oversight designed around today's systems has to cope with tomorrow's, as autonomous systems take on more of the decisions people used to make. Foresight asks what a control looks like when the system is considerably more capable than the one it was written for. A board can start with two questions: which assumption in our AI plan would hurt most if it were wrong, and what would tell us early?