Why Most National AI Strategies Stall in Year Two

The predictable arc

National AI strategies follow a recognisable trajectory. Year one produces a document, a coordinating body and genuine enthusiasm. Year two produces pilots, some of which work. Year three produces a review explaining why scaling has not occurred, usually attributing it to funding or talent.

The attribution is generally wrong. Funding is rarely the binding constraint at this stage, and talent shortages are a symptom rather than a cause. What actually failed was that the strategy specified outcomes without specifying who would deliver them, with what authority, and against what accountability.

Three structural gaps

The first is the pilot trap. Pilots are designed to demonstrate feasibility and are therefore run under favourable conditions — a motivated team, a clean dataset, a supportive department. Scaling requires the opposite conditions, and nothing in a successful pilot demonstrates that scaling will work. Organisations that treat pilot success as evidence of scalability are misreading their own data.

The second is procurement. Public procurement is optimised for specifiable deliverables and price competition. AI systems require iteration, and the vendor cannot honestly specify the outcome in advance. Procurement rules designed for construction produce contracts that make good AI delivery structurally impossible.

The third is data. Strategies assume data availability that does not exist. The data is held in incompatible systems, of unknown quality, with unresolved sharing authority between departments. This is a two-to-three year remediation programme that almost no strategy funds, because it is invisible and produces nothing announceable.

What the successful programmes did differently

The jurisdictions that moved past year two share several features.

They built a delivery unit with real authority — the ability to direct departmental resources rather than merely convene meetings. They reformed procurement specifically for iterative work before attempting large projects. They funded data remediation as a named programme with its own budget line rather than as a component of something more attractive. And they defined a small number of measurable outcomes rather than a comprehensive framework, accepting narrower scope in exchange for actual delivery.

The pattern is consistent: they treated delivery capacity as the thing being built, and the AI applications as what that capacity would subsequently produce.

A test worth applying

Before publishing a strategy, ask a specific question of each commitment: who exactly will do this, what authority do they have over the people whose cooperation they need, and what happens if they fail?

Commitments that cannot survive this question are aspirations. Including them is not harmless, because they consume the credibility that the deliverable commitments depend on. A shorter strategy that is believed outperforms a comprehensive one that is not.