We want to start with AI — what do we do first?

Not a tool evaluation. Not a company-wide training day. Not a policy document.

Pick one decision that a named person already owns, and improve it end to end. One decision, one owner, one number, six to eight weeks.

Why not start with a policy?

Because you do not yet know what you are governing. Policies written before any real use describe an imagined system, and they age badly the moment the first team does something the author did not anticipate. Write the policy after the first pilot, when you have concrete decisions to govern — you will write a much better one, in less time.

The exception: if you operate in a regulated sector or handle special-category personal data, get the data-protection position settled before anything touches production. That is not a policy exercise, it is a precondition.

Why not start with training?

Because enthusiasm without a use case decays in about three weeks. Train the team that is running the pilot, on the pilot. Broad training lands far better once people have seen a colleague's work change.

What "end to end" means

The pilot is not done when the model gives good answers in a test. It is done when the decision is actually being made differently, by the person who owns it, in their normal workflow — and someone can show the number moved.

That last step is where most pilots quietly stop. Budget for it explicitly.

A realistic first eight weeks

Weeks	What happens
1	Pick the decision and the owner. Write down the baseline and the counterfactual.
2–3	Build the narrowest thing that could work. Resist scope.
4–6	The owner uses it for real. Fix what breaks. Keep a log of failures.
7–8	Measure against the baseline. Decide: extend, change, or stop.

"Stop" must be a permitted outcome, and must be said out loud at the start. A programme where no pilot is ever allowed to fail produces pilots that are never allowed to be measured.