AI strategy: why you write it last

The claim this piece rests on

In Gartner's maturity model (2025), strategy heads the seven dimensions. Our experience taking AI to production in mid-sized companies leads us to place it at the end of the starting sequence. It is worth explaining why, because this is not a minor disagreement.

An AI strategy written before you have an instrumented case rests on somebody else's assumptions. The savings are sector benchmarks. The priorities come from what gets read, not from where your bottleneck sits. The timelines are whatever the vendor considers reasonable. The resulting document is coherent, presentable, and changes no decision, because it contains no information the company did not already have.

After one measured case, the same strategy fits on two pages and carries your own numbers: what one of your processes really costs, what can be recovered, where the real obstacle was — almost never the one assumed — and what kind of work pays off in your house. That does change decisions.

The caveat that cannot be omitted: this sequence is for mid-sized companies without severe sector supervision. In banking, life and health insurance or healthcare, strategy and governance move to the front, because the cost of an error is not a lost pilot but a regulatory file, and risk classification constrains what can even be piloted. Saying so does not weaken the thesis: it bounds it.

So why does it get written so early?

Because there is pressure, and the pressure is legitimate.

A board asks what the company is doing about AI, and "we are measuring one case" sounds thin against a three-year plan with a diagram. A competitor announces an initiative and the comparison is uncomfortable. A vendor offers to write the strategy for free as a way in, which solves the calendar problem in exchange for a strategy that reflects their catalogue.

Our practical recommendation is not to resist that pressure but to change the deliverable. What goes to the board in the first quarter is not a strategy, it is a hypothesis with a verification date. One case, one baseline, one decision date, and the written criterion for continuing or stopping. It is shorter, harder to fabricate, and demonstrates more judgement than any three-year plan.

What a useful AI strategy contains

When the moment comes — after the first case — five elements and no more:

1. Where your company competes. Which processes differentiate and which are commodity. It determines what gets built and what gets bought, and it carries the largest economic consequences of the five.
2. What we are not going to do. The list of explicit exclusions, with reasons. It is the section that resists being written and the one that saves the most time later.
3. How much, and from where. Budget and which line it comes from. A strategy without a source of funds is a wish list.
4. What stays in-house. Which knowledge, context and capability cannot live at a vendor. With AI this decision weighs more than with any previous technology, because the asset that accumulates — your company's context — is exactly what differentiates you.
5. How we will know it is working. Two or three indicators, already measured, not aspirational.

If the document runs past five or six pages, it is almost always because it includes an explainer on what AI is. That is an annex, not a strategy.

Diagnosis: where you are

Stage 1 · Intent. There is declared will and perhaps a document, but no decision has changed because of it. Observable signal: nobody has said no to anything citing the strategy.

Stage 2 · Strategy with evidence. At least one measured case exists and the document carries your own numbers. Observable signal: there are explicit exclusions, and someone has used them to reject a proposal.

Stage 3 · Living strategy. It is reviewed on a cadence against what the portfolio is teaching, and it has changed at least once on evidence. Observable signal: the current version is not the first, and it is known what changed it.

The first 90 days, if you already have a measured case

Weeks 1-2 · Harvest the learning from the case. Not just the result: what turned out differently from expected. That delta is the most valuable material you hold and it usually evaporates if not written down.

Weeks 3-4 · Differentiating/commodity map of candidate processes. This is a business conversation, not a technology one, and it is best led by someone who knows the customers.

Weeks 5-8 · Exclusions. Choose what will not be done in the next twelve months, and write the reason. If this section stays empty, the exercise has not happened.

Weeks 9-12 · Document and review cadence. Five pages and a date to look at it again.

What NOT to do

• Do not commission the strategy from the vendor who will execute it. Even if it is free. Especially if it is free.
• Do not write it for three years. The honest horizon today is twelve months with a half-yearly review.
• Do not copy a sector leader's strategy. Their bottleneck is not yours, and the part that matters is precisely the part they do not publish.
• Do not confuse strategy with a tool roadmap. A list of technologies sorted by quarter is a purchasing plan.

What writing it too early costs

It costs the programme's credibility. A strategy with figures that do not materialise — because they were sector benchmarks, not yours — burns the sponsor's political capital. And it costs opportunity: the quarter spent drafting the document is the quarter not spent producing the case that would have made it useful.

There is an additional, less visible cost: a premature strategy fixes the wrong priorities and makes them hard to move, because they are already approved. We have seen companies defend a use case for a year against their own evidence, because it was in the plan presented to the board.

Seventh and final layer of the series on the seven layers of AI maturity.

Frequently asked questions

When should you write an AI strategy?

After the first measured case. Before that, the document rests on sector benchmarks and changes no decision; after it, it fits on two pages and carries your own numbers. The exception is heavily supervised sectors — banking, life and health insurance, healthcare — where risk classification constrains what can be piloted and strategy moves to the front.

What should an AI strategy include?

Five things: which processes differentiate and which are commodity, what is explicitly excluded and why, how much is invested and from which budget line, what knowledge stays in-house, and two or three already-measured indicators. If it runs beyond five pages, it usually contains an explainer on what AI is.

Can the vendor write our AI strategy?

They can, and it is better not to let them. A strategy written by whoever will execute it tends to reflect their catalogue. The risk is higher when the service is free, because then the vendor's return sits precisely in what the strategy recommends.

How many years out should AI be planned?

The honest horizon today is twelve months with a half-yearly review. Technology and costs move too fast to commit decisions three years out, and a three-year plan in this area is usually a presentation document rather than a decision tool.