AI Strategy & Roadmap Is the Cheapest Part of an AI Project

A room agrees to do AI. There is a slide with six ideas on it, and everyone present likes at least one of them. Budget is approved in principle. The meeting ends early, which everybody reads as a good sign.

Nothing was decided in that room.

Six ideas and a budget is not a plan. It is six plans, none of them costed, competing for the same engineers — and the decision about which one actually happens has quietly been handed to whoever picks up the work first. The distance between that slide and a roadmap is where most AI money goes.

Anyone can produce the list. The list was never the hard part.

Give any capable team an afternoon and they will hand you twenty candidate AI projects. Ideation is free and always has been. The work is selection and sequencing — deciding which of the twenty happens first, which happens never, and what each one costs before anyone commits to it.

Skip that, and the project does not fail loudly. It fails the way Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 — and the reasons Gartner gives are worth reading slowly: escalating costs, unclear business value, inadequate risk controls.

Read them again. Not one of those is a discovery you have to build something to make. Each is a question with an available answer, asked too late by organisations that started building because building felt like progress.

A pilot's cost tells you almost nothing about the bill

The most expensive assumption in AI planning is that the demo scales linearly. It does not.

Gartner expects at least half of GenAI projects to exceed their budgets by 2028, blaming poor architectural choices and thin operational experience, and its researchers put the gap bluntly: a production-ready GenAI system can be orders of magnitude more expensive than the pilot that justified it.

There is a structural reason software companies get caught by this specifically. Per-seat pricing was built on an assumption that has held for thirty years — serving one more user costs approximately nothing. AI breaks that assumption, because every call costs real money in compute, and the cost scales with the thing you were hoping would grow. Gartner now expects AI coding costs to overtake the average developer's salary by 2028 as token consumption climbs and licensing shifts to consumption — and notes that without a governed operating model, those costs outrun the productivity gains the tools were bought for.

So the question a pilot answers — does this work? — is the less important one. The question that decides whether you should fund it is: does it still work at a hundred times the volume, at your margin, at your price? That number does not appear on its own. Somebody has to sit down with your volumes and produce it.

An order is a decision. A list is a deferral.

Sequencing sounds like project management. It is closer to risk control.

S&P Global's survey of more than 1,000 companies found the average organisation scrapped 46% of its AI proofs of concept before they reached production, and that the share of companies abandoning most of their AI initiatives had jumped to 42%, from 17% a year earlier. The obstacles they named were cost, data privacy and security.

Every one of those is knowable before a line of code exists. They are not discoveries. They are the questions a sequencing exercise asks on purpose, early, while the answer is still free.

And ordering does real work that a list cannot. Dependencies mean the slowest item usually has to start first — data access, owner sign-off and legal review do not compress, whatever the engineering does. The project that unblocks three others outranks the one with the better demo, even though the demo is what the room remembers. A list has no way to express any of that. It just sits there, implying everything is equally urgent, until someone picks the fun one.

What has to be in it before it counts as a roadmap

A document with quarters across the top and project names inside it is a wall chart, not a roadmap. The difference is whether the thing can be proved wrong. A real one carries, for each phase:

• A verdict — build, buy, or wait. Wait has to be a permitted answer, or the exercise is theatre.
• A cost calculated on your volumes, not a vendor's benchmark: cost per action, and what that does to the margin on the product it sits inside.
• A payback point. Not "efficiency gains" — the month the phase stops costing more than it returns, and what has to be true for that to land.
• A metric agreed before launch, so "is it working?" has an answer that is not a demo in a sprint review.
• A risk assessment attached to the phase itself, because the cheapest place to find out something is unworkable is a document.

This is not an invented checklist. When McKinsey looked at what separates organisations getting bottom-line value from AI from those merely using it, two of the practices most tied to adoption and return were establishing a clearly defined road map with phased rollouts, and tracking well-defined KPIs for what gets built. Fewer than one in five organisations were doing the second one at all.

That is a striking place to find an advantage — not in a better model, which your competitors can license on the same afternoon you can, but in having written down what success would look like before starting.

Changing your mind is cheap exactly once

Here is the argument for doing this work first, reduced to its smallest form.

Moving a phase in a planning document costs a meeting. Moving it after the build costs the build. Every figure in this piece — the cancelled projects, the scrapped proofs of concept, the budgets overrun by an order of magnitude — describes money committed before anyone had established what the thing would cost at scale, what it was worth, or which piece needed to happen first.

A roadmap is not the paperwork you complete before the interesting part starts. It is the one stage of an AI project where no and not yet are still inexpensive answers — and where, if the honest verdict on a project is that it should not be built this year, that verdict costs you a document instead of a year.