Every published figure below was read from the supplier’s own page on 21 September 2026 and is linked in Sources. The method is one you can run internally, and the last two sections say when you should not run it at all.

At a glance

  • £30,000 to £500,000 is the published UK range for producing an AI plan. Ours is a deliberately narrower fourteen-day version at £2,000.
  • A roadmap is a ranked list of decisions with numbers and owners, not a list of tools with quarters beside them.
  • Four steps: map the constraint, price every option including the rejects, rank by economic case, name an owner.
  • The one section that proves a roadmap is real is the rejected list.
  • In the largest published study of the gap, only 5 per cent of task-specific AI tools reached production. Planning is where most of them are lost.

What is an AI roadmap and what is a tools list?

A roadmap is a sequenced set of commercial decisions. Each item says what changes, what it costs, what it returns, who owns it, and what has to be true before it starts. Read out loud, it sounds like a capital plan.

A tools list is the other document, and it is far more common. It names products and assigns them to quarters. Copilot in Q1, a chatbot in Q2, “agents” in Q3. It contains no numbers that anyone measured and no names against any line. Graph.digital calls this pattern the tools-list roadmap and puts it first in its four failure modes of AI planning, which is a fair call.

The test is simple. Cover the product names on your document. If what remains still tells you what the business will do differently and what that is worth, you have a roadmap. If the page goes blank, you have a shopping list.

Why are most roadmaps written before anyone has priced the constraint?

Because the pressure to have a plan arrives before the evidence does. A board asks what the AI plan is, and the honest answer, which is that nobody has measured which process is costing the most, is not a comfortable thing to say in that meeting. So a plan gets written from what is available: vendor briefings, a competitor’s press release, and whichever department shouted loudest.

The result is a document ranked by enthusiasm. It is not wrong so much as unfalsifiable, because no item on it carries a number that could turn out to be untrue. And it is expensive in a specific way: the first project gets funded, it is the one somebody was excited about rather than the one that was costing the most, and when it delivers little the conclusion drawn is that AI does not work here.

Pricing the constraint first inverts that. You cannot rank options you have not costed, and you cannot cost an option without reading the process it changes.

Step 1: map where capacity and cash are held up

Start with where work waits, not with where AI is interesting. In most operations of £1m to £50m turnover, the constraint is one of four things: a queue in front of a single person, a hand-off between two systems that do not talk, a document that has to be produced by hand, or a decision that only one person is allowed to make.

Walk the operation and write down every place work sits still. For each, note how long it waits, how often, and what it is waiting for. Do not attempt a solution yet. The output of this step is a list of places money is held up, in order of how much is held up there.

The discipline is to keep the list about the business rather than about technology. If an item on it names a product, it belongs in step two.

Step 2: price each opportunity, including the ones you will reject

For every candidate, write three numbers: what the current way costs each year, what changing it would cost to build, and what it would cost to run afterwards. Use a loaded hourly rate and state it. A £28,000 salary is roughly £18 an hour loaded, and being ten per cent out will not change a ranking.

Then price the ones you already know you will not do. This is the step that gets skipped and it is the step that makes everything after it credible. A ranking where only the winners were costed is not a ranking, it is a preference with arithmetic attached to it after the fact. When a finance director asks why the invoicing project beat the customer-service project, the answer has to be two numbers, not one number and an opinion.

Mark each item with its data state, honestly. One clean export is a morning. Five systems that half agree with each other is the project.

Step 3: rank by economic case, not by enthusiasm

Sort by annual value released divided by total first-year cost, then apply two filters that have nothing to do with the arithmetic.

Reversibility. Prefer the item you could stop doing in a fortnight without consequence. Early in a programme, the option to change your mind is worth more than a marginally better return.

Evidence value. Prefer the item that will produce a number you can show the board at the end. A project that saves slightly less but proves the method is worth more than a larger one that finishes as an anecdote.

The top item should be something narrow enough to finish and visible enough to be believed. If your first item is a platform, the ranking has gone wrong.

Step 4: who owns each item?

One named person per item. Not a department, not a steering group, not “IT”. The owner is whoever can change how the work is done and will be asked about it in six months.

If no owner can be named for an item, take it off the roadmap. That is not bureaucracy; it is the single most reliable predictor in this whole exercise. MIT’s Project NANDA report puts misalignment with day-to-day operations among the main reasons tools stall between pilot and production, and an unowned item is that failure written down in advance.

The owner’s first job is to agree what “working” looks like before anything is built, in a sentence with a number in it. If success is not measurable, neither is completion.

What does an AI roadmap cost in the UK?

Published figures, all read today. Helium42 puts a focused single-use-case pilot at £30,000 to £80,000 including consultancy, training and tooling, and a full enterprise programme at £100,000 to £500,000. Graph.digital says a commercially grounded twelve-month roadmap with keep, kill and scale decisions and a governance structure can be in place within eight to ten weeks, and quotes capital freed by its portfolio review at £80,000 to £200,000, explicitly labelled as illustrative.

Two things are worth noticing about that market. The first is that most of the cost is the reading, not the writing: the document is short, and the weeks go into understanding the operation well enough to be specific. The second is that several suppliers publish costs for other people’s projects and no cost for their own work.

Ours is published. The fit call is free. The fourteen-day AI Constraint Audit is a fixed £2,000, and the ranked next step it produces is a roadmap for the processes it looked at. Details, including what it excludes, are on the AI Constraint Audit page.

What does a fourteen-day version leave out?

It is narrower on purpose, and you should know the edges before you buy it.

It covers the processes we can read in three sixty-minute working sessions plus the evidence you provide. It does not attempt an organisation-wide portfolio review of every department. It does not produce a governance framework, a target operating model, or a change programme. It does not include vendor selection for products we are not building. And it looks twelve months ahead at most, because anything further is a guess dressed as a plan.

What it does produce is four documents you own outright, written so that another supplier could quote from them. That portability is the test. If a roadmap only makes sense to the firm that wrote it, it was a sales document.

What does a finished roadmap look like?

Five sections, and none of them is long.

A constraint map, showing where capacity and cash are held up today, with the measured numbers beside each. An opportunity map, listing every candidate, including the ones being rejected. Economic cases, one per candidate, each with its assumptions written beside it so they can be argued with separately from the conclusion. A ranked next step, with a named owner and a definition of working. And a rejected list, with the reason each item was set aside.

Four competitor pages we read describe a roadmap in detail. None of them publishes one. That gap is worth noticing when you are choosing who writes yours: ask to see a structurally real example, redacted as heavily as they like, before you commission a document you have never seen the shape of.

When should you not build a roadmap?

When you have one process that is obviously the problem and everyone already agrees on it. Go and fix that. A roadmap to justify a decision already made is a delay with a cover sheet.

When nobody will be given time to execute it. A plan without capacity behind it produces a document and a quiet year.

When the business is mid-restructure or mid-acquisition and the process map will be wrong in three months. Wait until the shape settles, or map only the part that will not move.

And when the honest total of everything you could save is smaller than the cost of the exercise. That is arithmetic, and it is the first thing we check.

What this does not apply to

This is written for UK businesses of roughly £1m to £50m turnover: an operations director at a 40-person distributor, a contracts manager at a construction firm, a finance lead at a clinic group. It assumes a single management layer that can make a decision without a committee.

It does not apply to enterprises with formal portfolio governance, where the roadmap has to enter an existing process rather than replace it. It does not cover public-sector business cases, which follow HM Treasury’s Five Case Model and its own approval gates. It does not cover research programmes, where the output is knowledge rather than a working system. And it is not a governance or compliance framework: where AI outputs affect individuals, the ICO’s guidance on AI and data protection applies and sits outside this method.

When we are the wrong choice

If you want a document to present rather than a sequence to execute, somebody else will write you a better-looking one.

If the roadmap has to conclude that the project already chosen is the right first move, do not commission a ranking. The ranking sometimes disagrees.

If you need an organisation-wide portfolio review across many departments, our fourteen days is the wrong shape and we will say so on the call.

If there is no one who can be named as an owner for anything, the constraint is not AI readiness. It is authority, and no supplier can sell you that.

Sources

Read on 21 September 2026.

The short answers

Each answer is the opening of the section it links to, so nothing here is written for a crawler that a reader cannot also see.

What is an AI roadmap, and how is it different from a tools list?

A roadmap is a sequenced set of commercial decisions. Each item says what changes, what it costs, what it returns, who owns it, and what has to be true before it starts. Read out loud, it sounds like a capital plan.

Why are most AI roadmaps written before anyone prices the constraint?

Because the pressure to have a plan arrives before the evidence does. A board asks what the AI plan is, and the honest answer, which is that nobody has measured which process is costing the most, is not a comfortable thing to say in that meeting. So a plan gets written from what is available: vendor briefings, a competitor's press release, and whichever department shouted loudest.

What does an AI roadmap cost in the UK?

Published figures, all read today. Helium42 puts a focused single-use-case pilot at £30,000 to £80,000 including consultancy, training and tooling, and a full enterprise programme at £100,000 to £500,000. Graph.digital says a commercially grounded twelve-month roadmap with keep, kill and scale decisions and a governance structure can be in place within eight to ten weeks, and quotes capital freed by its portfolio review at £80,000 to £200,000, explicitly labelled as illustrative.

Who should own each item on an AI roadmap?

One named person per item. Not a department, not a steering group, not "IT". The owner is whoever can change how the work is done and will be asked about it in six months.

What does a finished AI roadmap actually look like?

Five sections, and none of them is long.

When should a business not build an AI roadmap at all?

When you have one process that is obviously the problem and everyone already agrees on it. Go and fix that. A roadmap to justify a decision already made is a delay with a cover sheet.

How long does it take to build an AI roadmap?

Published UK engagements describe eight to ten weeks for a full portfolio review and twelve-month plan. A narrower version that ranks and prices a handful of processes can be done in two weeks. Ours is fourteen days, with three sixty-minute working sessions inside it.

What is the difference between an AI roadmap and an AI strategy?

A strategy says where you are trying to get to and why. A roadmap says what you are doing first, what it costs, who owns it and what you are deliberately not doing. If the document has no rejected list and no numbers, it is a strategy that has been labelled as a roadmap.

Can we build an AI roadmap without external help?

Yes, and the four steps on this page are the whole method. Outside help earns its fee in two situations: when the ranking has to survive a finance director who does not care about AI, and when no one internally can be neutral about their own department's pet project.

What does AGMM charge to produce one?

The 30-minute fit call is free. The AI Constraint Audit is fixed at £2,000 over fourteen days and ends in four documents you own: a constraint map, an AI opportunity map, economic cases and a ranked next step. That last document is a roadmap by another name.

What if the roadmap says do nothing this year?

Then you have saved a year of spend and you keep the four documents. It is an uncomfortable result to present and it is a legitimate one. Our audit fee is not credited against a build precisely so that this answer costs us the same as any other.

How far ahead should an AI roadmap look?

Twelve months for the sequence, and no further than one quarter for anything with a fixed price against it. The model and tool market moves faster than an eighteen-month plan can absorb, so anything beyond the first two items should be a ranked intention rather than a commitment.

Should the roadmap be owned by IT or by operations?

By whoever owns the process being changed, with IT as a named dependency. Roadmaps owned by IT tend to become infrastructure plans, and roadmaps owned by a steering group with no operational authority tend not to be executed at all.

Do we need a data strategy before an AI roadmap?

No, but the roadmap has to be honest about data state per item. Step two prices data preparation for each opportunity separately, which is usually enough. A standalone data strategy before any use case is known is how organisations spend a year and produce a document.