The tool is purchased, the pilot leads to a green dashboard, and eighteen months later the same team is back on the spreadsheet with a subscription still on the books.
No part of that is a model failure. It is a set of missing decisions, and there are usually four of them: who is responsible for the outcome, which number is supposed to move, what the system is allowed to do, and who acts when it is wrong. An AI implementation roadmap is the document where those four decisions live, dated and signed. It forecasts no outcomes. What it does is remove the ambiguity that lets a bad pilot survive past the point where anyone would defend it in an honest meeting.
Three examples, all of them ordinary.
A UK distributor pilots demand forecasting. It beats the spreadsheet by a wide margin in the pilot. The forecast never lands in the system the planners place orders from, so twelve months later the software subscription renews and the planners still use the spreadsheet.
A professional services firm buys 400 assistant seats. Usage spikes in week two. At the next quarterly review no person can say which of the 400 users does anything differently, or how the firm would find out.
An insurer ships a claims triage model. No one specified who retrains it when the mix of claims shifts. Accuracy decays quietly, and the first person to notice is a handler with a queue.
Four documents carry the name, and one of them schedules decisions
Four documents go by similar titles, and only the second one orders decisions.
| The document | What it decides | What it holds at day 90 | Who signs it | What it does when the model changes |
|---|---|---|---|---|
| AI strategy | where the organisation intends to compete | themes, ambitions, a narrative for the board | chief executive or board | nothing, because it was never built to execute |
| AI implementation roadmap | which decisions get made, in what order, by whom | a dated register of decisions, owners, gates and kill criteria | the executive who owns the outcome | triggers the retrain, re-approval or retirement decision |
| Acceptable use policy | what staff may put into a tool, and what they may take out | a signed acknowledgement per user | head of risk or legal | nothing, unless the roadmap points at it |
| Training plan | who can operate the tool | attendance records | head of people | nothing |
| Vendor contract | what the supplier owes you | service levels and exit terms | procurement | restarts the commercial clock and leaves the outcome untouched |
If your current document has no dated register with identified proprietors and kill criteria, you have the preamble. Write the register first, then the deck around it.
What an approved roadmap hands over
Seven things come out of it. Each one is a sentence someone can be held to.
- A named business owner for each outcome, holding the authority to stop the work.
- A baseline measurement of the number that is meant to move, taken from the source system before anything is built.
- A written definition of what the organisation counts as AI use, and the name of the person who signed that definition.
- Kill criteria for each phase, agreed before the phase begins.
- A retraining trigger: the metric, the threshold, and the person who acts on it.
- A costed exit, covering what it takes to switch off, migrate or revert.
- A renewal decision that sits in the same diary as the outcome review.
All other items in a typical roadmap deck, including the architecture diagram and the vendor shortlist, are supporting material for that list. The vendors selling the tooling make the strongest version of this case themselves. Microsoft's AI Strategy Roadmap states plainly that no single roadmap fits every organisation and lists five drivers of AI value: business strategy, technology and data strategy, AI strategy and experience, organisation and culture, and AI governance. Read those five as a checklist and the deck stops being a presentation.
Checks to run before a single line is written
The document is cheap to write and expensive to write early. These are the conditions that must be met first.
- You can name the one business number the first use case is meant to move, and say what that number is today.
- The person with that number has seen the draft and will sign it.
- You have written down what your organisation counts as AI use, and named the person who signs that definition off.
- You know which regulator, if any, has a stated interest in the output.
- You have a list of every AI tool already running, including the ones bought on a team credit card.
- You have decided whether this is a portfolio document or a single initiative document, and you have told people which.
- Someone outside the delivery team has read your success criteria and tried to break them.
- You have agreed what stop looks like, in writing, before there is anything to stop.
- You know who chairs the first gate and who makes the decision.
Checks to run before the first pilot is named
A pilot is a decision to spend money on evidence. These are the conditions that make the evidence worth having.
- The process the pilot touches has a written description, and that description existed before the pilot did.
- The baseline comes from the system of record, with a date on it. If the team cannot produce last quarter's numbers from the system, the pilot will produce no proof.
- The pilot has a kill date in the calendar: on that date the work stops unless a named person renews it.
- Any data the pilot accesses has a specified person who has the right to it, and that person has confirmed the legal basis for this use.
- The integration point is identified: which system of record receives the output, and what happens if it never arrives.
- You have written down what the model may do and what it may refuse to do, in sentences a non-specialist can check.
- The running cost after the pilot is estimated and budgeted. Pilot pricing and production pricing are different numbers.
- The team that will operate the thing in production has been in the room since the first workshop.
Phase gates, and the statements that clear them
Gates are where the roadmap earns its keep, because they are the only points at which stopping is cheap. The table below sets the checks against what each one protects.
| When the check runs | What it protects | What a fail costs | Who can clear it |
|---|---|---|---|
| Before writing | the credibility of the document | months of work with no decision attached | the executive sponsor |
| Before pilot naming | the pilot's right to exist | a pilot that proves capability and nothing else | the process owner |
| At each phase gate | the budget | the next phase, spent on the wrong thing | the steering group |
| Weekly while live | the evidence | a surprise at the quarterly review | the delivery lead |
| After go-live | the outcome | a licence that renews itself | the business owner |
| Annually | the document itself | a roadmap that describes last year's estate | the sponsor |
At the gate itself, six statements have to be true.
- The previous gate's kill criteria were tested against the live system.
- The business owner confirms the number moved, or explains in writing why it did not.
- The support model is written: who is paged, at what hour, with what authority to switch the system off.
- The retraining trigger is defined as a metric, a threshold and a named person.
- The exit is costed, and the cost sits in the current financial year.
- Governance has been exercised once for real. A live decision has already gone through it.
The NIST AI Risk Management Framework Roadmap is useful here for one specific reason. The framework is voluntary, rights-preserving, non-sector specific and use-case agnostic, so a gate designed around it works across business units without a rewrite. The same Roadmap identifies AI RMF profiles as a primary method for organisations to show how they applied the framework, so a written profile of your own gate becomes evidence an auditor or a supplier can examine.
Weekly checks while a pilot is live
Weekly checks are short and boring by design. Their job is to stop a quiet failure from reaching the quarterly review undiscovered.
- The output reaches the system of record at the agreed rate.
- A human has overridden the model at least once, and the override is logged.
- The person who would notice a slow failure has looked at the output this week.
- Usage is measured by tasks completed.
- Cost per transaction is tracked, including inference and human review time.
- Questions from the front line reach someone who can change the model inside a defined window.
The vendor guides agree about the shape of this. Splunk's roadmap guide says that model drift is a major long-term challenge, as accuracy drops when real-world data and user behaviour change below a model, so you must assign a date and an owner for retraining and recalibration. The same guide reports that 78 per cent of organisations had implemented AI in some form by 2024, and that 92 per cent planned to expand AI investment over the following three years. That second number is the pressure that creates unowned pilots in the first place.
Ninety days after go-live
Ninety days is the first point at which the outcome can be judged separately from the launch. Five questions get answered in writing, by the business owner.
- Does the number the use case was meant to move differ from the baseline, and by how much?
- Which steps in the process changed? A dashboard change is not a process change.
- Who is still doing work the model was supposed to remove, and why?
- What has the system been asked to do that no one planned for?
- Has the retraining trigger been checked, and what did it read?
This is also where UK public sector guidance is worth borrowing from inside a private company. The UK Government Analysis Function publishes a collated list of approved AI guidance, best practice and standards that prioritises the Data and AI Ethics Framework and the Artificial Intelligence Playbook for the UK Government, and you should check departmental guidance first and consult IT or digital teams when any point is uncertain. The pattern transfers: one central framework, team-level depth beneath it, and a named place to look when the two disagree. An AI implementation roadmap without that hierarchy ends up with three policies, two owners and no decision.
Renewal, and the two numbers that decide it
Renewal is the cheapest gate and the one most often skipped. Two numbers decide it: the movement in the business number against the baseline, and the fully loaded cost of running the system for the year just finished.
- The business owner can state the movement in the number without opening a dashboard.
- The fully loaded cost includes inference, human review, integration maintenance and the internal time spent answering questions about it.
- The retraining trigger has fired at least once since the last review, or you know why it has not.
- A named person can switch the system off, and that person still works for you.
- The exit cost in the roadmap matches the exit quote you can get today.
- If the answer is retire, the date is already in the diary.
Retirement gets planned at signature, because that is the only moment at which anyone has the attention to plan it.
Where the published guidance and the vendor guides converge
Five independent sources land on the same short sequence, which is worth saying out loud when a colleague asks why the checklist runs to twenty-eight items.
The figures that usually appear in the business case come from vendor research, and you should keep them with their source included.
| Figure | What it measures |
|---|---|
| 70 per cent | AI projects that fail to deliver expected business value through weak strategic alignment and planning |
| 18 to 24 months | typical timeline for an enterprise AI implementation |
| £2.1 trillion | projected AI business value by 2030, from PwC UK economic analysis |
| 6 | phases in the implementation method the guide sets out |
The figures in the table above come from HP's AI roadmap guide for UK businesses, which sets out a phased method running from strategic alignment and opportunity identification through infrastructure design, data strategy, model development, deployment and sustainable governance.
Read across the sources and the disagreement is mostly vocabulary. NIST wants the framework voluntary and reusable. The UK government list wants one central ethics framework with departmental depth below it. Microsoft wants leadership vision treated as the strongest single driver and states that no two organisations share a roadmap. Splunk wants drift and retraining scheduled on a calendar. HP wants six phases with a readiness assessment in front of them. What an AI implementation roadmap actually does, once all of that is stripped back, is small and unglamorous: The AI implementation roadmap assigns a date, a name and a number to each decision, so that at some point in the following year somebody can be asked about it and will have an answer.


