How to build an AI roadmap: the four-week method we use
An AI roadmap is a ranked list of the AI use cases worth funding in your organisation, with a business case, the data needed, the risks and an order of delivery for each line. Done well, it takes two to four weeks and ends with a plan your executive committee can approve or reject line by line.
Most of the roadmaps I see in Belgian companies are something else: a list of tools to buy. Copilot for everyone, a chatbot for customer service, “something with agents” for operations. Nobody has checked whether the data exists, who would own the result, or what it would change in the P&L. Six months later the licences are paid and the pilots are still pilots.
Here is the method we use instead.
Start from the business, not from the tools
The first workshop has no AI in it. We ask the executive committee three questions: which numbers matter most this year, where do people spend time on work they dislike, and which decisions are made with too little information. The answers become the frame every AI idea is judged against.
This is also where we agree on what “worth funding” means. For one client it is margin; for another it is hours freed in a team that cannot hire.
Collect ideas wide, then narrow fast
In week one and two we interview each department for an hour. People know where the friction is. A finance team will tell you exactly which reconciliation eats three days a month, and a sales team which follow-ups never happen.
A mid-sized company usually produces dozens of ideas. In one recent executive programme we collected 210. That is fine. The point of the next step is to throw most of them out with a reason.
Artefact, the group Agilytic is part of, published a good example: with Bora Brasil it narrowed 83 ideas to 33, then 24, then five use cases for the first year, in nine weeks (case study).
Score every idea on the same five criteria
Criterion | Question we ask |
Value | What changes in the P&L or in hours, per year, if it works? |
Feasibility | Has this been done before with current AI, at this scale? |
Data readiness | Does the data exist, is it accessible, is it good enough? |
Risk | What happens when the AI is wrong? Is it high-risk under the AI Act? |
Adoption effort | How many people must change how they work? |
The table looks obvious. What makes it useful is that the same people score every idea, in the same session, so a favourite project does not get a softer grade.
Look at the data before you promise anything
For the top ten ideas we open the actual data: the shared drive, the CRM export, the ticket history. Good ideas fail here more often than anywhere else. The customer emails are not stored anywhere searchable, or the product data lives in three spreadsheets that disagree. Better to learn it in week three than in month three of a project.
Decide build, buy or use what you already pay for
Many use cases do not need a project at all. If your company runs on Microsoft 365, Copilot may cover meeting summaries and first drafts once people are trained. Teams that write, analyse or code a lot often get more from ChatGPT Enterprise or Claude. Agents that act in your ERP or CRM usually need building.
The roadmap makes this call per line, so you do not buy a platform to solve one problem or build what a licence already does.
Sequence twelve months, in three kinds of moves
We sort the short list into quick wins (productivity in the current process), differentiators (something competitors cannot copy easily) and transformational bets (a new way to serve customers). A good first year has two or three quick wins to build trust and fund the rest, one differentiator, and at most one bet.
Each line gets an owner in the business, not in IT, and one number that says whether it worked.
Write the rules at the same time
A roadmap is also where you decide what AI may and may not do in your company: which data can go into which tool, who approves an agent that acts on its own, how you check its work. Under the EU AI Act some uses carry obligations (recruitment and credit scoring, for example). Flag them now; they change the timeline.
What the executive committee receives
The ranked list, with the scores and the reasons ideas were dropped
A one-page business case for each of the three to five selected use cases
The tool decision per use case
A twelve-month sequence with owners and measures
A draft AI use policy if you do not have one
The first proof of value, scoped and ready to start
After the roadmap
The first proof of value usually takes four to eight weeks on real data. What survives goes into production, which is where most AI projects stall. That is the stretch our forward deployed engineers cover: they work inside your team until the use case runs and your people own it.
We have built roadmaps like this for an automotive brand drowning in disorganised data (case), a bank that wanted better conversion (case) and a health insurer’s digital transformation (case). Another leadership team took a different route and started with Copilot for the executive committee itself (case).
On the last day, the most useful thing an executive committee can do with the list is cross things out. Whatever survives that conversation has an owner.
Frequently asked questions
How long does an AI roadmap take?
Two to four weeks for a mid-sized company. Larger groups may need a roadmap per business unit.
Who should be involved?
The executive committee for the first and last sessions, one person per department for interviews, and someone who knows where the data lives.
Do we need a roadmap before rolling out Copilot or ChatGPT?
Not to start, but without one the rollout tends to stay at email and meeting summaries. The roadmap tells you where the tool should go next.
How often should it be updated?
Every six months. Models and prices change fast; your list of ideas should too.
Julien Theys, co-founder of Agilytic. Talk to us about your AI roadmap.