Many companies approach AI through isolated projects: an assistant here, an automation there, a trial launched by a motivated team. Each makes sense on its own. But without an overall view, initiatives multiply, tools pile up, the same security questions come back every time, and nobody really knows what AI delivered after a year.
A roadmap is not meant to predict everything. It is meant to decide in which order to move, with what resources, and how to know whether you are making progress. Here is the framework we use with our clients to build a realistic twelve-month plan.
Before you start: who is in charge?
A roadmap without an owner remains a document. Before the first quarter even begins, three roles must be clear:
- An executive sponsor, who sets priorities and unlocks resources.
- An operational lead, who coordinates projects, tracks indicators and keeps the roadmap up to date.
- Business champions in each team involved, who know the processes and drive adoption.
Support functions must also be involved early: IT for integration and security, the DPO or legal team for compliance. Involving them from the start costs a few meetings; discovering them at the end costs months.
Quarter 1: take inventory and spot quick wins
The first quarter is about understanding the real situation. Which AI tools are already in use, officially or not? Which processes take the most time? Where is the useful data, and what state is it in?
This inventory produces a list of possible use cases. We score them on three simple axes:
- Value: time saved, errors avoided, impact on customers or revenue.
- Feasibility: availability and quality of data, integration complexity, maturity of the team involved.
- Risk: data sensitivity, consequences of an error, regulatory exposure.
High-value, easy-to-implement, low-risk cases are the ideal candidates to start with. The quarter can also deliver a few quick wins: a business-grade tool made available to everyone, a one-page usage policy, a first training session. They build momentum without waiting for the big projects.
A good roadmap does not only say what to do. It also says in which order, and what you accept not to do right away.
Quarter 2: deliver a first measurable project
The second quarter is for the first real project. Just one, well chosen, on a limited scope. Before starting, you set the baseline and the success threshold: “processing time for this type of request halves”, “data entry errors on this flow all but disappear”.
This first project has two goals. The first is obvious: deliver a concrete gain. The second matters just as much: serve as a template. It lets you refine the method, validate technical and security choices, and show internally what a well-run AI project actually produces. A small proven success is better than an ambitious large project still under way at the end of the year.
Quarter 3: industrialize and govern
Once the first project is in production, the third quarter lays the foundations for doing more, faster and more safely.
- A shared technical foundation: access to models under business contracts, secrets management, logging, test environments. What was built for the first project becomes reusable.
- Light governance: a simple process to propose, assess and approve a new use case.
- Baseline compliance: an inventory of uses against the AI Act, GDPR checks for any processing of personal data, documentation of choices.
- Training: tailored to each team’s real use, not a generic theoretical course.
Quarter 4: scale up, with judgment
The last quarter widens the portfolio: two or three new projects chosen from the first-quarter list, re-scored in light of what was learned. Some use cases that seemed like priorities at the start of the year turn out to be less attractive, others emerge.
It is also time to review the year: measured gains, real costs, team adoption, any incidents. This review feeds the next roadmap, and it is what will justify, or not, speeding up.
Knowing when to stop a project
A healthy roadmap also plans exits. A project should be stopped or redirected when the success threshold set at the start is not reached despite reasonable adjustments, when usage or maintenance costs persistently exceed the gains, or when teams do not adopt it. Stopping a project is not a failure: it is what lets you focus resources on what works. Companies that never allow themselves to stop end up with a collection of half-used prototypes.
A concrete example
A services company of around 150 people starts with no formal AI strategy, but with several tools already used in a scattered way. In the first quarter, the inventory surfaces about fifteen use cases. Three are kept as priorities; a business assistant is made available to everyone with a usage policy. In the second quarter, the first project, sorting incoming customer requests, goes live with a quantified goal. In the third, the technical foundation and an approval process for new projects are put in place, and each department gets tailored training. In the fourth, two new projects start, while a third, judged less profitable after analysis, is dropped. By the end of the year, management knows exactly what AI delivered, and where to invest next.
Our recommendation
Do not try to write the perfect plan: an AI roadmap gets revised every quarter, because tools, costs and your own understanding change fast. What matters is having an owner, a list of honestly scored use cases, a first measurable project, and the discipline to take stock regularly. With these four elements, twelve months are enough to move from scattered initiatives to a structured approach.