By working convention, this article calls an enterprise AI roadmap the document that orders over time three kinds of lines: the artificial intelligence use cases the organisation has decided to fund, the shared workstreams on which they depend (data, platform, evaluation resources) and the decision points at which each case will be continued, reworked or stopped. The word has other uses: it can refer to an entire strategy, or to the schedule of a single project.
Its construction follows an order specific to AI. The use cases are settled first, then linked to the data and foundations they require, then classified with regard to the regulatory framework. Each is then divided into stages, because its result remains uncertain until it has been trialled. Finally come the schedule, set according to resources, and the review rule. Each step delivers a document and has someone accountable for it. This order and these accountabilities are a proposed method, to be adjusted to each organisation’s governance bodies. What applies to any roadmap (milestones, levels of certainty of dates) is set out in our article on the roadmap of an IT project.
The enterprise AI roadmap starts from use cases already chosen
The input to construction is the list of use cases selected at the end of prioritisation, each with its business case, work described in our page on AI strategy. If this list is missing, or if it rules out none of the ideas gathered, construction stops and refers back to that choice.
The starting point also matters: what the organisation already knows how to put into service limits what it can include in the first periods. Our article on enterprise AI maturity describes this assessment.
This step produces one line per selected case: the process concerned, the business owner, the target indicator and the sourcing model. The latter weighs on duration: a function supplied by a vendor and a custom-built system do not occupy the same place in a schedule. The management that arbitrated is accountable for the list, each business owner for their line.
Data and shared foundations determine what can start
For each line, the assessment covers the data sets or documents required, their owner, the right to use them for this purpose and their quality. Depending on the case, the data is ready, or it requires a preliminary workstream, which becomes a line in the roadmap with its own accountable person, or it does not exist, and the case leaves the roadmap.
Foundations are assessed in the same way: model access gateway, indexed document base, evaluation tools, operations pipeline. Entered as a separate line, placed before the first case that awaits it, a shared foundation has a single budget and a single accountable person.
The output of this step is a dependency map: for each case, the data and foundations it awaits, with their status. Data owners are accountable for their part, the department that builds the foundations for its own. The questions to ask about each data set are set out in our article on data governance for AI.
Regulatory classification comes before dates
The obligations of Regulation (EU) 2024/1689, known as the AI Act, vary with the risk level of the use and with the organisation’s role, provider or deployer of the system. Two technically similar use cases may therefore not require the same work before they go into service. Each line receives, before being dated, the role held by the organisation and the regime under which the system falls. When personal data is processed, the GDPR, Regulation (EU) 2016/679, also applies, whatever this regime.
The work that this classification entails becomes tasks in the line, with their duration: for a high-risk system, the documentation and the organisation of human oversight, for example. The target go-live date, for its part, is compared with the regulation’s application timetable, examined in our article on the AI Act obligations for businesses.
At the end, each line carries a written classification and the list of work that follows from it. The business owner is accountable for it together with the legal or compliance function, through the process described in our page on AI governance. This classification remains a legal qualification, which the organisation settles with its advisers.
Stages separated by decision points
An AI project carries uncertainty about its very result. The quality the system will reach on the organisation’s data cannot be read in any specification: it is measured by trial. A go-live therefore cannot be dated like the delivery of a piece of equipment.
Each line is divided into stages. Our approach has four: Frame, Prove, Industrialise, Operate. Between two stages sits a decision point, with a threshold written before the trial (what the system must achieve, on which set of cases, measured by whom) and three possible outcomes: continue, rework the stage, stop. The roadmap dates these decision points. Beyond the next one, a line’s deadlines remain conditional and are displayed as such.
This step gives, for each line, its stages, the threshold of each one and the date of the next decision point. The business owner is accountable for the threshold, the body that funded the case for the decision. The calculation of costs and gains that it then examines is covered in our article on the ROI of artificial intelligence.
The order is set by scarce people and by recurring cost
The lines are then placed in time. Dependencies give a first order: a data workstream before the case that awaits it, a foundation before its first user. Two constraints correct it.
Capacity, first. The business experts who build the evaluation sets and judge the answers may be the same people who run the process concerned, and lawyers, security and data teams are called on by several lines at once. The roadmap reserves these people for each period. Arbitration is made alongside the organisation’s other projects: reading capacity by critical skill and the body that decides are the subject of our page on project portfolio management.
Budget, next. Each case put into service adds an operating expense that lasts as long as it does: model usage, monitoring, re-evaluation. The roadmap therefore displays, alongside the build budget, the cumulative recurring spending it commits.
The schedule of lines and the table of reservations result from this. The person who maintains the roadmap proposes them, the management body arbitrates them.
A review on a fixed date, and at each trial result
An AI roadmap is reviewed at the pace of the body that arbitrates it. Between two sessions, events reopen it: a decision taken at a decision point, a delayed data workstream, a change of classification, a supplier’s withdrawal of a model version, a change in its prices.
The entry rule is written down with it. A new use case goes through the same qualification as the others, then takes the place of a line or waits.
Each review delivers the next version, headed by the movements since the previous one (lines added, removed or moved) and the reason for each. The person who maintains the roadmap is accountable for it.
The first version does not need to be long. Two or three cases, their data, their classification and the date of their next decision point may be enough to open the first arbitration session; further lines are added as trials deliver their results.