Take a manufacturer of building materials. Maintenance is trying out a failure prediction model, customer service a writing assistant, accounting the automatic reading of supplier invoices. The executive committee wants to know the enterprise’s AI maturity. Each division answers on the basis of its own trial, and the three answers do not tally.
An enterprise’s AI maturity refers, in this article, to its ability to take a use case from idea to operated service, then to start again with the next one. It is placed by recording signs that a third party can verify, in five dimensions: uses, data, putting into service and operations, oversight, skills. The survey gives a profile, and the gaps between dimensions designate the next step. The grid that follows is a proposed method, which claims to follow no published model.
On what basis should an enterprise’s AI maturity be judged?
The expression belongs to common usage: its definition depends on the grid you adopt. The one in this article holds to one principle. A sign is a piece of evidence that a third party can ask to see: a record of decisions, a dated and signed document, a service in operation with its measures, the name of a designated person. A questionnaire in which each division scores itself gathers an opinion. In our example, the three divisions would give themselves a good score, each on the basis of its trial.
The number of trials launched, licences distributed or euros spent describes an activity. It indicates neither whether a use case has reached service, nor whether the next one will get there with less effort.
The grid distinguishes three levels for each dimension. On trial: the practice exists in one team and depends on the people who carry it. Established: it is written down, it has an owner, and at least one service in operation complies with it. Repeatable: the next use case benefits from it without the team having to rebuild it. For the oversight dimension, a standard can serve as a benchmark: the ISO/IEC 42001 standard, presented in our entry, specifies requirements for an AI management system.
The signs to record, from uses to skills
| Dimension | Evidence to request | Sign of a practice on trial | Sign of a repeatable practice |
|---|---|---|---|
| Uses | The list of use cases and the record of decisions taken about them | Trials launched by the divisions, with no shared list or recorded stop | A maintained list, in which each case carries a business owner, an indicator and a dated decision |
| Data | For a case in progress: the data source, its owner, the latest measured quality status | Extracts prepared by hand by the person running the trial | Data served by a maintained pipeline, with the written agreement of its owner for this use |
| Putting into service and operations | For a system in service: the evaluation set, the result of the latest version, the rollback procedure | A demonstration on chosen cases, with no measurement after release | An evaluation set rerun at each change, cost and quality tracked, a person designated for operations |
| Oversight | The inventory of AI systems and the usage rules given to staff | No inventory, tools used without review | An up-to-date inventory, a written classification per system, a validation process with a known lead time |
| Skills | The names of those who build, operate and validate each system | Knowledge held by a single person or by a supplier | Written roles, business experts involved in evaluation, a service that survives a departure |
The “established” level is recognised between the two: the evidence exists for at least one service, without yet holding for the next.
The levels are not averaged. A single score would erase what the survey is looking for, because the lowest dimension limits what the others can deliver: a well-maintained list of use cases puts nothing into service without data or an operations pipeline. The result is therefore presented as a profile, one line per dimension, with the evidence that justifies each level.
What each gap in the profile calls for as a next step
When convincing trials do not go into service even though their data and their owner are there, the gap lies in the third dimension. The next step is a production release and monitoring pipeline, built on a first real model: our MLOps and LLMOps page describes its components.
Widespread uses without an inventory signal an oversight gap. Undeclared uses, defined in our entry on shadow AI, are identified first. The classification of systems and the validation process then fall under AI governance.
When it is data that holds back the use cases, the work is limited to the scope of the cases chosen: an owner, a verified right of use, a measured quality. These roles fall under data governance.
The reverse profile can occur: oversight and an operations pipeline in place, facing a list of cases with no owner or indicator. The work then comes back to the choice of use cases, the subject of our support in AI strategy.
There remains the case in which knowledge depends on one person or one supplier. The next step is writing down roles and transferring skills, to be carried out before any new case. One text touches on this dimension: Article 4 of Regulation (EU) 2024/1689 requires providers and deployers to take measures to promote the development of AI literacy among their staff, as our entry on the AI Act recalls. The record of these measures is among the evidence to request.
These next steps are then placed in time, with their dependencies: this is the role of the enterprise AI roadmap, which starts from the recorded profile.
The limits of the survey, and when to redo it
A profile holds for a scope and for a date. In a group, each entity has its own, and a consolidated profile would hide the one that can serve as a support for the others.
The target level is set according to the use cases chosen. An enterprise that confines itself to the AI functions of its off-the-shelf software does not need to build the operations pipeline of one that trains its own models; the inventory and usage rules concern it just as much. Aiming for the repeatable level in every dimension, with no use case requiring it, ties up resources.
Nor does the survey say what AI brings in. A repeatable practice can serve use cases of limited scope: measuring costs and gains is covered in our article on the ROI of artificial intelligence.
It is redone at each revision of the roadmap, with the same grid and the same evidence requested. Placed next to the first, the second survey shows what has changed level and what has stayed in place.