Build an AI skills matrix
AI skills matrix: link tasks, autonomy levels, controls and evidence to build a training plan that is genuinely useful.
By Pablo Nastar

Knowing how to write a prompt is not enough to prove that a person knows how to use AI in their job. They must also choose the admissible data, spot a fragile output, find a source and call on the right validator. An AI skills matrix makes these actions observable without ranking people according to their ease with a fashionable tool.
The matrix serves to decide a training path and a level of supervision. It must neither become an individual ranking nor claim to guarantee the organisation's compliance.
Contents
- Start from tasks and risks
- Define observable levels
- Collect evidence without an exam effect
- Turn gaps into paths
- Maintain the matrix over time
- Frequently asked questions
Start from tasks and risks
A skill only exists in a context: task, data, output, impact and responsibility. A person can be autonomous to summarise a public text and need supervision to analyse a contractual document.
Inventory the real actions
Ask the teams where AI already intervenes or could intervene: research, rephrasing, minutes, document analysis, visual production or automation. For each task, note the input, the output and the person who makes the final decision.
The AI skills audit helps collect the uses and needs. The matrix comes after that diagnosis: it describes what the organisation expects, not what people say they like.
Link mastery to context
The consolidated text of the European AI regulation asks providers and deployers to foster AI literacy, taking into account the knowledge, experience and context of use. In its article 4 it does not impose an identical level on every individual. The matrix must therefore stay proportionate to the roles.
Define observable levels
A useful level describes an action that can be seen and a limit that the person knows how to recognise. Avoid « beginner », « intermediate » or « expert » without an associated behaviour.
| Level | Observable behaviour | Supervision |
|---|---|---|
| Discovery | recognises authorised uses and forbidden data | direct accompaniment |
| Framed use | follows a procedure, cites the source and reports a doubt | systematic review |
| Limited autonomy | chooses the method within a defined scope and documents the control | sample-based review or on sensitive cases |
| Supervision | designs the test, handles the exception and updates the rule | assigned governance |
Write a sheet per task
The sheet describes the expected level, the exercise, the evidence, the blocking errors and the person who evaluates. « Produce minutes » becomes, for example, « check the list of decisions in the authorised recording, assign the actions and flag the uncertain passages ».
Do not confuse tool and skill
An interface changes quickly. The ability to protect data, verify a statement and document a decision remains transferable. Mention the software used during the exercise, but do not make it the only criterion of mastery.
The NIST AI Risk Management Framework organises risk management around governance, mapping, measurement and management actions. It is a voluntary resource, not an individual skills certification.
Collect evidence without an exam effect
The assessment must resemble the work, while using dummy or authorised data. A vocabulary questionnaire does not show whether the person will know how to stop a fragile use.
Use three types of evidence
An observation shows the action in situation. A deliverable lets you examine the sources, corrections and reservations. A short interview reveals the reasoning when the rule does not perfectly cover the case. Combine them according to the criticality of the task.
Hypothetical implementation. A group could receive a dummy document with a missing piece of information and data that must not be uploaded. Each person would choose the scope, prepare an output and then explain their control. The assessor would score the evidence, not the beauty of the text. This scenario is not an Educasium client deployment.
Prevent stigmatisation
Announce the purpose: build a path, not punish a person. Limit the collected data and give access to the criteria before the exercise. A weakness on a task does not define the employee's overall professional value.
Turn gaps into paths
Each gap must lead to an action: training, tutoring, procedure, restriction or tool improvement. Not everything is solved by an extra module.
Prioritise by exposure
A skill critical for a person who distributes client deliverables can come before a rarely used function. Take into account the frequency, the severity of an error and the detection capacity. The skills development plan stays linked to real roles.
The AI skills development plan helps turn needs into actions. For managers, the guide on checking AI outputs specifies the supervision posture.
Design a progression
Step 1: choose a priority task. The scope must be narrow enough to be assessed.
Step 2: define the expected action. Write an action and an observable limit.
Step 3: prepare an exercise. Use a representative and authorised case.
Step 4: collect the evidence. Keep only what serves the training decision.
Step 5: propose the action. Training, guided practice or restriction must address the cause.
Step 6: re-assess. Repeat a different exercise to check the transfer.
Explicit absence of client experience. Educasium does not claim here a deployed matrix or a progression result at a client. The model must be calibrated by the organisation, then tested on its tasks and responsibilities.
Maintain the matrix over time
A dated matrix becomes obsolete when the tool, the contract, the process or the risk changes. Each sheet has an owner and a review date.
Trigger the useful updates
Revise after an incident, a new function, a data change or a change of role. Remove the skills that no longer match the work. Keep a history of the criteria so as not to compare assessments built on different versions.
Limits and counter-cases
A matrix that is too detailed becomes impossible to maintain. A matrix that is too general guides no decision. It also does not make up for an absent usage policy, an unframed tool or blurry responsibilities. If the organisation does not know which task it authorises, it must settle that point before assessing people.
Also document the skills that deliberately stay out of scope so that the absence of a line is never interpreted as an implicit authorisation.
Frequently asked questions
How many levels should be planned?
Use the minimum needed to distinguish the responsibilities. Three or four levels are often enough to express discovery, framed use, autonomy and supervision. What matters is the observable difference between them. Avoid artificial precision that cannot be justified.
Should all tools be assessed?
No. First assess the actions and environments actually authorised. A transferable skill is worth more than a list of interfaces. A new tool may require an additional control, especially if the data or functions change. The matrix stays focused on the work.
Does the matrix prove compliance with the AI Act?
It can document part of the literacy measures, but it does not by itself demonstrate compliance. The context, the systems, the roles and the other obligations must be examined. Have the interpretation validated by the competent people. Do not present training as a regulatory certification.
How to avoid the exam effect?
Explain that the goal is to build the right support. Use situations close to work and allow consulting the procedures. Assess the decision and the request for help, not the memory of a course. Then give concrete feedback and a possibility of progression.
Make the matrix a tool for progression
The AI skills matrix becomes useful when it links a task, a piece of evidence and a responsibility. If it changes neither the path nor the supervision, it is just one more table.
The introduction to AI and prompting training can be adapted to your team's priority actions. Training eligible for OPCO/FIF PL funding subject to conditions. Qualiopi-certified organisation. To frame the levels and exercises, contact Educasium.