A prompting course can teach an employee how to request a clearer answer. It does not necessarily teach them whether that answer is suitable for a financial report, which information they may enter, or who must approve the next action. Those are different learning needs, and they depend on the work the employee is responsible for.
The European Commission’s guidance on the EU AI Act’s Article 4 links AI literacy to the systems an organisation provides or deploys, the risks involved and the knowledge of the people using them. It does not prescribe a universal course or require a certificate. The practical question is therefore not simply whether employees have completed training. It is whether they understand enough to carry out their particular AI responsibilities.
Who needs AI literacy and what it should cover
The Commission’s AI literacy questions and answers explains that Article 4 concerns providers and deployers and people directly dealing with their AI systems. Coverage extends beyond employees to people operating or using systems on the organisation’s behalf, potentially including contractors, service providers and clients under its organisational remit.
Start by identifying the systems people actually encounter and what they do with them. Include people who review AI-generated work, not just those who enter prompts. For external personnel, establish which activities they perform on your behalf and what guidance they need.
The Commission recommends considering general AI understanding, the organisation’s role, system risks, people’s existing knowledge and the context and purpose of use, including the people affected. Whether the organisation is a provider or deployer is a separate legal assessment. Where development or integration makes that unclear, examine when ERP AI integration can create provider responsibilities alongside the workforce assessment.
This gives the programme a shared foundation without making every learner follow the same curriculum. Explain which AI systems are used, their opportunities and limitations, and the organisation’s rules. Then develop the knowledge needed for each responsibility.
Match learning to the decisions people make
The Commission’s proportionality guidance calls for adapting literacy measures to the organisation’s role and system risks. It identifies additional measures as potentially relevant for people dealing with high-risk systems. A useful way to apply that principle is to distinguish occasional use, review and operational oversight.
Occasional users need boundaries they can apply
Illustrative situation: an employee uses an approved assistant to draft an internal announcement. Their learning should cover what information may be entered, how to check the draft against the intended message and when to ask for help. Practise with a draft that sounds convincing but includes an unsupported commitment. The employee should recognise that polished wording is not approval to make that commitment.
Keep the exercise close to the task. Someone drafting routine messages needs to understand the relevant limitations and information rules, not necessarily how to configure an AI workflow.
Reviewers need to know what constitutes adequate evidence
Illustrative situation: a finance employee receives information extracted from a document by AI. Their responsibility is not just to judge whether the result looks plausible. Teach them to compare relevant fields with the source, recognise uncertainty or missing information and escalate unresolved discrepancies before the information is used.
The system-quality question belongs in financial document extraction acceptance testing. The literacy question is whether the reviewer understands those acceptance rules and can apply them. A clear rule is of little practical value if the person reviewing the output cannot recognise when it has been breached.
Managers reviewing AI-generated work also need to understand what they are approving. Ask them to distinguish between correcting presentation, verifying underlying information and authorising a business decision. These are not interchangeable checks.
Configuration and oversight teams need operational understanding
Illustrative situation: a team supervises an assistant that proposes changes to business records. Teach the team which actions remain drafts, what requires approval, how to recognise unexpected behaviour and whom to contact when the workflow must pause.
Keep the learning connected to the actual boundaries between drafting, approving and executing actions. Understanding an output’s limitations is a knowledge requirement. Knowing when and where to escalate is a procedural responsibility. Having authority to approve its use is an organisational decision. Training should explain each without suggesting that course completion grants approval authority.
For high-risk systems, the Commission’s guidance also points to the separate Article 26 obligation concerning sufficiently trained staff and human oversight. It says relying on instructions for use alone is insufficient.
Combine practical experimentation with accountability
A Microsoft ERP community discussion offers a useful adoption example. Participants described giving leaders protected time and support to build something relevant to their work, followed by discussion of champions who could help adoption spread. That approach gives learning a concrete business task rather than leaving it at a general demonstration.
The discussion also connected adoption with safeguards. A participant described an acceptable-use policy excluding personal information from AI tools, including bank-account information, and making employees accountable for the work they produced using AI.
Use that combination when designing learning: practical experimentation, accessible support and clear boundaries. Let champions help colleagues understand permitted uses and raise questions, while keeping formal approval responsibilities explicit. An enthusiastic colleague should not become the default approver simply because they know the tool well.
Exercises should therefore include a reason to stop or ask for help, not just a successful result. Ask learners to explain what they would check, what remains uncertain and who owns the next decision.
Keep useful records without turning certificates into compliance
The Commission says there is no required universal training format and no need for a certificate. Organisations can keep internal records of training or other guidance initiatives. These points leave room for task-based workshops, guided practice and targeted briefings rather than requiring a particular course product.
Keep the statutory literacy obligation separate from the supporting records you choose to maintain. Useful records can include:
- The systems and responsibilities considered in each role assessment.
- The learning materials and guidance made available.
- Participation and practical exercises completed.
- Questions or misunderstandings that require follow-up.
- The reason for updating learning when responsibilities or systems change.
A completion record shows that an activity took place. To assess whether the learning fits the job, also ask whether people can explain an output’s limitations, recognise an escalation point and identify the authorised approver. That is a more useful basis for improving the programme than treating possession of a certificate as the answer.
If you need help connecting workforce responsibilities with your wider obligations, discuss EU AI Act compliance support with Digiventi.