Role-Based AI Literacy: Assess Needs Before Assigning Training

A procurement team summarises quotations, customer service drafts replies with an assistant, and a manager decides which tools may be introduced. All use AI, but the consequences differ. Assigning an identical course misses those differences. Before distributing content, answer three questions: which decision does the system support, who checks the output, and what must that person know to do so? A role-based learning-needs review starts here. It connects actual tasks with observable behaviour instead of broad labels about competence.

Separate the legal context from your chosen method

Our existing guide to AI literacy under Article 4 explains the regulatory context. This article adds a practical worksheet for assigning learning measures in German and European organisations. It does not introduce a mandatory examination, an official qualification or a ranking of employees by productivity. The purpose is a reasoned decision about the support a role needs for a particular AI use case. A separate assessment remains necessary to determine the legal position of the system itself.

The current European Commission FAQ describes measures supporting AI literacy while considering knowledge, experience, education and the context of use. It explicitly states that Article 4 does not require measuring individual employees’ knowledge. Nor is a particular certificate required. The case-based review below is therefore an organisational method you choose, not a statutory test. It helps identify learning gaps and target support. Neither an internal score nor a completed course should be presented as automatic confirmation of full legal compliance.

Start the worksheet with a task, not a department

Create one record for each relevant task. Capture the role, specific work assignment, system used, permitted inputs, affected people, expected output, human review and escalation contact. Add the observed learning need, agreed measure, responsible owner and next review trigger. These fields are a design proposal, not a statutory form. One entry might describe customer service drafting a complaint response: only approved case data may be used, and the assigned reviewer checks any commitments before sending the message.

Describe the use precisely enough for another manager to understand its boundaries. “Marketing uses AI” is insufficient. “Marketing drafts text variants from approved product information, with a specialist checking claims before publication” can be assessed. Also record whether the system merely writes, prioritises content or prepares decisions. Our guidance on safe AI use at work helps clarify permitted data and tool boundaries first. Training cannot retrospectively substitute for an unresolved decision about whether a particular tool or use is approved.

Use a small number of genuine workflows without collecting unnecessary personal content. An anonymised process example or synthetic practice case is often sufficient. Ask employees where uncertainty arises: entering data, judging outputs, obtaining approval or reporting an error. Include occasional tasks and deputy arrangements explicitly. Someone may work confidently on routine tasks yet need extra support for a one-off sensitive assignment. Infrequent use is not automatically low-consequence use, particularly when the output reaches customers or influences another person’s treatment.

Distinguish four roles, even when one person fills several

Distinguish users, subject-matter reviewers, system owners and decision owners. Users need to recognise input boundaries and typical failures. Reviewers need criteria for judging outputs and authority to reject them. System owners need knowledge of configuration, access, changes and available supplier information. Decision owners establish permitted purposes, resources and escalation routes. These roles are a practical organisational aid rather than a new legal classification. In a small company, one person may hold two or three roles, but each responsibility still needs explicit consideration.

For each role, examine existing knowledge in relation to the actual system. Technical experience does not automatically demonstrate understanding of confidentiality or harmful business decisions. Conversely, an experienced administrator may review outputs effectively without knowing how to develop models. The distinction between awareness and specialist qualifications is also useful in ISO 27001 training: general awareness and a specialised assignment require different learning objectives. Apply that principle to AI instead of assigning the same technical introductory course to every person.

Base learning needs on possible consequences

Next consider the consequences of a plausible mistake. Does a draft remain internal, become public, change a contractual commitment or influence how someone is treated? Also consider input sensitivity and the likelihood of detecting errors in time. This operational prioritisation is not a classification into statutory AI risk categories. An internally urgent learning need does not itself establish that a system is legally high-risk AI. Keep the operational worksheet and the formal system assessment connected, but do not merge their labels.

A useful decision aid uses three questions rather than a falsely precise overall score: could a mistake affect people or important interests, can the role reliably detect it before it takes effect, and is an accessible specialist named for uncertain cases? Where the first answer is yes and the second is no, supervised practice takes priority over more reading. Where no contact exists, the first problem is organisational. Another course cannot close that gap; the process owner must establish the missing route.

Where inputs contain personal data, the data decision is a separate learning objective. The role should explain which information is needed for the approved purpose and which information must not enter the tool. Connect this task to existing GDPR training rather than expressing privacy rules differently in every AI course. Approving the substance of a draft response does not automatically mean that all underlying data may be processed. Keep these decisions separate in the worksheet and identify the appropriate contact for each.

AI-generated

Use short cases to identify the actual gap

Give employees a concrete task containing one uncertainty. A generated product description includes an unsupported performance claim, a summary omits a contract exception, or an assistant requests additional customer information. Do not ask only for “correct” or “incorrect”. Ask the person to explain how they recognise the issue, which source they would check and when they would stop processing. This distinguishes a knowledge gap from missing access, time pressure or unclear responsibility, each of which needs a different response.

Agree simple observation criteria beforehand: permitted inputs identified, output compared with a suitable source, uncertainty acknowledged and the correct reporting route selected. Record the result as a specific development task, such as checking contract exceptions against the original before approval. Avoid the sweeping judgement “not AI literate”. Employees should understand the purpose of the observation and who can access it. Use practice data rather than genuine personnel assessments or confidential customer files that are unnecessary for the exercise.

Three cases show why one course is not enough

In the first case, customer service uses AI to draft replies. An employee identifies invented facts but accepts an unauthorised goodwill commitment. The learning need concerns the connection between text review and decision authority, not writing better prompts. A suitable measure is a supervised complaint exercise using actual approval boundaries. In a transfer task, the employee explains which commitment they may make in a different example. Our discussion of AI shopping assistants helps connect transparency and human review to customer-facing decisions.

In the second case, procurement summarises lengthy quotations. An employee understands the data rules but misses that a delivery date is conditional. They need a comparison exercise using the original document, the summary and a deliberately omitted exception. The solution is not unlimited confidence in a second AI answer. The subject-matter reviewer must locate and explain the relevant contract passage. The worksheet should therefore specify which statements always need to be traced back to the original source before a decision is taken.

In the third case, a manager approves a new assistant without establishing who investigates errors. The learning need concerns governance and accountability. A decision exercise should cover purpose, affected roles, boundaries, escalation and a review trigger. ConformBase offers a leadership AI governance course for this area; individual system assessment and organisational approval remain the company’s responsibility. Compare the current course scope with the documented gap. A participation certificate substitutes neither for the decision itself nor for the reasoning supporting it.

AI-generated

Only now assign learning measures and check their effect

Match each gap to an appropriate measure. Missing basic concepts may be addressed by a short learning unit. Uncertain output review requires examples and feedback. Unclear boundaries require an authoritative working instruction. Missing escalation routes require a process decision. The Commission’s repository of AI literacy practices presents different formats but cautions against assuming automatic compliance from copying an example. Choose according to the task and starting knowledge, not the size of a course catalogue or the attractiveness of its certificate.

After the measure, use a comparable but different case. If the person merely repeats the model answer, transfer remains uncertain. If they can choose the relevant source, explain a boundary and escalate a concern, you have better evidence of practical value. Record problems that lie outside training too. A correct answer in a practice setting achieves little if the real workflow provides no access to the original document or makes the required reviewer effectively unavailable. Learning and operational design need to reinforce each other.

Connect the measure, role, system version and outcome to your training records. A concise note can capture the trigger, observed gap, practice case, agreed support and owner. Limit access and detail to the actual purpose. The record should show why a measure was chosen and what remains unresolved afterwards. It need not create a permanent personality profile. Keep participation, subject-matter review and operational approval distinct so that a green course status is not accidentally interpreted as unrestricted authorisation to use a system.

Use changes as review triggers rather than relying on repetition

Define concrete review triggers: a new tool, changed purpose, new data types, different approval rights or an observed mistake. Add a scheduled organisational review where useful for your process. Do not describe that schedule as a universal statutory annual deadline under Article 4. Integration into the overall mandatory training plan prevents duplicate assignments and exposes interfaces with privacy and information security. Where a change affects only part of the workflow, update the affected roles instead of automatically repeating every course for everyone.

Start with a manageable pilot: one common task, one accountable business team and a few typical uncertainties. Complete the worksheet, observe a case and address the most important gap. Expand only after that. The outcome is not a claim that every employee is now comprehensively competent. It is an inspectable connection between task, risk, support and responsibility. That connection turns automatic course allocation into a reasoned learning decision that can be updated selectively when tools, processes or responsibilities change.