Who to pick for the first AI training pilot group

What the questionnaire actually assesses

An AI training pilot group often gets assembled by availability or by who happens to be interested. The question of who to pick then narrows into judging people: who is capable, who deserves it. The questionnaire below assesses the work task a candidate currently performs. It does not rate the candidate as a person or their track record to date. The same person can come out suited for one task and unsuited for another; what is being judged is always the pairing of task and candidate, not the candidate alone.

The questionnaire has three blocks. The first two ask about the task and about standing within the team; a weaker answer on one question can be offset by a stronger answer elsewhere. The third block holds two exclusion questions that work differently: a yes answer removes the candidate regardless of the rest of the profile. The rule for turning answers into a finished group follows after the questionnaire.

Block 1: questions about the candidate’s work

These questions ask about the work itself. They do not ask how long the candidate has done it, or how well they do it today.

1. Does the candidate have a repeating task with a checkable output? Yes means the task can be observed more than once during the pilot, and the outputs can be compared against each other. The candidate is suited to the pilot. No means a one-off or hard to compare task will not let the pilot show much. Look for a different task from the same person, or consider a different candidate.

2. Does someone else currently check the output of this work? Yes means there is already a person capable of judging output quality, and the pilot can watch the difference through their eyes. Include the candidate. No means that without an existing check, nobody on the pilot can judge whether the output got better, got worse, or simply changed shape. Either set up a check before the pilot starts, or leave the candidate out for now.

3. Can the state before AI use be captured? Yes means the time, the steps, or the shape of the output can be recorded before the pilot begins, giving something to compare against later, a habit also worth carrying into how often to repeat AI training and when a refresher is enough, which depends on the same kind of before-and-after record. No means without a captured starting state, the pilot will prove little beyond a participant’s impression. Consider a short preparatory phase before confirming the group.

Block 2: questions about standing within the team

4. Will colleagues see the result? Yes means the result appears in a shared output or a downstream step of the process, so the pilot can shape how the wider team feels about the tool. No means the result stays isolated and nobody else sees it. This does not exclude the candidate on its own; it just limits what the group can tell you about the rest of the team.

5. Can the candidate genuinely set aside time for the pilot? Yes means they have room in their schedule for practice and short feedback without putting their regular work at risk. No means without set aside time, the pilot turns into overtime or a step that keeps getting skipped. Get confirmed time from the manager before including them. Agreement with the idea of a pilot is not the same thing.

6. Is the candidate’s manager willing to accept a temporarily changed way of working? Yes means the manager knows the task will run differently for a while and understands how the pilot will be assessed. No means without that agreement, the pilot’s outcome risks never being adopted once it ends, even if it goes well. Settle this before including the candidate. Negotiating it after the pilot is too late.

Block 3: exclusion questions

These two questions work differently from the four above. A yes answer removes the candidate from the pilot regardless of how well they answered the questions above.

7. Is the candidate the only person who can cover a critical task during the pilot? Yes means exclude the candidate. The pilot must not put an unrepresented piece of operations at risk. Look for a different candidate, or wait until cover is in place. No means cover exists, and this question does not exclude the candidate.

8. Would the pilot expose data that must not go into the AI tool? Yes means exclude this specific task from the pilot, or adapt it to use synthetic or anonymised material instead of real data, a distinction that comes up more broadly in is it safe to put company data into ChatGPT. Which data may or may not go into an AI tool is decided by the company’s internal policy, or a data protection officer, not by this questionnaire. No means this question does not exclude the task.

Turning the answers into a group

The completed questionnaires produce a list of tasks the pilot can carry. In practice CIAD recommends three to six candidates in a single group: few enough that every questionnaire and every output can be checked individually, and enough to cover more than one type of repeating task. This is an operational recommendation from CIAD. It is not drawn from a measured optimal size, and the AI Act does not prescribe one.

A group is worth assembling once it covers at least two different task types, ideally from different teams. A pilot built from a single task type shows little about whether the findings transfer anywhere else.

The exclusion rule cannot be offset by anything else. Answers in blocks 1 and 2 balance against each other: a weaker answer on one question can be offset by a stronger one elsewhere. Answers in block 3 are not offset by anything. A candidate can have an ideal repeating task, a visible result, and manager support, and still not belong in the pilot if they are irreplaceable for a critical operation, or if their task would require putting data into the AI tool that internal policy does not allow there. These two yes answers are the only ones in the entire questionnaire that remove a candidate regardless of the rest of their profile.

Once the group is set, the next decisions are channel, length and logistics, and the common ways a training programme goes wrong once it is running, a topic covered in the mistakes that ruin company AI training and in general or industry-specific AI training: how to decide.

Personal data in selection and pilot tracking

The questionnaire itself assesses a task. Once answers turn into an actual record of who was picked, who was left out and why, and whose output gets tracked before and after the pilot, that record becomes processing of personal data in an employment relationship. GDPR Article 88 allows for specific rules covering processing in an employment context, and Articles 5 and 6 require a defined purpose and a legal basis for any processing. No part of this questionnaire secures those conditions on its own.

Who may compile a selection report, how long records are kept, who has access to them, and whether the selection process itself is fair to employees who were not chosen, are questions for the company’s HR and legal team to assess. This article does not determine whether a specific selection method complies with GDPR; that assessment belongs to your own HR or legal adviser, applied to your specific situation.

The AI Act and training order: what the law actually requires

Companies sometimes justify pilot group selection by pointing to the AI Act, as if the law dictated who gets trained first. The current wording, since Regulation (EU) 2026/1744, prescribes no order and no pilot phase at all. It requires providers and deployers to take measures supporting the development of AI literacy among people who work with the operation and use of an AI system on the organisation’s behalf, taking into account their knowledge, experience, education, training and the context of use. It does not set a number of phases, a format, or who goes first. According to the European Commission’s current Q&A, there is also no obligation to measure individual employees’ knowledge, and no certificate is required; an internal record of training is enough if an organisation chooses to keep one.

A pilot group built from the questionnaire above is an organisational choice about how to trial an AI tool in a limited, controlled way on selected tasks. Who gets chosen first says nothing about that person’s importance to the company, and nothing about any order the law expects for training the rest of the team.

Sources and limits

The questionnaire above is a CIAD editorial tool. It has no demonstrated predictive value as a measured instrument. None of the three cited sources tested whether a specific combination of answers predicts a successful pilot; the closed source library contains no such research. The recommended group size and composition in the section on forming a group are a CIAD operational recommendation, explicitly labelled as such. No study offers that figure as a finding. The section on the AI Act draws on the current wording of Article 4 after Regulation (EU) 2026/1744 and on the European Commission’s Q&A, both verified as of the dates given for those sources; it is a summary, not legal advice. The section on personal data cites the specific GDPR articles that apply to selecting and tracking employees, but it does not conclude whether the described process complies with the law: assessing the specific design of the processing, its legal basis, and the fairness of selection toward employees who were not chosen remains for your own HR and legal team.