A team member can explain retrieval-augmented generation, name several AI tools, and describe an agent. Then a customer request arrives, and the same person is unsure which information to provide, whether the draft is accurate, or who should approve it.
That is the problem behind the message: your team doesn’t need more AI buzzwords. The next learning goal should identify work people need to perform and the decisions they must make while performing it.
Role-based AI training connects learning to the tasks, tools, responsibilities, and review requirements of a particular role. Choosing the right practice task means finding something useful, bounded, supported by approved information, and possible for a reviewer to check.
This guide helps department leaders and learning teams select those tasks. It includes a selection process, a completed hypothetical example, and a reusable exercise brief.
Start with the work your team needs to perform
Ask employees what they repeatedly prepare, review, explain, or reconcile. Collect examples of the work, then identify where the difficulty occurs.
“Learn prompting” names a technique. “Draft a support response from an approved policy and flag questions the policy cannot answer” names a task and its boundaries.
Microsoft Learn’s guidance on role-based skilling journeys starts from the tasks, decisions, and tools associated with a role. That is a useful planning principle: the work should shape the exercise.
For the broader program around these exercises, use LuMay’s workforce AI upskilling plan. Here, the immediate decision is narrower: which task should this group practise next?
Collect candidates before choosing a demonstration
Invite a department lead, someone who performs the work, and someone who reviews it to describe a recent example. Each sees different constraints.
The employee can explain missing inputs and repeated steps. The reviewer can identify errors that make an output unusable. The department lead can explain which improvement matters.
Record the trigger, inputs, output, reviewer, and common exception for each candidate. Include the existing way of completing it. A conventional template or clearer form may solve the problem without AI.
Keep candidates small enough to discuss concretely. “Automate customer support” contains many tasks. “Prepare an escalation summary from one ticket thread” is a usable starting point.
Use five checks to select a practice task
The following checks are a proposed planning aid. They help teams explain a selection decision; they are not a validated scoring system or a universal readiness standard.
1. The task matters to the role
Choose work learners recognize and expect to perform again. Establish what becomes easier or more reliable if they develop the skill.
A task need not happen daily. A less frequent reporting activity can still be worthwhile if the output matters and the team will have an opportunity to repeat it.
Selection evidence: a role owner can explain when the task occurs, who uses the output, and what currently requires improvement.
2. The inputs are approved and understandable
Prepare a source pack learners may use in the chosen environment. It could contain synthetic records, public documents, or internal material approved for that purpose.
Removing names alone does not establish that a document is safe to use. Sensitive details can remain in free text, attachments, or combinations of facts. Follow the organization’s handling requirements.
Selection evidence: the data owner and tool owner have confirmed the permitted material and environment.
3. The scope has a clear stopping point
State what the learner will produce and which decisions remain outside the exercise.
A useful output might be a draft response, an exceptions list, or a proposed set of questions. It should be clear when the work is ready for review, when information is missing, and when the learner should stop.
Selection evidence: the task brief names a deliverable, exclusions, and escalation conditions.
4. A reviewer can check the result
A fluent answer is not an acceptance criterion. Identify the evidence a reviewer will use to decide whether the output is accurate, complete enough, and suitable for its purpose.
D2L’s employee generative AI training guide connects role-specific activities with observable learning outcomes and output verification. Apply that principle by defining the review conditions before the exercise begins.
Selection evidence: a reviewer can explain what would cause the result to be accepted, revised, or rejected.
5. Practice can happen without unintended live actions
Set up the exercise so a mistake does not automatically send a message, change a customer record, approve a request, or trigger another consequential action.
Use the restrictions supported by the training environment and verify them. A written instruction to avoid an action does not establish that the tool lacks permission to perform it.
Selection evidence: the facilitator understands the available tools, access boundaries, and recovery procedure.
Compare candidates with a task-selection matrix
Illustrative example: a support team is choosing its first supervised AI exercise. These are proposed training choices, not measured LuMay customer results.
| Candidate | Useful learning outcome | Preparation needed | Decision |
|---|---|---|---|
| Draft a reply from an approved return policy | Use evidence, identify missing facts, and prepare a reviewable response | Synthetic request, current policy, and reviewer | Select once preparation is confirmed |
| Automatically approve refunds | Would combine eligibility decisions with live financial actions | Separate authorization, integration, and control design | Exclude from this introductory exercise |
| Summarize a long escalation thread | Preserve important facts and distinguish unresolved questions | Approved thread and reference summary | Keep as a follow-up task |
| Invent a customer-service use case from a blank prompt | Practise brainstorming | Clear purpose and evaluation conditions | Do not choose for the current evidence-review objective |
The decision is not that one task is always superior. The first candidate fits this group’s stated objective and has a checkable output. The refund task requires a different level of authorization and preparation.
Turn the selected task into an exercise brief
Write one brief that the facilitator, learner, and reviewer can all use. Avoid leaving important constraints in a separate conversation.
| Field | Completed hypothetical example |
|---|---|
| Role and task | Support representative drafting a response to a return request |
| Inputs | One synthetic customer message and one approved policy version |
| Output | A draft reply, source references, and a list of missing facts |
| Allowed AI assistance | Summarize the request and propose draft wording |
| Human responsibility | Check policy support, identify uncertainty, and approve the final wording |
| Excluded actions | No message sending, refund approval, or customer-record changes |
| Review conditions | No invented eligibility, accurate policy references, and appropriate treatment of missing information |
| Escalation | Conflicting policy or an unresolved exception goes to the designated reviewer |
The learner should explain why a statement belongs in the response, not only show which prompt produced it. Include an imperfect draft or ambiguous case so the exercise requires judgment.
Check whether the task teaches a reusable skill
After the first attempt, change a relevant condition. Use a different request, remove an important fact, or introduce a source conflict. Keep the learning objective comparable.
Ask the learner to identify what changed, what remains supported, and which parts require another decision. This tests whether the person can adapt the method rather than reproduce a demonstration.
Keep the review proportionate. The goal is to see how the learner frames the task, checks evidence, revises the output, and recognizes boundaries.
For detailed before-and-after measurement, use LuMay’s corporate AI training scorecard. Task selection and training effectiveness are related decisions, but they require different evidence.
Prepare the handoff into supervised training
Before scheduling the exercise, confirm that participants can access the approved tools and source pack. Identify who will answer policy questions and who will review the work.
If any of those conditions is unresolved, record it as preparation work. Do not interpret a missing login or unavailable reviewer as an employee skill deficit.
LuMay’s corporate generative AI workshop agenda explains how a prepared exercise can fit into a supervised session. For specialist IT scenarios, see the hands-on workshops for IT teams.
Copy this task-selection brief
- Role and recurring task: Who performs it, and when?
- Current work problem: What needs to improve?
- Approved inputs: Which records and policy versions can be used?
- Required output: What must the learner produce?
- Allowed AI assistance: Which parts can the tool support?
- Human decisions: What remains with the learner or reviewer?
- Acceptance evidence: How will the result be checked?
- Stop conditions: What requires clarification or escalation?
- Practice boundaries: Which live actions are unavailable?
- Selection decision: Use now, prepare first, or defer—with a reason.
Bring your team’s learning goals
Choose one role and one task your team needs to perform more confidently. Bring an approved example, its reviewer’s expectations, and the questions people currently struggle to answer.
Use those details to select a practice task before choosing the presentation or demonstration. Bring your team’s learning goals.





