Design a practical AI capability demonstration

Purpose and fit

Use this to see how a skill works in practice after learning or before supervised use on a specific task. Create a demonstration in a safe environment with existing approved material or optional clearly labelled simulated variations. Unlike a broad learning plan, this resource drafts one practice exercise and a blank record for later observed actions. Explain its developmental purpose, record access and how learners can challenge an inaccurate observation.

Before you start

Use an existing authorised, anonymised task example and the acceptance criteria you already apply. A prior output or source excerpt is sufficient to start designing a practice activity; synthetic material is optional, not a prerequisite. If you cannot supply sources, request an outline only and leave the observer key unresolved. Any deliberately altered example must be a clearly labelled practice copy, not a change to the original record.

Expect a draft learner brief, evidence-linked observer key and blank feedback record, not a run exercise or demonstrated competence. Each issue in the key must be traceable to supplied material or explicitly labelled as an optional simulated variation.

Required inputs

• One representative task, its acceptance criteria and the specific capability to practise.
• An approved tool or offline approach and any existing authorised source excerpts or prior output; missing material limits the draft to an outline until supplied.
• Any known issues in the example, such as an unsupported claim or ambiguous instruction; optional simulated variations can be proposed in a separate labelled practice copy.
• A domain-informed observer, accessible instructions, agreed accommodations and time for a second attempt.

Practical workflow

1. Define the target action narrowly: for example, checking material statements against sources before accepting an internal summary. State what the exercise cannot assess, including unrelated tools or task types.
2. Prepare a learner brief with the goal, permitted inputs, available support and boundaries. Keep a separate observer key locating evidenced issues or explicitly labelled simulated variations and acceptable responses. Label the packet as a simulation.
3. Agree observable criteria before the session. Use actions such as “opens the cited source” or “holds the unsupported statement for review.” Include an acceptable choice to avoid AI when it adds no useful support.
4. Run the exercise without live sends or changes. Invite the learner to explain key choices, verbally or in another accessible format. Allow the support that would genuinely be available in the intended workflow and record it.
5. Review the artefact and process together. For each criterion, mark “observed,” “not observed” or “not assessable here,” citing the relevant action. Distinguish an incorrect result from a tool failure or unclear exercise instruction.
6. Let the learner revise and retry with a changed example. Agree one next practice activity and any supervision needed for the bounded task. Keep actual work authorisation with the existing authorised owner.

Expected output

A learner brief, observer key and feedback record. Copy the following template:

Exercise ID/version: [ ]. Task and intended recipient: [ ]. Capability in scope: [ ]. Exclusions: [ ]. Safe inputs/tool: [ ]. Allowed support: [ ]. Time/access adjustments: [ ]. Expected artefact: [ ]. No-live-action boundary: [ ].

Criterion: [ ]. Observable action: [ ]. Evidence/location: [ ]. Observation status: [ ]. Support used: [ ]. Learner explanation or correction: [ ]. Feedback: [ ]. Retry change: [ ]. Next practice: [ ]. Record owner/access: [ ].

Example criterion: “Resolves conflicting source dates before using a figure.” Evidence might be a source comparison and an explicit hold on the disputed figure, rather than simply a correct number.

Human checks

Test the exercise with a reviewer before using it. Check that the source packet supports the observer key and that more than one reasonable approach can succeed. Ask the learner whether the setup reflected the real task. Keep personal learning evidence access-limited; AI can help draft exercises using sanitised material, but should not autonomously score or rank employees.

Limitations

A simulation samples behaviour under artificial conditions. Observing a skill once does not establish reliable performance under workload or justify access expansion. This proposed learning method is not a professional credential, validated personnel assessment or substitute for required training.

Data, model variability and your responsibility

Use relevant data that exists or that you provide, with accuracy and completeness sufficient for the task. Different AI models and versions may produce different results, and responses may vary between runs. You are responsible for checking and validating the output against your own requirements, source information and acceptance criteria before relying on it.