Frameworks & Templates4 min read

Demo, Simulation or Live Work? Choose the Right AI Practice Format

Match AI practice to the consequence of the task: demonstrate the method, simulate the judgement, then use live work with review.

Bokili EditorialPublished October 30, 2026· Verified October 30, 2026

A learner progresses from observing an AI demonstration to a safe simulation and then supervised live work.

A useful AI practice exercise can still be given in the wrong format. A trainer may demonstrate a workflow that employees need to perform, or send learners straight into live work when a safe simulation would expose mistakes first. The content may be sound, yet the learning design asks for too little evidence or too much risk.

The practical rule is simple: match the practice format to the consequence of the task. Use a demonstration to make the method visible, a simulation to make judgement observable, and supervised live work to test performance in context. Moving between those stages should depend on evidence, not enthusiasm.

The practice format is part of the skill

Skills England’s 2026 employer guide says effective AI training should link to real tasks and decisions, give learners work-relevant scenarios, and include feedback, reflection and repeat practice. Its supporting cases also warn that approaches do not transfer identically across organisations. That makes format a design decision: the closer practice moves to live work, the more local context, permission and review matter.

	Demonstration	Simulation	Supervised live work
Learner action	Observe and explain the method	Perform it on a safe fictional case	Use it on an approved real task
Main evidence	Can describe the sequence	Can make and justify decisions	Can perform under real constraints
Failure cost	Very low	Contained	Potentially real
Best next move	Recall check	Feedback and another case	Review, record and release decision

Do not confuse realism with readiness

A realistic exercise is useful only when the learner has enough support and the organisation has defined what may enter the exercise.

Use the demonstration–simulation–live ladder

Three practice formats, three kinds of evidence

Demonstrate

Show one bounded workflow from input to checked output. Make the decision points visible, including when the tool should not be used.

Simulate

Give the learner a fictional but plausible case. Change one condition, collect the reasoning, and offer feedback before a real consequence exists.

Use live work with review

Choose an approved low-risk task, name the reviewer and release rule, then record what changed between the AI draft and the accepted result.

Worked example: a supplier briefing

Imagine a procurement team learning to prepare a one-page supplier briefing with AI. In the demonstration, the facilitator uses fictional public information and narrates each choice: permitted sources, claims that require evidence, missing facts and the final review. The learner can see the method, but observation alone does not prove they can apply it.

Reading is a start. Practice makes it stick.

Start learning

Move the same skill through three formats

1
Watch one complete example
The learner marks where the facilitator checked evidence, excluded sensitive information and rejected an unsupported claim.
2
Run a simulation
The learner receives a fictional supplier pack containing one stale document, one conflicting number and one missing owner.
3
Explain the release decision
Before feedback, the learner states what can be used, what needs checking and what must stay out.
4
Try one approved live brief
A reviewer checks the source pack, the edits and the final use before the brief is shared.
5
Capture the delta
Record which judgement changed between simulation and live work so the next exercise targets a real gap.

The skill is not merely writing a concise brief. It is selecting permitted evidence, marking uncertainty and deciding whether the output is ready. Each format reveals a different part of that skill. The live task should therefore be the last test, not the first exposure.

Set gates for moving to the next format

Progression gate

The learner can explain the workflow without copying the demonstration.
The simulation includes a decision, not only a prompt-writing step.
Feedback names the behaviour to repeat or change.
The live task is approved, bounded and low enough risk.
A reviewer and release rule are named before work starts.
The learner can show what changed between the draft and accepted result.

Not every skill needs to reach live work during training. A high-impact or infrequent task may stay in simulation until the organisation has stronger controls. Conversely, a familiar low-risk task may move quickly to supervised use. The movement rule should reflect consequence, reversibility and the quality of available review.

Know what each format cannot prove

	Useful evidence	Still unproven
Demonstration	The method is visible and understandable	Independent performance
Simulation	The learner can decide in a controlled case	Performance with live data and pressure
Supervised live work	The workflow works in the learner’s context	Consistent performance without support

Use the WORK transfer test to decide when generic learning needs company context. Add a teach-back test before independent practice, and use a busy-day rehearsal once the basic behaviour is stable.

Choose a practice format in ten minutes

Write one AI-assisted work behaviour, not a broad topic.
Name the consequence of getting it wrong.
Choose the lowest-risk format that can reveal the required judgement.
Write one piece of evidence that allows progression.
Name the reviewer and release rule if live work is involved.

Good AI training does not race learners towards real work. It collects the right evidence at the right distance from consequence. Start by making the method visible, let people practise judgement safely, and move into live work only when the organisation can support it.

Build practice, not exposure

Bokili helps teams turn role-specific AI behaviours into short practice missions. The format can change while the decision, evidence bar and feedback remain clear.