Implementation Playbooks5 min read

AI Training for Employees: Add a No-AI Control Run

AI training for employees should include one fair control run. Compare the same task with and without AI so learners can judge quality, total effort and task fit.

Bokili Editorial· Verified November 10, 2026
ShareX
The same employee task completed with and without AI, then compared on quality, time and review effort

AI training for employees often assumes that every exercise must use AI. That teaches operation, but not judgement. An employee can produce a polished answer without learning whether AI was the right method, whether review erased the speed gain, or whether an ordinary template would have been clearer. Add one no-AI control run: complete the same bounded task with and without the approved tool, then compare both methods against the same standard.

The goal is not to make employees compete with a model or to prove that one lane always wins. It is to teach task fit. NIST’s AI Use Taxonomy starts with human goals and outcomes, then describes how AI contributes to the task. The UK Government AI Playbook likewise says use cases should be led by business and user needs and should offer a significant advantage over existing techniques. A control run turns those principles into practice.

The same employee task completed with and without AI, then compared on quality, time and review effort
A control run lets learners compare methods before deciding where AI fits.

Why AI training for employees needs a control lane

A learner who only sees the AI-assisted route cannot observe the counterfactual: what would have happened through the existing process. The comparison need not be a formal impact evaluation. It should be a small, fair exercise that makes the work visible. Use the same brief, source pack, time window, acceptance criteria and reviewer for both runs.

This also prevents training from becoming the default answer to every performance gap. US Office of Personnel Management guidance asks what performance is required, why it is not happening and whether training is the best solution. Sometimes the better intervention is a clearer template, a simpler approval, better source material or a change in job design. A control run gives that possibility a place in the lesson.

Run the same task through two fair lanes

The CONTROL exercise

  1. 1

    1. Choose a bounded task

    Use a familiar, low-risk task with one clear output, such as a six-line supplier update from fictional notes.

  2. 2

    2. Lock the brief

    Give both lanes the same source pack, audience, deadline, data boundary and acceptance criteria.

  3. 3

    3. Run without AI

    Use the current method: reading, a template, search, spreadsheet or professional judgement. Record elapsed work and review.

  4. 4

    4. Run with approved AI

    Use the organisation’s approved tool and the same materials. Keep prompts, corrections and human checks visible.

  5. 5

    5. Compare total effort and quality

    Count preparation, prompting, checking, correction and hand-off—not only generation time. Score both outputs with one rubric.

  6. 6

    6. Decide the fit

    Choose AI, the existing method, a hybrid or no use. Record the reason and what evidence could change the decision.

The same-person comparison is useful for learning, but it is not a scientific productivity claim. Order can affect the result because the employee learns the source pack during the first run. Alternate the order across learners or use two equivalent fictional cases when fairness matters. The point is disciplined judgement, not a publishable effect size.

Worked example: a supplier update

Reading is a start. Practice makes it stick.

Start learning

An L&D designer prepares two equivalent fictional supplier cases. Each contains a delivery date, one missing assurance document, a changed price and an unresolved question. The required output is a short manager update with confirmed facts, open issues, source references and a clear decision boundary.

In an illustrative run, the existing template takes sixteen minutes and meets five of six criteria. The AI-assisted lane produces a draft in seven minutes, but checking and correcting it takes ten more because it turns the missing assurance document into a confident claim. The final output also meets five criteria. These invented figures do not prove that AI is slower. They show why generation time alone cannot decide the method.

Without AIWith approved AI
Brief and sourcesIdentical fictional packIdentical fictional pack
Work recordedReading, drafting, template usePrompting, drafting, iteration
Review recordedFact and format checkFact, source, omission and invention check
Quality standardSame six-point rubricSame six-point rubric
Possible decisionKeep, simplify or combineAdopt, revise, limit or reject

The debrief is where the learning happens. Ask which step improved, which new failure appeared, whether the output was easier to review and what would happen on a fresh case. The answer may be a hybrid: use the template to lock facts and structure, then use AI only to shorten or adapt the checked draft. It may also be “do not use AI for this task”.

Score the method, not enthusiasm

Use a short rubric with observable criteria: factual accuracy, completeness, source fidelity, decision boundaries, usability and total effort. Add one stop condition for unsafe data or high-consequence judgement. Do not grade confidence, prompt style or willingness to use AI. The employee is demonstrating method selection, not loyalty to a tool.

For more consequential deployments, use stronger evaluation design. The UK guidance on evaluating AI interventions recommends establishing a baseline and considering efficiency, accuracy and quality. A ten-minute learning exercise is not a substitute for that work. It is a way to build the habit early: compare against something, include review cost and keep the human goal stable.

Build a no-AI control run
  1. Choose one repeated, low-risk task and write the useful outcome.
  2. Prepare one fictional source pack and a six-point acceptance rubric.
  3. List the existing non-AI method and the approved AI-assisted method.
  4. Decide how you will record preparation, execution, checking and correction.
  5. Run or schedule both lanes with the same reviewer.
  6. Finish with four options: AI, existing method, hybrid or no use—and one sentence of evidence.

Connect the exercise to a practical learning path

Use the control run after a three-task employee baseline and before scaling a workflow pilot. Pair it with task decomposition when the work has several reviewable stages, and use the work-transfer guide to test whether the chosen method survives a fresh case. Bokili for HR and L&D supports this kind of short, role-aware practice with feedback and visible progress.

Good AI training for employees does not end with “the learner used AI”. It ends with a reasoned method choice. A fair control run teaches employees to notice where AI adds speed, where it adds review, where a hybrid works better and where the existing process should stay.

Practise this in Bokili

Turn what you just read into a ten-minute mission.

  • Compare one task with and without AI

    Run two fair lanes and choose the method using quality, total effort and review evidence.

    Start learning
  • Start with a three-task employee baseline

    Observe task fit, input safety and output checking before choosing training.

    Start learning
  • Teach task decomposition

    Split complex work into reviewable passes with human checks.

    Start learning
  • Measure one work transfer

    Test whether a skill survives a fresh but comparable case.

    Start learning

Sources

  1. AI Use Taxonomy: A Human-Centered Approach — National Institute of Standards and Technology
  2. Artificial Intelligence Playbook for the UK Government — Government Digital Service
  3. Guidance on the Impact Evaluation of AI Interventions — HM Treasury and Evaluation Task Force
  4. Planning & Evaluating — US Office of Personnel Management
  5. AI training for HR & L&D leaders — Bokili
ShareX

Reading is a start. Practice makes it stick.

Bokili turns skills like this into ten-minute missions for your whole team, with instant feedback and progress you can see.

Start learning

Keep reading