The demonstration goes perfectly. A manager types a question, the assistant produces a polished answer, and everyone gets a login. On Tuesday, a customer asks something the reference documents do not cover. A staff member copies the confident answer and sends it.
The missing lesson was how to recognize the boundary of the system's knowledge.
Useful AI training prepares people for the whole task: choosing the right source, protecting information, checking the answer, handling an exception, and knowing who owns the next decision. Your work product in this chapter is a role-based practice plan and a separate record of which employment uses require further review.
Mesa Equipment Service and the training examples are fictional. No employee has taken these exercises, received a performance rating, or been selected for a job through this guide.
Teach one job task at a time
Begin with a task someone already understands. A coordinator might draft an answer from an approved service policy. A purchasing employee might compare an invoice with an order and receiving record. A manager might review an exception and decide who should resolve it.
Ask the people doing the work where mistakes occur. If the policy itself is missing, another prompting lesson will not supply it. If the tool is unavailable on a worker's device, a lecture about motivation will not fix access.
The CDC's Quality Training Standards recommend assessing needs, defining learning objectives, using relevant practice, providing feedback, and supporting learners afterward. Those standards were developed for CDC training; the approach here adapts those ideas to business tasks and does not claim CDC certification.
Write an objective someone can demonstrate. “Understand AI” is difficult to assess. “Draft a service-fee answer that cites the current policy, preserves unknown availability, and remains unsent for review” gives the trainer something concrete to inspect.
Match the lesson to the role
People need a shared foundation, then practice that fits their responsibility.
| Role | Practice task | Evidence of readiness |
|---|---|---|
| Everyday user | Draft from an approved source packet | Supported facts, missing information identified, correct next step |
| Reviewer | Inspect a plausible but flawed draft | Finds the specific error and explains its consequence |
| Document owner | Resolve a policy conflict | Publishes an authorized source update through the actual process |
| Workflow owner | Review failed and ordinary cases | Accounts for quality, review effort, exceptions, and recovery |
| System administrator | Configure the permitted workspace | Demonstrates access boundaries and the approved change process |
These roles may overlap in a small business. The work still needs to be assigned. A person who knows how to write a prompt does not automatically have authority to approve discounts, change records, or release a workflow.
For a solo operator, separate the roles in time. Draft first. Review against the source in a second pass. Use independent qualified review when the decision needs expertise you do not have.
Run a short practice session with visible evidence
The practice pack proposes a sixty-minute session: ten minutes to explain the task and boundaries, ten to demonstrate source checking, twenty for an individual attempt, fifteen for review and correction, and five to assign follow-up.
That is a planning suggestion, not a required pace. Make room for the person's access needs and the task's difficulty. Provide working time, an approved account, readable instructions, and a way to ask questions without sharing private details with the whole group.
Demonstrate the reasoning a reviewer should perform. Open the source, identify its version, compare each factual claim, and point out what the document does not establish. Show an ordinary failure as well as a good answer.
Then let the learner attempt a new case. If the trainer supplies every correction before the learner can respond, the exercise measures the trainer's ability to finish the work.
Use Mesa's service question as a worked exercise
A fictional customer asks: “What does an inspection cost, and can you confirm Friday?”
Mesa's current SERVICE-01 policy establishes a $45 inspection fee, credited toward a repair approved within thirty days of the inspection. Availability must be checked before a booking is confirmed. An older SERVICE-00 document lists $35 and is superseded.
The source packet supplies no Friday slot for this request. A good practice answer can explain the fee and credit terms and say availability needs checking. It cannot confirm Friday. Office hours do not establish a technician's free time.
The trainer should ask the learner to identify the current policy, mark the missing booking evidence, and prepare an unsent draft. The question is not whether the learner can make the answer sound friendly. It is whether the answer preserves the business's actual commitments.
The companion cases add an obsolete price, an unsupported discount request, and text inside a supplied document that tells the assistant to ignore the policy. That text is part of the material being examined; it is not authority to replace the review rules.
Score the work, then teach the gap
Use a small rubric that separates serious errors from polish.
| Review area | What to look for |
|---|---|
| Source use | Current, relevant documents and traceable references |
| Factual accuracy | Correct price, terms, units, and other supported facts |
| Unknowns | Missing evidence stays unresolved |
| Authority | Drafting does not become approval, sending, or booking |
| Clarity | The reader understands the answer and next step |
For the supplied exercise, an invented booking, unauthorized discount, or use of restricted information is a critical failure. An unclear sentence may need revision without carrying the same consequence. Do not hide a critical error inside a high average score.
Record the failed criterion, the correction, and another suitable practice attempt. Describe the task performance, not the person's character. “Confirmed a time absent from the source” is actionable. “Not trustworthy with technology” is an unsupported personal judgment.
The thresholds in this training pack are proposed teaching rules. Passing them does not grant access or certify someone for every business workflow. A real workflow owner must assess readiness for the specific scope.
Check transfer to work after the class
A learner may handle the classroom case well and struggle when the source is longer or a customer asks two questions at once. Arrange a later review of an appropriate work sample, with access and privacy controls suited to the task.
CDC guidance on evaluating training effectiveness distinguishes learning from applying that learning at work. It also explains why a post-training score alone cannot establish how much knowledge changed without a baseline. A positive satisfaction survey is useful feedback, but it is not proof of competence.
Measure attendance, practice completion, demonstrated skill, and later application separately. In the constructed cohort, eight people were invited, six attended, five completed practice, and four demonstrated all required criteria. That means 75% attended, 62.5% completed practice, and 50% of invitees demonstrated the criteria. Among the five completers, the last rate is 80%.
None of those figures proves that training caused improvement: the packet contains no baseline assessment or follow-up workplace observation. The counts are fictional, not results from Randy's employees.
In a team or customer community, invite people who voluntarily demonstrate useful skills to help others practice. Review actual contributions and ask whether they want a mentoring role; offer support and a way to step back. AI may organize examples or suggest training topics, but an opaque engagement score should not decide who becomes a leader.
Use AI as a practice partner
AI can turn an approved procedure into varied practice questions, explain a term more simply, or produce a deliberately flawed draft for review. Check the teaching material before assigning it. An invented answer key can teach the wrong policy with great confidence.
Create practice cases from the supplied approved policy and task objective.
Vary the questions while keeping policy facts unchanged.
Include one case with missing information and one with a source conflict.
For each case, give the trainer an answer key with source references,
critical errors, and useful feedback. Keep the learner version separate.
Do not invent policy, employee traits, accommodations, or permissions.
All scenarios and outputs remain teaching drafts for review.
Offer a non-AI route when practicing the underlying skill. The worker should still be able to find the source, identify an unsupported claim, and use the fallback process if the tool fails.
Keep hiring assistance separate from hiring decisions
Drafting a job description from approved duties is different from ranking applicants. Turning an approved onboarding procedure into a lesson is different from deciding whom to promote or dismiss.
For this guide's starter workflow, use AI to prepare materials for review: a clearer description of actual duties, a draft interview guide, or a checklist of evidence an authorized reviewer needs. Do not introduce automated candidate ranking, personality inference, or employment decisions as a side effect of the training project.
US employment rules apply to selection procedures, including technology-based ones. The EEOC's guidance on employment tests and selection procedures explains the need for appropriate job-related validation and makes clear that vendor documentation does not remove the employer's responsibility. A general model's confident score is not validation for a particular job.
Before any consequential employment use, establish the relevant jurisdiction, job requirements, purpose, data, validity evidence, review process, and means of correcting errors. Address applicable discrimination, privacy, notice, and other requirements with qualified review. A small set of practice prompts does not resolve those questions.
Make access part of the process
An applicant's difficulty using an assessment tool may reflect a barrier in the tool. Do not treat it as evidence that the person cannot perform the job.
The US Department of Justice's AI and disability guidance describes how hiring technologies can screen out qualified people with disabilities. It explains the importance of assessing relevant job skills and providing reasonable accommodations unless doing so would create undue hardship. It also recommends clear information about the technology and an accommodation-request process. This is US ADA guidance; it is not a complete statement of every jurisdiction's employment rules.
For internal training, provide accessible instructions and appropriate alternatives, and ask what would help the person participate. Keep requests involving personal information in the proper confidential channel. Do not ask a chatbot to infer a diagnosis, disability, or protected characteristic from a worker's writing, voice, photograph, or practice results.
Finish with a named owner, a follow-up date, and a record of what remains untested. Training has done its job when people can use the tool, question it, and recover from its mistakes within their actual authority.
Next: Turn individual skills into a workable team change in Chapter 30.