How Should a Small Business Train Staff to Use AI?

A small business should train staff on one real job, not on a tour of AI features. By the end of the session, the team should know what information may enter the tool, what a useful answer looks like, who checks it, and where the approved work goes.

Bottom line: Choose a repeated, low-risk task with an obvious reviewer. Run the same example together, compare the outputs with the source, and leave one person responsible for improving the routine. A prompt library without that workflow is just another folder nobody owns.

This guide includes a filled 75-minute exercise for turning approved call notes into an internal follow-up brief. The example is illustrative. It is not a Prairie AI client result or a tested performance claim.

Pick the workflow before the tool

Good training starts with work the team already understands. A customer follow-up draft, meeting summary, weekly report, quote handoff, or document checklist can work because staff can compare the AI output with a source record.

Avoid the first-session temptation to use live personal or confidential information. The Office of the Privacy Commissioner of Canada advises businesses to limit the sharing of personal, sensitive, and confidential information and to build privacy safeguards into AI use. Use a fictional or properly redacted example until the business has approved the tool, account, and data path.

The Government of Canada's generative AI guide is written for federal institutions, not private businesses. Its risk distinction is still useful for planning a workshop. Editing a draft that receives another review is lower risk than summarizing client information or using AI in service delivery. Start where a person can check the answer quickly and safely.

A filled 75-minute team exercise

Picture a five-person service company. The office receives call notes from technicians and needs a short internal brief before anyone contacts the customer. Three staff members attend the session: the office manager, a dispatcher, and the owner.

Fictional source note: Prairie Street Bakery had an intermittent freezer alarm. A technician visited Tuesday, found the temperature normal, and cleaned one loose sensor connection. The technician asked the office to confirm whether the alarm returned after 48 hours. The note does not approve another visit, state a price, or promise a response time.

Expected internal handoff:

• •Follow-up needed: Check with Prairie Street Bakery after Thursday afternoon to ask whether the freezer alarm returned.
• •Known: Temperature was normal during Tuesday's visit. The technician cleaned a loose sensor connection.
• •Missing or unapproved: No price, return visit, cause, warranty decision, or response time is confirmed.
• •Owner: The office manager reviews the source note, then decides whether to call the customer or send the item back to the technician for more detail.

The point is not to reproduce those sentences exactly. A useful answer must keep the timing, preserve the difference between an observation and a diagnosis, and avoid inventing a price or promise.

Minutes 0 to 10: write the boundary

Before opening an AI tool, ask the group to complete these lines:

• •Input: approved technician call notes with customer names replaced by fictional names for training.
• •Output: an internal follow-up brief.
• •Reviewer: office manager.
• •Final location: the existing follow-up queue.
• •Stop and ask a person when: the note mentions safety, an angry customer, pricing, warranty, personal information, a contract, or a promised date.

If the team cannot agree on those five lines, stop. The problem is the workflow, not the prompt.

Minutes 10 to 25: make a first attempt

Give everyone the same source note and the same instruction.

Training instruction: Turn the source note into an internal follow-up brief. Include what happened, the next check, missing information, and the staff role that should review it. Do not diagnose the equipment, promise a visit, state a price, or write a customer-facing message. If the source does not support a detail, mark it missing.

Each person runs the instruction in the approved training account. Save the three answers beside the source note. Do not spend the whole session polishing prompts. The variation between answers is part of the lesson.

Minutes 25 to 45: review against the source

Use this review card for every answer:

• •Source match: Can each factual statement be traced to the note?
• •Missing facts: Does the answer admit that price, diagnosis, warranty, and a return visit are not approved?
• •Next action: Does it name the 48-hour check without turning it into a customer promise?
• •Owner: Does it route the brief to the office manager?
• •Tone and format: Can a colleague scan it without mistaking it for a message ready to send?

Mark each line pass, fix, or unsupported. An answer fails if it invents a diagnosis, says the problem is resolved, promises a visit, or hides a missing fact behind confident wording. A clean writing style does not rescue unsupported content.

Minutes 45 to 60: test the exception

Add one sentence to the source note: "Customer says the alarm is affecting food safety and wants a technician back today."

Run the same instruction again. The expected change is not a more persuasive reply. The brief should flag the safety concern and same-day request for immediate human review. The training succeeds when staff notice that the ordinary process no longer applies.

This second round teaches a habit that a feature demo rarely does: the prompt may stay the same while the correct workflow changes because the source changed.

Minutes 60 to 75: assign the next week

End with a short operating plan:

• •The office manager owns the review card and gathers corrections.
• •Staff use the routine on five approved, low-risk notes during the next week.
• •Every brief remains internal until the normal reviewer approves the next action.
• •The team records unsupported details, missing source information, and time spent reviewing.
• •After five examples, the owner decides whether to keep, revise, or stop the routine.

Five examples will not prove a return on investment. They can show whether the task is stable enough for further testing and whether staff can catch a bad answer before it leaves the office.

Choose a workflow champion, not an AI enthusiast

The champion should know the work and have enough authority to correct the routine. An office manager, dispatcher, program coordinator, estimator, or operations lead may be a better choice than the person who writes the cleverest prompt.

Give the champion a narrow job:

• •keep the source example, instruction, and review card together
• •collect failed or awkward outputs without hiding them
• •update the instruction only when the team can explain the problem
• •route tool, privacy, security, and policy questions to the right owner
• •remove the routine if staff stop checking the source

The champion does not approve every possible use of AI. They own one training workflow and the evidence from it.

Decide whether the training worked

After one week, review the five examples. Count how many briefs passed without a material correction, how many contained unsupported details, and how long review took. Ask whether another staff member could follow the card while the champion was away.

Keep the routine when the source is consistent, reviewers catch errors quickly, and the brief lands in an existing queue. Revise it when the same field is missing or staff disagree about what counts as approved. Stop when the task exposes information the tool is not approved to receive, the output cannot be checked efficiently, or nobody owns the exception path.

Saskatchewan's April 2026 Economy Committee record noted that training and development came up in industry adoption discussions and described funding for a hands-on boot camp intended to help up to 10 Saskatchewan small and medium-sized enterprises move from experimentation to operational use. That is useful local context, not proof that any particular workshop works. The team still needs evidence from its own job.

If you already have a task and staff group in mind, Prairie AI's hands-on AI training can build the session around your approved tools and real workflow. If the team cannot yet agree on the process to practise, start with an AI workflow audit and map the handoff before booking a broad workshop.