Train your team on AI by replacing the one hour demo with four weeks built on their own repeating work: week one documents how a task is done today, week two runs it with AI and keeps both versions, week three turns the corrections into a shared note, and week four is a real demonstration. Nobody needs a training department to run it, only an owner or manager willing to facilitate four short sessions and collect four kinds of paperwork.
Key Takeaways
- A demo needs follow-up practice on real tasks before you can assess anyone's skill.
- Each week has one named outcome and one artifact the owner actually collects.
- Real skill shows up as the ability to name a specific past mistake and produce the check that caught it.
- The person touching customer information needs a different depth of training than the person drafting internal notes.
- The plan's output becomes the evidence behind your written AI use policy, not a separate exercise.
Why does the one hour demo fail to produce real AI use?
A demo shows the owner what a tool can do. It does not show an employee how to do their own job with it, because the demo was never run on their own task, their own data, or their own mistakes. Without a practice assignment, an employee can watch, nod, and return to the same workflow without ever showing whether they can use the tool correctly.
The failure is not enthusiasm. A team can leave a demo wanting to try the tool. The failure is that wanting to try something and knowing how to check its output on a real task are two different skills, and only the second one changes how work actually gets done. A sales team that gets excited about an AI tool in a kickoff call and then drifts back to the old process within weeks is the same pattern described in MetaTechAi's piece on getting a sales team to keep using an AI system after launch: initial interest is not the hard part, and the gap between a launch meeting and sustained daily use is where the owner needs a repeatable practice and review routine.
A four week plan fixes this by making the team's own recurring work the material, not a canned example. Nobody is being trained on a hypothetical; they are being trained on the exact task they will still be doing in month two.
What happens each week, and what does the owner collect?
Each week has one outcome and one piece of paper (or shared file) the owner is responsible for collecting before the next week starts. If the artifact is missing, the week did not happen, no matter how the conversation went.
| Week | Named outcome | Artifact the owner collects |
|---|---|---|
| Week 1 | Each person picks one repeating task and writes down how they do it today, step by step, with no AI involved | The before process |
| Week 2 | Each person runs that same task with AI and keeps both the AI-assisted version and the original version | The paired outputs |
| Week 3 | The team reviews where the AI output needed correction and writes the fix into one shared instruction note per task | The shared instruction note |
| Week 4 | Each person demonstrates the task end to end, including the check step, in front of the owner or the team | The demonstration record |
Picking the task matters as much as running the weeks. A repeating task is one the person already does at least weekly, with a clear start and a clear finished state, such as drafting a follow up email, summarizing a call, formatting a weekly report, or writing a job note. A one time project, like researching a new vendor, does not repeat often enough to produce a useful before-and-after comparison inside four weeks.
What is the owner's facilitation procedure?
This is a working procedure for the person running the plan, whether that is the owner or a single manager. It is written as a sequence to run, not a description of a program that already exists elsewhere.
- Before week one starts, list every repeating task across the team and assign exactly one task to each person, matched to work they already do.
- In the week one session, have each person write their current method for that task as a short numbered list, with no AI step included, and collect every write-up before the session ends.
- Read each before-process write-up before week two begins, and flag any task description too vague to test, such as "handle customer emails," and ask that person to narrow it to one email type.
- In week two, have each person run their task with AI at least once on a real instance of the work, saving the AI output next to their own version from week one rather than replacing it.
- Midway through week two, ask each person a single question in a short check-in: "What is one thing the AI got wrong on this task?" If someone has no answer, inspect their checks and use a fictional example with a known error to test the review habit before week three.
- In the week three session, go through each pair of outputs as a group, and for every correction found, write the fix into one shared instruction note for that task type rather than a private note only one person can see.
- Close week three by writing a short, separate record of any step in the task that still needs a human regardless of the correction, such as final approval on anything sent to a customer.
- In week four, have each person demonstrate their task end to end in front of you, including the step where they check the AI's output, not just the step where they generate it.
- After the week four demonstration, file the shared instruction notes and the human-required record next to your written AI use policy, then set a date to repeat the plan on the next task.
Running this on one task per person keeps the plan small enough to actually finish. Adding several tasks at once makes it harder to inspect each person's work and resolve corrections before moving on.
What proves someone has real skill, not just enthusiasm?
Enthusiasm is easy to produce and easy to mistake for competence. Someone who likes a tool will say so, show you an impressive output, and move on. Skill shows up in two narrower places, and both are things you can ask for directly.
The first is whether the person can name a specific thing the AI got wrong on a past attempt, not a generic caveat like "it's not always right." A real answer sounds like "it listed a deadline that was never discussed" or "it summarized a call but missed the part where the customer asked for a refund." A vague answer is a reason to ask for a demonstration of the review step.
The second is whether the person can produce the check they ran, not just describe it in general terms. That might be a side-by-side comparison against the source document, a second person's sign-off, or a short note explaining why a number was correct. If someone claims to review AI output but cannot show you what reviewing looked like on their last attempt, you do not yet have evidence that the review step is reliable. These two questions are the entire test. Ask them during the week three review session, not just at the end of the plan, since that is when the habit either exists or it does not.
Who needs a different depth of training than everyone else?
Not every task carries the same risk, and training depth should follow the risk, not the org chart. The person handling customer, patient, or financial information needs more than the person drafting an internal note or an agenda, because the cost of a missed error is different in kind, not just in degree.
For someone touching customer information, the week two and week three work should include a step on what may go into the tool in the first place, a named reviewer who signs off before anything reaches a customer, and a documented check specific to that data, such as confirming a name or account number against the source record before sending anything out. Start practice with fictional or appropriately redacted inputs unless the account and data have been approved for that use. For someone drafting internal notes, meeting summaries, or first drafts of routine replies, a lighter review by the person themselves is usually proportionate, provided the inputs are nonsensitive and the result will not drive a consequential decision. Internal notes can still contain confidential information, so the input rules apply to everyone.
The guide to AI tools for teams covers how to match a tool to a task by what the task actually requires, which is the same judgment call applied one level earlier, before training even starts. If your plan needs to cover the fuller range of roles and task categories this applies to, the AI training for employees overview walks through that broader picture.
How does this plan connect to the AI rules you already have?
If your business already has a one page AI use policy or a short set of staff rules, this plan is not a separate program running alongside it. It is how that policy gets tested against real work instead of sitting unread. The small business AI policy guide sets out the rules about approved tools, customer data, and named reviewers; the four week plan is where a team actually produces the evidence those rules assume exists, such as who checked an output and what they found wrong with it.
Two artifacts from the plan belong directly inside that policy. The shared instruction notes from week three become the specific, task-level guidance a general policy cannot write in advance, because they are built from an actual mistake the team found rather than a hypothetical one. The human-required record from week three becomes the plain list of what the policy means when it says a person must check the result before anyone relies on it; instead of a generic instruction, you get a specific list of steps for each task that still need a named person's sign-off.
There is also a risk management reference for this approach. The National Institute of Standards and Technology's AI Risk Management Framework Playbook sets out, under its governance function, that organizations should establish ongoing training for personnel covering applicable rules, the potential negative impacts of AI systems, and the organization's own policies, with the training tailored to different roles rather than delivered as one generic session (NIST AI Risk Management Framework; NIST AI RMF Playbook). That is a voluntary framework, not a law, but it describes the same shape as the plan above: role-based training tied to real use, not a one-time session.
For businesses whose AI use falls within the EU AI Act's scope, Article 4 adds an AI literacy obligation. The European Commission's AI literacy guidance explains that providers and deployers should support staff learning with attention to their knowledge, experience and the context of use. Having European customers alone does not settle which requirements apply to your business. Check your systems, role and operating context with qualified help. This four week plan is an editorial starting point for practice and recordkeeping, not a certification or a guarantee of legal compliance. Keep the training record available when reviewing those requirements; this article is not legal advice.
What should you do with this before next Monday?
Pick one repeating task per person, tell them week one starts with writing down how they do it today, and put the week four demonstration date on the calendar now, before the plan starts, so it does not quietly slip. The four artifacts, not the four conversations, are what make the plan real: if you reach week four without a before process, a paired output, and a shared instruction note for every person, the plan did not run, whatever discussions happened along the way.
If you want help thinking through which tasks to start with or how to scale this past four people, contact AI Guy and describe the repeating tasks your team already does every week.
What FAQs do small business owners ask about training employees on AI?
Do we need a training department to run this four week AI training plan?
No. The plan is built to run without one. The owner or a single manager facilitates it using the team's own recurring tasks as the curriculum, using tools already approved for the work. Set aside staff time for practice and review, and check whether any task needs specialist support before including it.
What if an employee cannot say what the AI got wrong on a past task?
Ask them to show their source checks before deciding whether they need more practice. A correct output is possible, so do not require an invented error. If no mistake appears, use a fictional practice output with a known omission and ask them to identify it and document the correction.
Does everyone on the team need the same depth of AI training?
No. A person handling customer or patient information needs stricter input rules, a named reviewer and a documented check step. A person drafting nonsensitive internal notes or an agenda can use a lighter review if the output will not drive a consequential decision. Apply the approved input rules to every role.
How does this plan connect to the AI rules we already wrote for staff?
The plan produces the evidence your policy assumes exists: a record of who checked what, what the AI got wrong, and which steps still need a human. File the shared instruction notes and the human-required list next to your written AI use policy so the policy stops being a one-page promise and becomes a working reference.