How to train employees to use AI on one real task

Two in three US firms with fewer than 250 employees now use AI, according to a 2026 survey of 3,732 of them by the US Chamber of Commerce (opens in a new tab). But in a July 2026 survey by Clarify Capital (opens in a new tab), only 26% of small business owners could point to a specific, measurable result from their AI tools. In many firms, the return stays low because staff don't know how to use an AI assistant well on their own work, or on work they share with colleagues. In Atlassian's 2025 survey (opens in a new tab), employees said AI made them 33% more productive, while 96% of the large companies surveyed hadn't seen a dramatic improvement in efficiency, work quality or innovation across the company.

This guide shows you how to train your staff to use the assistant in a way that pays off, with a one-hour session where people work through one real task they already do, learn to catch its mistakes and leave with a shared checklist. A week later, you look at how it went in their everyday work. The session works with any assistant your firm uses: ChatGPT, Claude, Gemini, Copilot, Grok or another.

Each step below follows the same example, a bookkeeper's monthly cash summary for the owner.

Which task should the session be about?

Start with a task your team repeats, one that passes five tests:

• It comes round often, weekly or monthly, so people use what they learned again soon.
• It has a clear result, so you can say what a good one looks like.
• Checking the assistant's work is quicker than doing the task by hand.
• A mistake is easy to spot and cheap to fix before anyone sees it.
• It uses information your firm allows in the tool.

Tasks that often pass: replies to customer questions that repeat, a summary of a long document, actions from meeting notes, or a first check of a spreadsheet for odd figures.

To see how this plays out, picture a bookkeeper who writes the owner a half-page summary of the month's cash: money in, money out, and anything unusual. It comes round every month, the result is clear, and every figure can be checked against the bank export. Financial work is a common place to start: in a Goldman Sachs survey (opens in a new tab) of small business owners who went through its 10,000 Small Businesses program, 42% said they use AI for financial analysis and business planning.

What do you need before the session?

Once you've picked the task, you need about an hour to prepare five things:

1. A task card. One page that says what the task is, who it's for, what a good result looks like, and what may go into the assistant.
2. Sample inputs with no client data. Last month's figures with the names changed, or made-up figures in the same shape.
3. A good example. A past result the person who receives it was happy with.
4. A flawed answer with mistakes planted in it. The section on catching a wrong answer, below, explains how.
5. Today's time. How long the task takes now, so you can compare in a week.

Also make sure everyone can sign in to the firm's account on the assistant, rather than a personal one.

For the bookkeeper, the task card looks like this:

How do you run the hour?

Split the hour like this, and give most of it to people working on the task:

• First 10 minutes: show the task card and the good example, and agree what a good result looks like.
• Next 20 minutes: each person asks the assistant to do the task with the sample inputs, then improves their request until the answer matches the card.
• Next 15 minutes: hand out the flawed answer. Everyone looks for the mistakes on their own, then you compare.
• Next 10 minutes: go through the checklist together and adjust it to this task.
• Last 5 minutes: each person names a piece of work they'll use it on this week.

Let people type their own requests, even when it's slower. In Atlassian's survey, employees learned to use AI best in small, hands-on groups working on a shared problem.

For the bookkeeper, a first request might be: "Here is last month's bank export and the list of payments due. Write a half-page summary for the owner: opening and closing balance, the three biggest changes, and anything overdue." The second try adds a fix for what the first answer got wrong, such as "Use only the figures in the export."

How do people learn to catch a wrong answer?

Atlassian found that 42% of employees trust AI answers without checking them when they're short of time. For the 15 minutes on the flawed answer, ask the assistant for the cash summary yourself before the session, then edit its answer so it contains three typical mistakes:

• a closing balance that doesn't match the bank export
• a supplier payment called overdue when it was paid
• a confident reason for a change that the figures don't support, such as "spending rose because of seasonal demand"

People often miss the invented reason, because it sounds as sure as the true figures beside it. Spend most of the discussion on it, and ask the group to find the line in the bank export that supports it.

What should people check every time?

The checklist on the task card has five questions, and people go through them for every answer:

1. Numbers: does every figure match the source?
2. Facts: is every claim in the source, or did the assistant add it?
3. Gaps: is anything on the task card missing?
4. Reader: would the person receiving it understand it without asking?
5. Information: did anything go into the tool that shouldn't have?

A person still checks every answer before it goes anywhere, and the checklist makes that check quick and the same for everyone.

What happens a week later?

Before everyone leaves the session, book a 30-minute follow-up for a week later, after they've used the assistant on their own work. At the follow-up:

1. Each person brings one example from their week and how long it took.
2. Compare the time with today's, the figure on the task card.
3. Look at what the checklist caught, and fix the step that went wrong most often: the request, the task card or the checklist.
4. Decide what's next: keep going, change the task card, or stop using AI for this task.

Say the bookkeeper's summary now takes 40 minutes instead of 90, and the checklist caught one wrong total in the first week, so the owner keeps the change and the team picks the next task. My note on telling whether an AI tool is actually helping covers how to measure a change like this over a longer stretch.

Common questions

How long does it take to train staff to use AI?

About an hour for the session and 30 minutes for the follow-up a week later, for each task. Start with one task, and add the next once the first one is part of everyday work.

Is a general AI course a good place to start?

It works better as a second step. Run the session on one task first, so people have already used the assistant on their own work when the course shows them what else it can do.

What if staff worry AI will replace them?

Say plainly what the time saved is for, such as work that's waiting or clients you couldn't take on before. In the session, people stay in charge: they write the request, check the answer, and decide what gets sent.

Which AI assistant should we use?

The one your firm already pays for, on its business plan, so the work stays in the firm's account. Before anyone uses it with client information, read my six checks before an AI tool touches client data.

Training on the tasks that matter most

Once the first task works, run the same hour for the next one. If you'd like help, I run this format as a working session in the Ongoing AI Advisor: 90 minutes on a real task from your firm, with a recap of the requests and steps we used. It starts after an AI Adoption Review, which finds the tasks worth training on first.