Train staff on the tools you actually chose

What it is

Give everyone who uses an AI tool at work short, practical training. The training covers three things:

What the tool is. It writes the words most likely to come next. It does not look facts up unless it is built to, and it can state false things with confidence.
How to use this tool for our tasks. Worked through on real examples from the team’s own work.
Our rules. What must not go in, what must be checked, and who to ask.

General talks about AI do not do this job. Neither does general training on IT or privacy.

What the evidence shows

The documented cases listed under “Company cases” below, one evaluation and one experiment are the basis for this advice.

Official surveys show how wide the training gap is.

Lack of skills is the main reason firms hold back. Among firms in the European Union that considered AI and did not use it, 70.9% said they lacked the expertise, according to Eurostat.
Few firms train most of their staff. In the United Kingdom, only 11% of firms with 10 or more employees had trained more than half their workforce, according to the Office for National Statistics.
The smallest firms train least. In Canada, among firms using AI, 24.0% of firms with 1–4 employees had trained existing staff, against 68.1% of firms with 100 or more, according to Statistics Canada.
People who were never shown the tool misread it. A lawyer at a New York firm had not used ChatGPT before. He told the court he assumed it was “like a super search engine”. He filed six decisions ChatGPT had made up.
General training did not transfer. At a government department in Australia, staff had e-learning on privacy and security. The regulator found no evidence of any attempt to train staff on how generative AI tools work, and said staff could not be expected to work that out from general material.
Staff acted on output they did not understand. At Rite Aid, the regulator said staff were not trained or warned that facial recognition alerts could be false. Staff accused shoppers on the strength of an alert.
Hands-on practice worked better than formal sessions. In a UK government pilot with 1,000 licences, training made a clear difference to how satisfied users were. Staff who taught themselves were more satisfied than those who attended the department’s formal sessions.
A small team was won over by trying it. At a small online shop, staff hesitated at first. The founder ran hands-on sessions so that everyone could try the tool.
Knowing how AI works changes what people hand over. In an experiment with 111 participants, people who were taught how AI works gave it the tasks that suited it and kept the ones that suited them. Their results improved.
In the European Union, supporting staff’s AI literacy is a legal duty. The AI Act says organizations that use AI “shall take measures to support the development of AI literacy of their staff”, including others who use AI systems on their behalf. The duty has applied since 2 February 2025. The text adds that the duty does not require a guarantee of “any specific level of AI literacy of any individual”.

Training is not enough alone. One large law firm had trained its staff, and a lawyer still filed eight made-up cases. Pair training with a check before work goes out.

How to do it

Train on the tool you bought. Use its real screens, not slides about AI in general.
Start with what the tool gets wrong. Show two or three real errors from your own trial: a made-up source, a wrong figure, a confident wrong answer. People remember these.
Use your own tasks. Take three tasks the team does every week and work through them together.
Cover the rules in ten minutes. Go through your one-page AI-use policy: what stays out, what gets checked.
Give practice time. Block an hour a week for the first month. Ask each person to bring one thing that worked and one that did not.
Name a go-to colleague in each team for questions.
For staff still learning the job, set the tool to give hints, not finished answers. In a randomised trial with nearly 1,000 secondary-school students in Turkey, students who practised with unrestricted GPT-4 did worse on the exam without it than students who never had AI. A version limited to hints avoided that loss.
Include everyone who touches the tool: managers, new joiners, temporary staff, contractors.
Check what stuck. A month later, give a short test with right and wrong answers. See test AI knowledge instead of asking staff to rate their own.
Repeat when the tool changes and for every new joiner.

What to measure

The share of users trained, including temporary staff.
Scores on a short knowledge test a month after training.
The share of licence holders who use the tool weekly, split by group (role, age, gender), to see who the training is missing. In a review of 76 sources covering more than 300,000 people, 39.3% of women used generative AI against 47.8% of men, and the gap has stopped narrowing. The authors report that training raised women’s use in a UK pilot.
Errors caught at the checking step, and whether they fall over time.

Where it fails

One talk, then nothing. Skills fade and the tool changes.
Training with no practice time. In the UK pilot the formal sessions did less than people’s own practice.
Only junior staff attend. In one law-firm case, the person who misused the tool was a partner.
Training without a check. A trained person under time pressure can still skip the rule.

By company size

Micro and small: one hands-on session led by whoever knows the tool best, then a standing ten minutes in the weekly meeting.
Medium: a session per team on that team’s tasks, and a named go-to colleague in each.
Mid-market and enterprise: training by role, a record of who has done it, and a short test. Temporary staff and contractors are covered before they get access.

Assessments for this practice

These are existing questionnaires, tests and guides you can use. Each card gives the question it answers, who answers it and whether you can use it freely. All assessments.

Company cases

Companies that did this

What happened without it

Companies that came round to it

Sources

Pinski, Adam and Benlian (2023). AI Knowledge: Improving AI Delegation through Human Enablement. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. What this study did and found
Bastani, Bastani, Sungu, Ge, Kabakcı and Mariman (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. What this study did and found
European Union. AI Act, Regulation (EU) 2024/1689, Article 4 (AI literacy), consolidated text of 27 July 2026. European Commission’s AI Act Service Desk. ai-act-service-desk.ec.europa.eu
UK Department for Business and Trade (August 2025). The evaluation of the M365 Copilot pilot in the Department for Business and Trade. assets.publishing.service.gov.uk
Office of the Victorian Information Commissioner (September 2024). Investigation into the use of ChatGPT by a Child Protection worker. ovic.vic.gov.au
Eurostat (December 2025). Use of artificial intelligence in enterprises. Statistics Explained. ec.europa.eu
Office for National Statistics (2026). Artificial intelligence in UK businesses: 2023 to 2026. ons.gov.uk
Statistics Canada (2026). Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026. 11-621-M2026010. www150.statcan.gc.ca
Cranney, Delecourt and Koning (updated August 2026). Global Evidence on Gender Gaps and Generative AI Over Time. Harvard Business School working paper; summary page. aiinstitute.hbs.edu