AI training for employees: what works, and why most of it fades

Why the usual AI course fades within a month

The typical corporate AI course is a recorded module or a webinar. It explains what a large language model is, shows a few impressive demos, and ends with a quiz. People finish it, feel briefly informed, and go back to the same inbox on Monday. Nothing in their actual week has been touched. Two weeks later the demo is a vague memory and the habit never formed.

The failure is structural, not a matter of effort. A demo shows what the tool can do for someone else. It does not show a credit controller what the tool does for the month-end chase, or a bid writer what it does for the Thursday tender. Transfer from a generic example to a specific job is the hard part of any training, and generic AI courses leave it to the learner.

The volume of training matters too. BCG’s AI at Work 2025 survey of more than 10,000 employees across eleven countries found that regular use rises sharply once people have had at least five hours of training, and that in-person coaching improves the result further. Only about a third of respondents said they had received adequate training at all. A 45-minute webinar is not five hours, and it is not coaching.

It is an organisational problem before it is a skills problem

Microsoft’s 2026 Work Trend Index, a survey of 20,000 workers using AI in ten countries including the UK, found that organisational factors accounted for more than twice the reported impact of individual ones (67% against 32%). Only 26% of the people surveyed said their leadership was clearly and consistently aligned on AI. BCG saw the same pattern from the other side: positive sentiment about the tools rose from 15% to 55% where leadership support was strong, and only about a quarter of frontline staff said they had that support.

Read those two findings together and the design of good training follows. It has to involve the manager as well as the team. It has to settle the questions that stop people using the tool at all, which are almost never about prompting and almost always about permission: can I put this client’s name in, is this covered by our confidentiality terms, who checks the output. And it has to leave something behind that the organisation owns, not a certificate that the individual owns.

What works: five things a session has to do

1. 1

   Start from a survey of the people attending

   Before the day, ask each person what they spend their week on, which tasks they repeat, which assistant they have access to, and what they are worried about. Ten minutes per person. The answers become the exercises. Nobody is taught on a demo document when their own board paper is available.

2. 2

   Build on real tasks, in the tool they actually have

   If the organisation has licensed Microsoft 365 Copilot, the session runs in Outlook, Word, Excel and Teams. If it is ChatGPT Team or Claude, it runs there, with the team’s real documents where the data rules allow. The first workflow is built in the first hour, tested on a real task, and written down before anyone moves on.

3. 3

   Agree the guardrails in the room

   What can go into the tool, what never can, and who signs off the output. One page, written by the team with the manager present, not a policy emailed afterwards. This is the step that turns quiet non-use into confident use, because it answers the question people were too polite to ask.

4. 4

   Leave a library the team owns

   Every workflow built on the day is captured where the team will find it: the prompt, what to attach, what good output looks like, and who owns it. Not a slide deck. A working document in the team’s own SharePoint or Drive, added to after the session.

5. 5

   Come back at week four

   A short review call against three numbers agreed up front: minutes saved per person per week, how many of the workflows are still in use, and how many people used the tool this week. Whatever fell away gets looked at honestly. Whatever stuck gets shared.

None of this is exotic. It is how any skill is taught when the teacher expects to be judged on whether it was used afterwards, rather than on whether the room enjoyed the morning.

Who to train first

Not everyone, and not the most senior people first. Pick one team whose week contains a lot of repeated written work: proposals, client updates, reporting, case notes, internal summaries. Sales, marketing, finance, operations and client service teams all qualify. Twelve people is the most a hands-on session can serve properly; beyond that it becomes a lecture.

Train the manager alongside the team, in the same room. Then run a separate, shorter session for the leadership group, because their questions are different: what to fund, what the policy needs to say, and how to tell whether the investment is working. Trying to answer both sets of questions in one room satisfies neither.

• A team of four to twelve with repeated written work and a manager who will attend.
• The assistant the organisation has already licensed, set up and tested the week before.
• A half day to a full day, in person, with laptops open and real documents to hand.
• A named person who will own the workflow library afterwards.

Two things you can do before booking anything

You do not need us for the preparation. Run the survey yourself, then use the assistant you already have to make sense of the answers and to draft the first version of the team’s rules. Both prompts below work in Copilot, ChatGPT, Gemini or Claude.

You are helping me plan a hands-on AI training session for my team of [number] people in [function, e.g. client services at a property consultancy]. Below are their answers to four questions: what they spend most of their week on, which tasks they repeat, which AI assistant they have access to, and what worries them about using it.

Do three things. First, group the repeated tasks into no more than six workflows, named in plain language, and tell me how many people mentioned each. Second, for each workflow, write one sentence on what a good first AI-assisted version would look like inside [the assistant we have licensed]. Third, list every worry raised, grouped by theme, and mark which ones are about data or confidentiality rather than skill.

Keep it under 500 words. Do not invent tasks that are not in the answers.

[Paste the survey answers here]

Draft a one-page set of working rules for a team of [number] in [function] at a [type of organisation, e.g. UK accountancy firm with 60 staff] using [assistant name] for day-to-day work. The rules must cover: what information may be entered into the tool and what must never be (name specific categories relevant to our work, such as client identities, personal data under UK GDPR, unpublished financials, anything under NDA); how outputs are checked before they leave the team; where finished prompts and workflows are saved; and who to ask when unsure.

Write it as short numbered rules a busy person will actually read, in UK English, with no jargon. Mark any rule that depends on our licence terms or IT settings with [CHECK] so we can confirm it. Do not present this as legal advice.

Here is what we already know about our constraints: [paste anything relevant: confidentiality clauses, IT policy extracts, the tool’s data handling terms]

If you are buying training, what to ask the provider

• Will you survey our people before the day, and will the exercises be built from their answers? If the agenda is fixed before they have met your team, it is a course, not training.
• Does the session run inside the tool we have licensed, on our documents, or on a demo environment?
• Who agrees the guardrails, and when? The right answer is your team, in the room, on the day.
• What do we own afterwards? A library your team can edit is worth more than a certificate.
• What happens at week four? If the answer is nothing, expect the fade this article opened with.
• Who is in the room from your side, and have they used these tools to do real work themselves?

The last question is the one people forget. Someone who has used Copilot to get through their own inbox, or Claude to pull a first draft out of forty pages of notes, teaches differently from someone who has read about it. The difference shows in the first ten minutes.