Why One-Off AI Training Doesn't Work, and What to Do Instead

One-off AI training rarely changes how people work, because one session on generic examples can't give people the regular practice that makes AI useful. What works is practice over weeks on people's own work, with AI champions in each team, a real project to learn on, and managers who ask about it.

If you've paid for AI training and seen little change in how your team works, this piece explains why. It covers what a new survey from Anthropic shows about who gets value from AI, what UK research says works, and five questions to ask a training provider before you sign.

What Anthropic's survey found

On 16 October 2026 Anthropic published The AI Expectations Gap, a survey of 10,980 US adults, run by Morning Consult in August, and 1,334 of Anthropic's own staff. Anthropic staff expect AI to be able to do 75% of their own work by 2030. US workers put it at 28%.

The two groups largely agree on what AI can do today. Where they split is on what it does for them, and that tracks how much they use it.

Asked how long a task AI is well suited to takes with AI, compared with doing it without, 59% of Anthropic staff say a quarter of the time or less. Among US adults who use AI at work, 26% say the same. And 29% of those US users say the task takes just as long or longer with AI, once they count the time spent checking and fixing its work. Only 5% of Anthropic staff say that.

The same pattern shows up in what people expect. About a fifth of the US sample use AI at work regularly. Of those, 56% expect AI to manage at least seven of eight test tasks by 2030, from routine emails to a Nobel-level discovery. Among all US adults it's 41%, and among people who've never used an AI tool, 28%. Anthropic staff, who build and use AI every day, are at 96%.

The report's own conclusion is careful. Expectations may shift as AI use spreads, and how far depends on whether AI comes to do as much of the work elsewhere as it does inside an AI lab.

Two caveats before going further. This is American data, and the staff poll was internal and unweighted: people who build AI for a living were always going to expect a lot from it. And I'm part of Anthropic's Claude for Startups programme, so read my take with that in mind.

Why regular use matters more than training

Strip it back and the survey says something every trainer already knows: the more you practise something, the more you get out of it, whether it's driving, a language or a spreadsheet. AI works the same way. The people who see the time savings are the people who use it every day, on real work, and have learned where it helps and where it doesn't.

That's a problem in the UK, because most people haven't had the chance. Deloitte's GenAI Workforce Survey, which asked 25,000 UK workers in May and June 2026, found that half of the people using generative AI at work have had no formal training at all, and only 20% describe themselves as highly proficient. Just over half the UK workforce, 51%, said training would encourage them to use it more often.

The Office for National Statistics sees the same thing from the business side. Only 10% of firms with ten or more staff that use AI say they use it extensively.

Why one-off AI training doesn't work

Most AI training is sold as sheep dip. Everyone goes through the same session once, comes out the other side with a certificate and a list of prompt tips, and goes back to a job that hasn't changed. Nobody follows up or checks whether anything is different a month later.

It's everywhere because it's easy to buy, deliver and report on, not because it works. A half-day workshop gives people a look at the tool. It doesn't give them the hours of use on their own work that the survey shows makes the difference.

The research on what does work points the same way. Boston Consulting Group's AI at Work survey of more than 10,600 people, published in June 2025, found that regular use is sharply higher among employees who get at least five hours of training, and higher again where they have in-person training and coaching. It also found frontline staff are more likely to use AI regularly when leaders visibly back it. A year on, BCG's 2026 edition found only 36% of employees feel they've had enough training.

UK government research says the same. What works for AI upskilling in the UK, published by the Department for Work and Pensions with Skills England in July 2026, puts it plainly: "AI training needs to last longer than a single session." It recommends revisiting training after three and six months, and found that "Many SMEs struggle to move from pilot use of AI to wider adoption."

What works instead: champions, projects and practice

Three things, run together, over weeks rather than a day.

1. Champion networks. A named person in each team who uses AI on their own work, has time set aside to help colleagues, and meets the other champions regularly to swap what's working. They're the colleague you can lean over to and ask, "how would you do this?", not a helpdesk. When Deloitte asked UK workers what support they wanted, access to internal AI champions sat alongside mentoring from experienced colleagues and knowledge-sharing forums. A champion network is how practice spreads from the keen few to everyone else.

The evidence for champions comes mostly from case studies rather than controlled trials, but it's consistent. The UK government's July 2026 report found that "Coaching and mentoring from local champions help learners who lack confidence or digital experience." In its case studies, Roche appointed AI champions within each function "to sustain momentum beyond the initial rollout", and in its regulatory team, voluntary fortnightly sessions where colleagues showed how they used AI drew around 150 employees each time. KPMG's lesson was that "Champions are essential to providing bottom-up specific use cases and support in the flow of work." Both are the companies' own accounts rather than independent audits.

2. Project-based learning. Learn on a real piece of work with a real outcome. A team picks one process, say the monthly report, meeting follow-ups or the procedures nobody has written down, and rebuilds it with AI over a few weeks. By the end, the process works differently and the team has had weeks of practice on it. The strategy piece covers how to choose which processes to start with.

3. Experiential practice, built into the week. Short, regular practice on people's own tasks, followed up at two, four and eight weeks. Managers ask one question in their regular check-ins: what did you use AI for this week? That question does more than any certificate, because it makes use normal and shows you who needs help. Gallup's survey of 19,043 US workers in May 2025 found that, in organisations investing in AI, employees who strongly agree their manager supports AI use are 2.1 times as likely to use it a few times a week or more. Only 28% of employees in organisations that have started bringing in AI strongly agreed their manager actively supports it.

Five questions to ask an AI training provider

Before you sign anything, ask these. A good provider will have the answers ready.

1. What happens after the session? If the answer is a slide deck and a feedback form, it's sheep dip.
2. Whose work do people practise on? It should be theirs, not a made-up case study.
3. How do you set up and support champions? Who picks them, how much time they get, and who looks after them.
4. What do you measure, and when? Look for use and time saved at four to eight weeks, not satisfaction on the day.
5. What do managers have to do? If the answer is nothing, it won't stick.

There's more on telling providers apart in how to choose an AI training provider, and on why so many leaders have stopped trusting the sector in Everyone Needs AI Training. Nobody Trusts the Consultants.

FAQ

Why doesn't AI training work?

Mostly because it's a one-off. People get one session on generic examples, then go back to work that hasn't changed, with no follow-up. Getting good with AI takes regular use on your own work, which a single session can't give.

What is sheep-dip training?

Training where everyone goes through the same one-off session and then goes back to work that hasn't changed, with no follow-up. It's quick to buy, deliver and report on, but it rarely changes how people work.

Does AI training actually work?

It can, but the amount and the follow-up matter. Boston Consulting Group found regular AI use is sharply higher among employees who get at least five hours of training, and higher again where they have in-person training and coaching. Anthropic's 2026 survey found the people who use AI most report the biggest time savings.

What does an AI champion do?

A champion uses AI on their own work, has time set aside to help colleagues in their team, and meets other champions to share what's working. They're the colleague people can ask, not a helpdesk.

How long should AI training last?

Think weeks, not hours. A programme that runs over four to eight weeks on people's own work, with follow-up and managers asking about it, does more than a single day. BCG found regular use sharply higher among people who'd had at least five hours of training, and a July 2026 UK government report recommends revisiting training after three and six months.

Where this fits

The strategy says what you're using AI for. The usage policy says what staff may do. The governance keeps it all current. None of it changes anything until people use AI often enough to get good at it.

That's why I talk about training and embedding, not training on its own. If you'd like to talk about what that looks like for your team, get in touch.