Every organisation I walk into has the same quiet gap. The licences are bought, the pilot is declared a success, and a slide somewhere says adoption is climbing. Then you sit down with the people doing the work and ask how they learned to use the tools. Most of them shrug. They watched a video at lunch, copied a prompt from a colleague, or just kept poking at the thing until it did something useful. Nobody taught them anything.
That shrug now has a number attached to it. Deloitte's GenAI Workforce Survey asked 25,000 UK workers about their experience, with fieldwork running through May and June 2026. Half of the people who use GenAI at work received no formal training at all. Only 20 percent were given training that was mandatory. The rest were left to work it out for themselves.
What the numbers say about who is actually learning
It is tempting to read a statistic like that and conclude that people are resourceful. They are. Self-taught users are often the most curious people in the building, and they have carried the first wave of adoption on their own initiative. But resourcefulness is not the same as capability, and the survey makes the difference plain. Only 20 percent of current users describe themselves as highly proficient. Four in five people who already use these tools regularly do not trust their own skill with them.
Think about what that means for the return you are expecting. A tool that is used at the surface, for a quick summary here and a rewritten email there, produces surface-level gains. The compounding value sits further down, in redesigned workflows, in checking and challenging outputs, in knowing which tasks to hand over and which to keep. That kind of judgement does not appear by osmosis. It comes from structured practice with feedback, which is exactly what half of users never received.
The appetite is there, too. 51 percent of the UK workforce say training would encourage them to use GenAI more. This is not a story about reluctant employees dragging their feet. It is a story about willing employees waiting for someone to show them how. When people say they would use a tool more if they were trained, they are telling you the bottleneck is not motivation. It is enablement.
Where training does happen, it is thin
For the people who did get trained, the format tells its own story. 59 percent learned online, 36 percent in person, and 18 percent through self-service portals. Online modules are cheap to roll out and easy to tick off, which is why they dominate. They are also the format most likely to leave someone able to describe a prompt and unable to write a good one under pressure on a Wednesday afternoon.
I have run enough workshops to know how this plays out. A person completes a forty-minute module, passes the quiz, and goes back to a job where the tool sits in a browser tab they rarely open. Three weeks later the knowledge has faded, because it was never attached to a real task they own. Training that sticks is training that happens against the work itself: their reports, their customer emails, their messy spreadsheets. Generic examples about drafting a poem or planning a holiday do not transfer.
In-person sessions do better, but only when they are built around practice rather than presentation. A room full of people watching a facilitator demo a tool is just a slower version of the video. The value comes when each person brings a live problem, tries an approach, sees it fail, adjusts, and sees a colleague solve the same problem differently. That is where the confidence comes from, and confidence is what converts a licence into a habit.
Your leaders are not helping, and most do not know it
The part of the Deloitte data that I find most telling has nothing to do with courses. It concerns what employees hear from the people above them. 65 percent of users say they do not hear leaders talking about GenAI with a clear understanding. Breaking that down, 30 percent hear incoherent messages, 18 percent hear nothing at all, and 16 percent report that leaders are not listening to them on the subject.
Workers take their cues from the top. If senior people speak about AI in vague, breathless terms, or avoid the subject, staff conclude that it is either a fad or a risk, and they behave accordingly. No training programme can fully compensate for a leadership group that has not done its own homework. I have watched executives sponsor an enterprise rollout while admitting privately that they have never used the tool for a real piece of work. Their teams can tell.
The fix here is not a keynote. It is for leaders to use the tools visibly, on real tasks, and to talk plainly about what worked and what did not. A manager who says, "I tried this on the board paper, it got the structure right and the numbers wrong, so here is how I check it now" does more for adoption than any all-hands announcement. That is a clear understanding, expressed in a way people can copy.
The stigma nobody budgets for
Then there is the number that rarely makes it into a business case. 23 percent of respondents believe there is a stigma attached to using GenAI. Nearly a quarter of workers feel it. They worry that using the tools will make them look lazy, or as though they are cutting corners, or that it will signal that their job can be done by a machine.
Stigma drives behaviour underground. People who feel judged do not share their prompts, do not ask for help, and do not admit when an output was wrong. You end up with a workforce quietly using AI in isolation, unable to learn from each other, and exposed to precisely the errors that open conversation would catch. The self-taught half of your users is not just under-skilled. It is also silent.
You cannot train away stigma with a module, but you can shrink it by changing what is normal. When a team lead shows their own AI-assisted work in a meeting and credits the tool openly, using it stops being a confession. When sharing a good prompt earns recognition rather than suspicion, people start sharing. Culture changes when the visible examples change.
What actually sticks
Putting the survey together with what I see in practice, a pattern emerges for the kind of capability building that lasts. It has four parts, and none of them is exotic.
First, tie learning to real work. Every participant should leave a session having improved something they will use that week, not having completed an exercise built for the room. Second, make it hands-on and social. People learn faster alongside peers who share the same constraints, because they see approaches they would never have tried alone. Third, give managers the skills first, so the messages employees hear are coherent and the examples they see are credible. Fourth, follow up. A single workshop starts a habit; it does not finish one. Short, regular check-ins where people bring a new problem keep the practice alive past the first fortnight.
Notice what this does to the numbers Deloitte reports. Hands-on, work-based training closes the proficiency gap, because people build real skill. Coherent leadership closes the communication gap, because managers know what they are talking about. Open sharing erodes stigma, because using the tools becomes visible and respected. One well-designed programme works on three of the survey's biggest problems at once.
A practical starting point for Monday
If you want to act on this without launching a major initiative, start by finding out where you stand. Ask your people three questions. How did you learn to use these tools? How confident are you, honestly? And do you feel comfortable telling colleagues when you use them? The answers will tell you whether you have a skills problem, a leadership problem, a culture problem, or all three, which is the most likely outcome.
Then pick one team and run a proper pilot. Bring their real tasks into a working session, have their manager attend and participate, and agree a simple follow-up rhythm for the following month. Measure what changes: time saved, quality of output, willingness to share. You will learn more from one honest team experiment than from a dozen vendor dashboards.
At Collective Campus we have been building this kind of practical, work-based capability for organisations for years, and our AI training is designed around precisely these principles: real tasks, live practice, and leaders who go first. But whoever you choose to work with, hold them to the same standard. If a programme cannot show you how people will apply what they learn on their own work the following week, it is a lecture, not training.
The cost of leaving it to chance
The uncomfortable reading of the Deloitte data is that most organisations have outsourced their AI capability strategy to individual curiosity. It has worked well enough to get started, which is why the adoption charts look healthy. But half your users are self-taught, four in five are not highly proficient, and a quarter feel they have to hide what they are doing. That is a fragile foundation for something you are betting competitive advantage on.
The good news is that the demand is already there. Over half your workforce is telling you training would make them use these tools more. They are asking for exactly what you can give them. The organisations that respond with practical, well-led, socially safe learning will pull ahead of those that keep hoping curiosity is enough. It was enough to get the first half of your users moving. It will not be enough to take them the rest of the way.
When the room is still guessing, run a design thinking workshop on a live problem. If you already have a bet, test it in a lean startup product sprint.
