I want to tell you about the worst AI training I have ever seen. A UK firm, around 60 staff, booked a full-day AI workshop for the whole company. External trainer, big room, pastries. The trainer spent the morning on the history of AI, then the afternoon demoing five tools nobody at that company would ever use. Everyone clapped politely, went home, and changed nothing about how they worked. The invoice was north of £5,000. The value delivered was zero.
I have seen that pattern repeat at every size of business, from five-person agencies to enterprises with proper L&D budgets. And the reason is almost never the trainer. The reason is the premise: that AI skill is a subject you can teach in a room, separate from the work.
The Evidence Finally Caught Up
For years this was just my opinion, formed by sitting through enough workshops to develop a twitch. Now the evidence is in, and it says the same thing.
In June 2026 the Department for Work and Pensions and Skills England published a research programme called Skills for AI. It drew on 23 workshops with around 150 organisations, ten case studies, and a UK employer survey with 536 responses. The headline finding: over 44% of organisations report their staff use AI tools daily, but staff mostly learned through trial and error, peer support, online videos, or the prompts built into the tools themselves. That creates what the researchers call uneven and risky practice. Your people are already using AI. Nobody taught them how.
Two more findings from that survey worth pinning to the wall:
Nearly all organisations, 97%, report providing some AI training. Yet the same employers identified big gaps in what that training achieves: 51% said provision lacked flexibility, and 34% said it was not practical or contextualised. Access is no longer the problem. Effectiveness is.
And the biggest barrier employers cited to staff actually participating was not scepticism about AI. It was cost at 42% and limited availability at 37%. Capacity, not interest.
The report’s conclusion reads like something a contractor would say after seeing that £5,000 workshop: good AI training is practical, task-based, and built into the work itself. They packaged it as six principles called PRIMES: practical, reachable, integrated, modular, expandable, sustainable. I would summarise it more bluntly. Teach people the tasks they actually do, in short sessions, at work, and keep it updated. That is the whole framework.
Why Almost All AI Training Fails the Same Way
Every failed training programme I have seen makes the same three mistakes.
Mistake one: it teaches tools instead of tasks. “An introduction to ChatGPT and Copilot” is the default AI course in this country, and it is mostly worthless. Tools change every few months. The task your operations manager does every Tuesday does not. When training starts from the tool, people leave knowing what a prompt box looks like. When training starts from the task, they leave knowing how to cut three hours off the quote turnaround process. The Skills for AI survey found 67% of employers still struggle to develop technical AI skills through training, and I would bet the gap is wider for the non-technical majority, because the courses on offer assume everyone wants to become a prompt engineer.
Mistake two: it is an event, not a habit. A one-day workshop is a box-ticking exercise. The half-life of AI expertise is short, roughly 18 to 24 months at current pace, and that applies to your staff as much as to specialists. Training that happens once a year is stale within three months. The employers getting results run short sessions, 30 to 90 minutes, tied to real work, repeated often. The DWP research is explicit about this: modular, stackable learning beats longer courses, especially for SMEs and frontline staff who cannot disappear for a day.
Mistake three: it ignores what people already do. Your team is already pasting things into chatbots. They learned by trial and error, and that means some of them are pasting customer data into public tools right now. Training that starts from “let’s learn AI” ignores the risky habits already in place. Training that starts from “show me what you already do with it” finds them in an afternoon.
The Developer Version of the Same Story
Here is my favourite piece of corroborating evidence, because it comes from the group most saturated with AI training and tooling: software developers.
The 2026 Stack Overflow Developer Survey, over 49,000 respondents, found AI coding tool adoption at a record 84%. Trust in those same tools hit an all-time low: only 29% trust the output to be accurate, down from 40% in 2024, and just 3% “highly trust” AI-generated code. The top frustration, cited by 66%, was AI solutions that are almost right but not quite. Developers have absorbed the training, use the tools all day, and still do not trust the output.
That is not a failure of training. That is what competence in AI looks like: fluency plus scepticism. The most skilled AI users in the world treat AI output as a draft that needs review. If your training is producing staff who blindly trust the output, or staff who refuse to touch it, you have failed in opposite directions.
The lesson for an SME owner is that the goal of AI training is not confidence in AI. It is calibrated judgement about when AI helps and when it is a liability. That is much harder to teach in a workshop, and much easier to teach alongside real work.
What Actually Works: A Cheap, Honest Pattern
If you run an SME and want your team using AI properly without burning £5,000 on a workshop, here is the pattern I recommend. It is close to what the DWP research landed on, minus the acronyms.
Pick three real tasks, not thirty. Find the repetitive, text-heavy, low-risk tasks in each role. Drafting quote responses, summarising meeting notes, tidying up product descriptions. Three per team is plenty to start.
Run short sessions on those tasks only. 45 minutes, in the working environment, using the company’s actual data and actual systems (with sensible rules about what goes into which tool). One task per session. Everyone leaves with something they did, not something they watched.
Make the best current user the teacher. Every SME I have ever walked into has one person who has quietly got very good at using AI in their role. Give them an hour a week to share what they do. Peer learning is already how your staff learned to use AI badly, per the DWP findings, so use the same channel to fix it.
Write down the rules. Two pages maximum. What data can go into which tools, what must never be pasted anywhere, who checks output before it goes to a customer. This is the bit nobody does and everybody needs.
Revisit every quarter. Tools change, tasks change, people change roles. A training programme is a habit, not a document.
None of that needs an external trainer. If you do want outside help, buy someone in to facilitate those sessions and pressure-test your two-page rules, not to give your team a history of artificial intelligence over pastries.
The Awkward Commercial Bit
There is an uncomfortable footnote to all this. Some of the research I have cited, particularly from commercial sources, also claims that structured AI training yields a 14-hour-per-week productivity boost, or that every £1 spent on licences should be matched with £0.30 on training. I would treat numbers like those with suspicion. They come from surveys of self-reported expectations, not measured outcomes. When a number sounds too good, it usually is.
The claim I am comfortable defending is the one the government’s own research supports: your staff are already using AI daily, learning informally, and developing uneven and occasionally risky habits. The cheapest useful thing you can do is direct that learning at real tasks with a few written rules, repeatedly, in the flow of work. That is a 1990s insight about training that happens to be true for AI, and it will still be true when today’s tools are landfill.
I have sat through a lot of AI training. The sessions that worked were the ones where I left having done my actual job faster. The sessions that wasted money were the ones where I left knowing the definition of machine learning. Your team is no different.