Most AI training teaches people to recognise a tool, not to work differently with it. On 10 June 2026, Skills England published new research called Skills for AI: What Works for AI Upskilling in the UK, the SKAI study, led by Dr Nisreen Ameen of Royal Holloway, University of London and drawing on insights from over 150 employers. The headline finding is the one every head of learning should sit with. Around 44% of UK workplaces now use AI every day, and yet, in the words of the accompanying release, adoption remains “uneven and often limited in impact”. Daily use is no longer the problem. Use that changes anything is.
To its credit, the SKAI research does not pretend the answer is more content. It proposes a framework called PRIMES, six principles for AI training that actually lands: Practical, Reachable, Integrated, Modular, Expandable and Sustainable. Read those six words slowly and notice what they have in common. Not one of them is about the slide deck. They are about whether the learning is hands-on, whether people can reach it, whether it sits inside the real job rather than beside it, and whether it survives past the launch week. The research has, in effect, described why a decade of upskilling spend produced 44% daily use and limited impact. The training was none of those things.
Why “We Trained Everyone” And “Nothing Changed” Are The Same Sentence
Walk into most AI upskilling programmes and you find the same shape. An introductory module, a webinar, a library of prompts, a certificate at the end. People click through, score the session well, and go back to their inbox. A month later the daily-use number looks healthy and the behaviour looks identical. The tool is open in a tab. The way the work gets done has not moved.
This is not an AI problem. It is the oldest problem in corporate learning, wearing a new badge. The reason it persists is that almost no programme is measured at the level where behaviour lives. The Kirkpatrick model, the standard framework for evaluating training since the 1950s and reaffirmed by Kirkpatrick and Kirkpatrick in 2016, sets out four levels. Level 1 is reaction, did people enjoy it. Level 2 is learning, did they know more on the way out. Level 3 is behaviour, are they doing the job differently a month later. Level 4 is results, did the outcome move. The overwhelming majority of AI training stops at Level 1, a satisfaction score collected in the room. It tells you the session was pleasant. It tells you nothing about whether a single analyst structured a brief differently the following week.
The Confidence Trap Sitting Inside The 44%
There is a reason satisfaction scores stay high while behaviour stays flat, and AI makes it sharper than ever. People are poor judges of their own competence, and a chatbot that answers fluently makes everyone feel more capable than they are. In Sidestream’s own academic behaviour-change work, building on research from UCL, Cambridge and Bocconi, participants consistently rated their own communication skill well above what trained observers later measured. The confidence was real. The skill was not yet there. Put a confident tool in confident hands and the self-assessment becomes useless. Someone can use AI daily, feel fluent, score the training nine out of ten, and still produce work no better than before. The happy sheet captures the feeling and misses the gap entirely.
This is why “Practical” and “Integrated” sit at the heart of the PRIMES list, and it is also why naming them is the easy half. Roediger and Karpicke showed in Psychological Science in 2006 that being tested on material lifts long-term retention by roughly 50% compared with re-reading it. A webinar and a prompt library are the re-reading kind of learning. They feel productive and fade within weeks. The behaviours an organisation actually needs from AI, knowing when to trust an output and when to challenge it, drafting then editing rather than accepting, judging where the tool helps and where it quietly harms, do not arrive through being shown a demo. They arrive through doing the real task, with feedback, against a standard.
What We Measure Instead
Because real behaviour change happens through lived experience that makes the memory stick, we build for Level 3 from the first conversation. In a Sidestream programme the central method is not a presentation, it is a rehearsal. People do the actual work the training is about, watched and coached, and the same task is run before and after so the change is observed rather than self-reported. We replace the happy sheet with a record of what people did. A team can show you their daily-use statistic. We would rather show you a manager handling the difficult version of the job differently than they did a month ago, because that is the thing a satisfaction average can never prove.
That is the difference between our immersive approach and conventional e-learning. An e-learning module can prove that someone clicked to the end and felt good about it. It cannot prove they behave differently under pressure. A simulation that puts a person in the real task, observed against named behavioural anchors, produces exactly the evidence the SKAI research is pointing at when it warns that adoption stays “limited in impact”. Impact is a behaviour question. You cannot answer it with a completion rate.
The SKAI study is right that AI training needs to be Practical, Reachable, Integrated, Modular, Expandable and Sustainable. The harder question is how you would ever know whether yours is. The only honest test is to measure what people do after the programme, not what they thought of it during. A course can tick all six PRIMES principles on paper and still change nothing, and you will only find out by looking at behaviour. The organisations that get a return on the 44% will be the ones that stop counting logins and start counting the difference in the work.
What We Do About It
Our training and development and immersive simulations are built for the programmes that have to justify themselves, because every programme now has to. If your organisation is somewhere inside that 44%, using AI every day with little to show for it, the question to start with is not which course to buy next. It is how you would prove that the last one changed how anyone actually works.
44% of UK workplaces use AI every day and the impact stays limited. Daily use is not the goal. Changed behaviour is. Measure Kirkpatrick Level 3, or keep counting logins.
The teams that turn AI access into AI advantage will not be the ones with the longest course catalogue. They will be the ones who can show, in observed behaviour, that the work itself got better. Book a call to look at what measuring at that level would mean for your programmes.
Book a free 30-minute diagnostic call → or read about our research-backed approach.
Source: Skills England, Skills for AI: What Works for AI Upskilling in the UK (SKAI), research led by Dr Nisreen Ameen, Royal Holloway, University of London, published 10 June 2026 with the accompanying GOV.UK release “New tools will help employers maximise AI productivity gains”. Daily-use figure of around 44% and the “uneven and often limited in impact” assessment are quoted from that release. Get in touch today. We are Sidestream.