At the start of an engagement we usually ask an organisation’s employees a plain question: What would help you most, over the next three months, to make better use of AI in your work? I expected the answers to be about skill. Better prompting. A proper course. A better tool.
Some of them are. But in organisation after organisation, two answers sit among the most common, and neither is about skill at all. Tell me what data I am allowed to put where. And: tell me which of my tasks I am allowed to do with it.
That is worth pausing on, because most of these organisations have already bought the training. The licences are in place, the workshops have run, the prompting guides are on the intranet. The people answering are able to use the tool. What they are waiting for is permission, and permission is not something you can train into an individual. It is something the organisation has to decide.
This is the gap most AI programmes fall into. They stall not on technology, and not on the skills of the people using it, but on an organisational learning gap: the inability of the organisation, as an organisation, to work in a new way. Which is why buying more training changes far less than it costs.
Employees are already using AI
Adoption is the one thing you do not need to fix. MIT’s GenAI Divide found that only about 40% of companies hold official large-language-model subscriptions, while in more than 90% of companies employees are already using AI tools on their own. Yet the same report puts the share of generative-AI pilots that show no measurable P&L impact at roughly 95%, and locates the binding constraint not in infrastructure, regulation or talent but in learning: most systems do not retain feedback, adapt to context or improve over time. 
MIT was describing systems. My reading, from the leadership teams I sit with, is that the diagnosis applies at least as well to organisations. The individual learns the tool in an afternoon. The organisation, which has to learn to do the work differently, may have learned nothing a year later. Training addresses the first kind of learning and leaves the second untouched.
Nobody can send an organisation on a course
Where does the value actually come from? The clearest external answer is McKinsey’s: of twenty-five organisational attributes tested, the redesign of workflows had the biggest effect on a company’s ability to see EBIT impact from generative AI. Only 21% said they had fundamentally redesigned even some workflows. The data is self-reported and correlational, so read it as direction rather than proof. But the direction is unambiguous, and it is the opposite of where the training budget goes.
Redesigning a workflow is not an individual act. It means deciding, together, who does what, in which order, with which information, and who is allowed to decide what. It changes handovers, approvals and the definition of done. No course teaches that, because it is not a skill anyone holds alone. It is a decision the organisation makes and then keeps.
That is the practical meaning of organisational learning. An organisation has learned something not when its people know it, but when the way the work is done has changed and stays changed. As I often put it to executive teams, an organisation is not what it decides. It is what it repeats. A training day changes what people know. It rarely changes what the organisation repeats.
The missing substrate: permission and safety
Which brings us back to the two answers from the survey. Why would capable people, holding a licensed tool they know how to use, ask to be told what they are allowed to do?
Because using AI at work without a written boundary is a private risk assessment, made alone, every time. Is this customer’s data safe to paste in? Will anyone hold it against me if the draft goes wrong? People then do one of two things. They abstain, and the organisation loses the value. Or they use the tool quietly, on whatever they happen to have, and the organisation carries a risk it never chose. Without governance the benefits of AI go to the individual and the risks to the organisation. Neither outcome is what the training was bought for.
The research on learning has been making this point for a quarter of a century. Amy Edmondson defined psychological safety as a team’s shared belief that admitting a mistake, asking a question or disagreeing will not be punished, and showed it to be a precondition for learning behaviour in teams. When Google later studied hundreds of its own teams, the strongest differentiator of the best ones was not intelligence, experience or composition but the same thing. And the fear that keeps people from asking is both real and misplaced: Brooks, Gino and Schweitzer found that people avoid asking for advice because they expect to look incompetent, when in fact those who ask are judged more competent, especially on hard tasks.
Clear rules about data and tasks are psychological safety made concrete. A boundary that says you may do this, up to here is an act of leadership; it turns every private risk assessment into a shared one, made once. An absent rule is not neutral. In most organisations it is read as you may not, and the more thoroughly instructed the culture, the more literally it is read. I have seen situations where every task has a written procedure, leadership openly wishes people would use more judgement, and a tool whose entire value lies in open-ended questioning sits unused — not because anyone found it difficult, but because nobody had given permission. The root cause there was never a skills deficit. It was a safety deficit, and a second training day would have deepened it.
The fair objection: surely skills still matter
They do, and the point is not to cancel the training. It is to stop expecting it to carry weight it cannot bear. Training is the floor of an AI programme, not its lever. Ordered correctly, it comes last: first the boundaries, which cost a page of writing and a leadership signature; then the redesign of the work, which the people who run it have to do together; and only then training, aimed at the redesigned workflow rather than at generic prompting. Training that precedes permission teaches people to do, faster, work they are not allowed to change.
The second objection is that rules will strangle experimentation. In my experience the opposite is true. What strangles experimentation is ambiguity. A clear boundary is what makes everything inside it safe to try, and rules written as a list of prohibitions do not stop use; they push it underground, where nothing is learned. Write the rule as a permission with limits, and experimentation goes up, not down.
Start here, this quarter
1. Answer the two questions in writing. What data may be put into which tools, and which tasks may be done with them. One page, phrased as permissions with limits, signed by leadership rather than IT. It is the cheapest intervention you have, and it removes the obstacle most organisations are trying to remove with licences and courses.
2. Ask your own people the three-month question. What would help you most, over the next three months, to make better use of AI in your work? Then listen for the difference between requests for skill and requests for permission. The ratio tells you which gap you actually have.
3. Redesign one workflow as a group, not one person at a time. Take a process that matters, put the people who run it in a room, rebuild it with AI inside, and settle the decision rights. Then measure the work itself, not the usage. One redesigned workflow teaches the organisation more than a hundred trained individuals.
4. Say out loud what a good failure looks like. Learning behaviour follows safety, not instruction. If leadership names the kind of mistake it expects, and will not punish, while people experiment inside the boundaries, the first mistakes become the curriculum instead of the reason the programme stopped.
The training was never the point
The uncomfortable thing about the two survey answers is how cheap they are to resolve, and how long they have gone unresolved in organisations that have spent real money on everything around them. People are not waiting to be taught. They are waiting to be told they may, and to be sure they will not be alone with the consequences when they try.
An organisation learns the way a team learns: when it is safe to ask, safe to be wrong, and clear where the edges are. Those are not training outcomes. They are leadership decisions, and until they are made, the most sophisticated tool in the building will keep doing what it does today: making individuals slightly faster at work the organisation has not yet learned to do differently.
About the Author
For nearly three decades, Lenni Laukkanen has helped organisations develop and grow through technology. He argues that AI is not a technology question but a leadership one: it rarely fails as a tool — it exposes ways of operating built for a world that no longer exists.
As founder of Astu Labs, trusted advisor to executive teams and keynote speaker, he helps leaders navigate the organisational changes required to capture value from AI. A runner, cyclist and adventurer, he holds that organisations, like athletes, change one deliberate step at a time. More at lennilaukkanen.fi.
*The views and opinions expressed by the author do not necessarily state or reflect the views or positions of Hyperight.com or any entities they represent.
