Why AI Training Doesn't Stick in Large Organisations

There is a number from the Microsoft Copilot rollouts that should stop any transformation lead in their tracks. When Gartner studied enterprise deployments, 90% of the people given the assistant said they would fight to keep it, and at the same time 72% said they struggled to fit it into their daily routine (Gartner, via Computerworld, 2024). Read those two facts together. People love the tool and cannot make it part of how they work. Only six percent of the organisations studied had finished a pilot and were planning a real scale-up. The licences went out. The behaviour stayed put.

This is not a Microsoft story; it is the shape of enterprise AI everywhere. MIT's NANDA initiative looked across hundreds of generative-AI deployments and found that around 95% of pilots delivered no measurable impact on the profit-and-loss statement, with only about five percent producing rapid acceleration (MIT NANDA, 2025). Their diagnosis is worth holding onto: the failure is not the quality of the models, it is the approach. Generic tools, they note, do not learn from or adapt to how people actually work. The technology arrived. The way of working did not move to meet it.

Faced with that gap, most organisations reach for the same instrument: training. More courses, more literacy modules, more prompt libraries. It is the reasonable response, and it mostly does not work. Understanding why is the difference between a programme that produces a completion certificate and one that produces a workforce that decides differently.

There is a layer underneath the keystroke that no feature tour reaches, and it is the layer that decides whether AI is worth anything. It is the habit of how a person engages the machine in the first place. Most people use AI like a vending machine. They have a task, they type a request, they take what comes back, and they repeat. No expert framing, no challenge to the request, no examination of whether the task was the right one. The machine does exactly what it is told, and what it is told is almost always the symptom rather than the cause, because the thinking behind the request was never examined.

The small minority who get real value operate a different default. They treat the machine as a thinking partner. They give it a specific expert role. They state the real goal rather than the surface task. And, the move that does most of the work, they make the machine question them before it answers, surfacing what they have not specified and challenging whether the task is even the right one. We call that move the Flip, and it is the hinge of the whole thing. Most people instruct AI to produce. The Flip instructs it to interrogate the request first.

Training shows people the buttons. It almost never changes the question they ask before they press one.

This is why training fails to change behaviour even when it is well delivered. The vending-machine habit is invisible to a feature curriculum, because the curriculum is about the tool and the habit is about the person. You can teach someone every capability of an assistant and leave the way they reason with it completely untouched. The gap is not technical. It is cognitive, and a course aimed at the tool will keep missing it.

We have a name for the mechanism underneath this: the Zero-Barrier Trap. The barrier to entry for AI is zero, so everyone in the organisation was already using the machine before any course ran, untrained, in vending-machine mode. A feature course therefore teaches people what they had already worked out by clicking around, which is why it feels redundant in the room and changes nothing afterwards. The barrier that was never zero is the barrier to value: business training, the discipline of reasoning with the machine on real work. Organisations keep clearing the barrier that does not exist and leaving the one that does standing, and that, more than any curriculum detail, is why the training does not stick.

The evidence on what works points the same way the diagnosis does. Harvard Business Impact studied how AI-fluent people actually became fluent and found they were twice as likely to have learned through experimentation rather than instruction. Their conclusion is blunt: without hands-on practice, learning "stays in the realm of theory, important, but inert" (Harvard Business Impact, 2025). Capability is built by doing the real work differently, not by being told about it. The same study found that the binding constraint on AI fluency was a lack of organisational support and reinforcement, not a lack of employee motivation.

That is what the Havruta Methodology is built to install, and it answers the two failures directly with two pillars. The Cognitive Pillar replaces the feature list with one portable habit, the 4-Lines: give the machine a specific expert persona, state the real goal, invoke the Flip so it questions you before it answers, and run the exchange one step at a time. It is simple enough that a finance analyst, a plant supervisor and a marketer can all run it on their own real work the same afternoon, which is the hands-on practice the evidence calls for.

The Brain Pillar answers the second failure. It replaces the one-off event with persistent, shared context, the verified ground a team reasons from again and again, so the habit is reinforced by the work itself rather than left to memory. This is also where Ground Truth lives, the discipline of anchoring the machine in your verified internal reality instead of its generic average. The two pillars together give people a test they can apply to their own output, what we call the Mirror Principle: if the output is generic, the reasoning was generic. A workforce that has internalised that has stopped needing the course, because it has learned to notice the gap itself.

None of this requires new tools or new budget. It works with whatever AI the organisation already has on its desks. The variable it changes is not what people can access. It is how they reason before they press a key.