All AI adoption is Change management

There is an article Robert Schaffer published in the Harvard Business Review in 2017 that deserves more attention than it received. It said something simple: all management is, at its core, change management. If sales need to increase, that is change. If a merger needs to be implemented, that is change. If an internal policy needs to change, that too is change. The idea behind it is that separating “change” from ordinary management is a fiction that creates more problems than it solves.

Nine years on, that sentence is more relevant than ever, describing with surgical precision why the majority of organisations are failing in their adoption of artificial intelligence.

Why most AI adoption initiatives fail to generate real business value

88% of organisations already use artificial intelligence in at least one business function, according to McKinsey. That figure sounds encouraging, it sounds like transformation underway, but BCG, after studying thousands of implementations, found that only 26% of companies have managed to move beyond proof of concept to generate real value. Almost everyone is doing something with AI, but only a very small fraction is doing something that generates impact.

The question that matters, then, is not whether your organisation has adopted AI. If you already have, the question is whether that adoption is generating something real or simply occupying space in quarterly reports. If you are still evaluating whether to do so, the question is whether you have clarity on what conditions you need to have in place before it is worth attempting.

Why AI adoption failures are usually organisational rather than technical

I have analysed many cases and when an AI initiative fails, the most common reaction is to look for the problem in the technology: the model was not good enough, the data was not clean, or the integration with existing systems was too complex. And yes, those problems exist and are legitimate, but the data suggests they are not what determines success or failure.

BCG found that approximately 70% of the obstacles in AI adoption are related to people and processes, not technology. Technical problems, including algorithm quality, account for barely 10% of the problem. Prosci reached a similar conclusion studying over a thousand professionals across different industries: human factors explain nearly two thirds of failures, whilst technical problems account for less than a fifth.

There is something else in the data worth highlighting because it completely inverts the usual logic. Organisations that do generate real value from AI follow a resource allocation rule that most ignore: they assign 10% of their resources to algorithms, 20% to technology and data, and 70% to people and processes, but most organisations do exactly the opposite. They invest in the platform, the vendor, the technical pilot, and then cannot explain why nobody uses what they built.

Why workflow design matters more than employee attitudes

The organisation that has already understood the problem is not technical will often immediately commit the next mistake: assuming the problem is the people. If employees are not using AI, it is because they do not understand it or are afraid of it. The solution, then, is to communicate better, convince them, train them, and measure “adoption” by counting how many people completed the training module.

The most recent academic research on AI adoption in the workplace suggests that logic is inverted. A study published in 2025 found something that should change how we think about this problem: the structural characteristics of the job role predict AI adoption more powerfully than any employee attitude towards the technology. It does not matter whether the employee believes AI is useful or not. What determines whether they will use it is how their work is designed: what level of autonomy they have, what variety of tasks they carry out, whether their daily workflow creates real opportunities to apply the tool.

So, if an employee is not using AI, it may not be because they do not want to, it may be because the design of their work gives them nowhere to fit it in. The organisation that trains and communicates without redesigning the work is treating a structural problem as though it were a problem of attitude. And that produces the same results: pilots that impress leadership but never scale, tools that work technically but that nobody uses, and employees who learn to work around the system rather than with it.

Why traditional change management models fail

And here we arrive at the core of the problem, to see it clearly, we need to ask something that few organisations ask themselves: what mental model are we using to manage this process?

The best-known change management frameworks, from Kotter to Lewin, share something in common: they are designed as projects. There is an initial diagnosis, an implementation phase, and a close-out where the change becomes embedded, that is, there is a beginning and there is an end. Researchers in change management have themselves begun to point out the tension this creates: under current conditions, the refreezing process never ends. Changes are present simultaneously at every level of the organisation, and the models acknowledge that reality but do not resolve it, they remain sequential projects applied to a phenomenon that is continuous.

And here lies the central problem with AI adoption: AI is not a one-off event or a merger with a closing date. It is a technology that evolves constantly, whose use cases are discovered in practice, and whose integration with real work takes time and requires ongoing adjustment. Applying the logic of a change project with a start and an end is applying the wrong model to the wrong phenomenon.

The case McKinsey documents with Lilli, its own internal AI tool, illustrates exactly how it works when managed well. There was no grand launch, nor a one-off adoption workshop followed by a project close-out. There was continuous incorporation into the onboarding of new employees, reminders in regular training sessions, and leaders who incorporated the question “did you ask Lilli?” into their team conversations. Integrating AI into daily workflows moves it from hobby to habit, so adoption was not an event, it was a habit built layer by layer, day by day.

What makes that sustained process possible is a culture where experimenting and getting things wrong carries a managed cost, not a prohibitive one. In sectors where operational risk is high, that does not mean unrestricted freedom to try anything: it means designing bounded spaces where learning can happen without compromising what cannot fail. Organisations that successfully integrate AI are not those that convince their employees to use it, but those that build an environment where trying something new is not penalised by default. That culture is built through daily practice, and that is why it is a condition of continuous change, not a step inside a project.

But this is not only a problem of the wrong change model

Here is the conclusion that the three previous mistakes prepare for, and which few organisations nonetheless articulate clearly.

Consider what happens when an organisation launches AI initiatives without a precise answer to this question: what specific value do we want to capture, how will we know whether it is happening, and what needs to change in how we work for it to be possible? What happens is predictable, the tools accumulate, activity is measured instead of impact, reports track how many employees completed the training, how many pilots are active, how many licences were purchased. And when someone asks whether AI is generating value, nobody can answer with precision because nobody defined what it would mean for it to do so. In 2025, 42% of companies are abandoning their AI initiatives, compared to 17% the year before. That is not the number of organisations that had bad luck with the technology, it is the number of organisations that executed without criterion.

Execution reflects what criterion allows and if criterion does not exist, execution has no direction in which to fail well. Criterion cannot be compensated for with better change management, more workflow redesign, or a more open culture of experimentation. If you do not know what you are looking for, none of those tools will help you find it.

This criterion does not only operate at the outset, it must be alive in every decision, because AI evolves, contexts change, and what was a reasonable bet in January may need revisiting in July. The capacity to evaluate continuously, to distinguish what is generating value from what is simply active, is what makes AI adoption something manageable. Without that capacity, the three previous mistakes are not avoidable: they are committed by default.

So, AI adoption is not a technical project, not a problem of attitude, not an event with a closing date, it is simply, management, and it always was. Schaffer said as much in 2017 about change in general, BCG confirms it in 2024, naming change management as the most important capability for succeeding with AI.

The question is not whether your organisation has AI or is planning to adopt it. The question is whether it knows how to manage it for what it is: a continuous process that requires criterion before tools, design before training, and the capacity for permanent evaluation above all else.

And it turns out that is not a new question, it is the same one that management has spent decades avoiding answering well.

How to diagnose where AI adoption is failing in your organisation

If you recognise any of these three mistakes in your organisation, the next step is not another programme, it is knowing with precision which of them is operating, and where.

To that end, I have developed two diagnostic instruments (link here). The first evaluates whether AI adoption in your organisation is generating real value or accumulating activity without measurable impact. The second evaluates whether the organisation has the structural conditions to capture that value. Together, they answer the question this article leaves open: not only why it fails, but exactly where it is failing in your case. They are the starting point for a conversation grounded in evidence, not assumptions.