AI Adoption Fails at Culture, Not Tech

The tools are bought, the licences are assigned, and the training session has happened. Yet usage is low, or it is happening quietly where nobody can see it. If that sounds familiar, the problem is probably not the software. AI adoption tends to stall when people are unsure what the change means for their jobs, their standing and their skills, and no one has given them a straight answer.

This blog explains where that hesitation comes from, how to tell different kinds of resistance apart, and what leaders can change in the way work, incentives and conversations are set up. It is written for managers and executives who are past the pilot stage and wondering why the results have not followed.

Why does AI adoption fail when the technology works?

It fails because using AI changes how people work, how they are judged and what they expect from their careers, and a tool does nothing to address any of that. A model can perform well in a pilot while the organisation around it stays the same, with the same targets, the same approval chains and the same unspoken rules about what counts as good work. Employees notice that mismatch quickly, and a culture transformation review that looks at how work really happens is often where the gap first becomes visible.

The numbers point in the same direction. A global study by the University of Melbourne and KPMG, covering more than 48,000 people in 47 countries, found that 58% of employees intentionally use AI at work. Only 47% said they had received AI training, and only 40% said their workplace had a policy or guidance on generative AI use. People are being asked to change how they work with little direction on how to do it safely, and many respond by being cautious or by working privately.

Technical problems do exist. Poor data, weak integration and unclear use cases all slow a rollout. They tend to be visible and fixable, though. The cultural problems sit underneath, are harder to measure, and are often misread as laziness or a lack of skill. That misreading is where most rollouts go wrong, because the fix for a fear is not the same as the fix for a skill gap.

What are employees actually afraid of when they resist AI?

Employees usually fear one of four things: losing their job, looking less capable, losing the value of expertise they spent years building, or being measured by a system they do not understand. These fears are different from each other. Treating them as a single problem called resistance tends to produce general reassurance that convinces no one.

Job loss. Mercer's Global Talent Trends 2026 research reported that employee concern about AI-related job loss rose from 28% in 2024 to 40% in 2026. The concern is widespread, and it does not depend on whether any layoffs have been announced.

Looking less capable. People worry that AI-assisted work will be read as less their own. In four preregistered experiments published in the Proceedings of the National Academy of Sciences, participants expected that using AI would make them look lazier and less competent, and they were less willing to disclose that use to managers. In a separate experiment, employees described as getting help from AI were rated as lazier than those who got help from other sources.

Losing the value of expertise. Experienced staff often build their standing on judgement that took years to develop. When a tool produces a plausible answer in seconds, some of them reasonably ask what their experience is now for.

Being measured. If it is unclear whether AI use is tracked, rewarded or penalised, people tend to assume the least favourable answer and say nothing.

Is it resistance, or a reasonable response to a real problem?

Often it is a reasonable response. Low usage can come from fear, from a valid objection to the tool, or from a skill gap, and all three look identical from the outside. Before choosing a response, it helps to ask what the behaviour is actually signalling.

• People tried the tool and stopped. The likely cause is that the output or the workflow does not fit the task. The first response is to review the use case with the team and fix it or drop it.
• Use is high but nobody talks about it. The likely cause is fear of being judged. The first response is to make open use normal and remove any penalty for it.
• Questions about what happens to roles. The likely cause is job security. The first response is to give a plain answer, even where the answer is incomplete.
• Polite agreement, no change in work. The likely cause is that targets and reviews reward the old way. The first response is to check what people are actually measured on.
• Confident use with frequent errors. The likely cause is a skill gap in checking outputs. The first response is to train people to verify, not only to prompt.

The last pattern matters more than it looks. The Melbourne and KPMG study found that 66% of employees rely on AI output without evaluating its accuracy. Some of what looks like enthusiasm is use without the judgement that makes it safe, so the response there is training in checking, not encouragement.

A useful habit is to ask people directly, in small groups and without managers in the room, what they would lose if AI worked as promised. The answers usually sort themselves into the categories above.

Why do employees hide their AI use?

Employees hide AI use because they expect to be judged for it, and in many workplaces they have reason to. The Melbourne and KPMG study found that 57% of employees hide their use of AI and present AI-generated work as their own. Hidden use is a sign that the culture has taught people that disclosure arries a cost.

The Duke University research in the Proceedings of the National Academy of Sciences adds a detail that points to a lever. In its hiring experiment, the outcome depended on the evaluator's own habits. Managers who used AI rarely tended to favour the candidate who did not use it, while managers who used AI often tended to favour the candidate who used it daily. In other words, how AI use is perceived depends partly on how familiar the person judging it is.

Hidden use has practical costs. Outputs go unchecked, useful methods stay with one person, and the organisation has no reliable picture of where AI is helping or causing errors. The goal is not to monitor harder. It is to make open use ordinary enough that people stop hiding it.

What do leaders do that makes AI adoption harder?

Leaders make it harder when their messages and their systems disagree. They announce AI as a priority while targets, headcount plans and performance reviews stay the same. They talk about efficiency without saying what happens to the time that is freed up. They expect people to learn new tools on top of a full workload.

Mercer's research points to a gap in attention. It found that 62% of employees feel leaders underestimate the emotional and psychological impact of AI, and that only 19% of HR leaders include those impacts in their digital implementation strategy. Fear is being treated as a side effect instead of a design input.

The gap between a sound AI strategy and an organisation able to carry it out is something we have covered in more detail in Your Strategy Is AI-Ready. Your Organisation Isn't., which looks at operating models, workflows and leadership behaviour.

How do you build trust before asking people to change?

Start by being honest about what is known and what is not, and make it safe to experiment and to get things wrong. People will not try a new way of working in front of colleagues if a mistake will be held against them. Honesty about job impact does not require certainty. It requires saying what has been decided, what has not, and when people will hear more.

The research base for this is well established. In a 1999 study in Administrative Science Quarterly, Amy Edmondson found that team psychological safety, a shared belief that taking interpersonal risks is safe, was associated with learning behaviour in work teams, and that learning behaviour linked psychological safety to team performance. Adopting AI is largely a learning problem, so the same conditions apply.

In practice, a few steps help:

• Say plainly which roles are expected to change and which decisions are still open.
• Separate a learning period from any evaluation, so early experiments are not scored.
• Let teams pick their first use cases instead of assigning them.
• Publish simple guidance on acceptable use. Only 40% of employees in the Melbourne and KPMG study said their workplace had one.

What role do middle managers play in AI adoption?

Middle managers decide whether AI use happens in daily work. They set priorities, approve time for learning and shape what counts as good output. If they are unsure, undertrained or worried about their own roles, the team usually reads that signal faster than any leadership announcement.

This is the layer that often receives the least preparation. A mandate arrives from above, questions arrive from the team, and the manager sits between the two without a clear brief or permission to adjust targets. The Duke findings suggest that a manager's own familiarity with AI also shapes how they judge others who use it, which makes personal practice relevant, not optional.

What tends to help is giving managers time to use the tools on their own work first, a short list of questions they can expect and a way to escalate concerns they cannot answer. They also need explicit permission to change targets where AI shifts what a reasonable workload looks like.

How do you redesign work instead of adding another tool?

Redesigning work means choosing a specific task, mapping its steps, deciding where a person checks the output, and adjusting roles and measures to match. Adding a tool to an unchanged process usually adds a step instead of removing one, and people notice when a new tool makes their day longer.

A drafting task shows the difference. If AI produces the first draft but the review stages, sign-offs and formatting rules stay as they were, the team now has a draft to check on top of everything else. If the process is rebuilt so that review focuses on judgement and the repetitive steps are removed, the benefit becomes visible to the people doing the work.

Incentives belong in this redesign. When rewards depend on hours logged or volume produced, people have little reason to change how they work, and they have some reason to avoid any approach that makes their effort less visible. Checking what people are measured on is often a faster diagnostic than any survey.

How can you tell whether the culture is shifting?

Look at behaviour, not licence counts. A change in culture shows up as people talking openly about how they use AI, asking different questions and proposing their own use cases. Login numbers can rise while fear stays exactly where it was.

Some signals worth tracking over a quarter:

• Whether teams share prompts, examples and mistakes without being asked.
• Whether the questions in team meetings move from "will this replace me" to "how do I check this output".
• Whether managers describe their own use of AI without prompting.
• Whether use cases are being proposed by teams instead of assigned to them.
• Whether unsanctioned tools are declining as approved options and clear guidance improve.

A short pulse survey with three or four questions, such as whether people feel safe admitting they used AI, whether they know what is expected of them, and whether their manager has answered their concerns, will usually tell you more than a usage dashboard.

Where should an organisation start?

Start with one team, one workflow and an honest conversation, then widen from there. Trying to shift a whole organisation at once usually repeats the original problem, because the people affected have no say in what changes. A small, well-chosen start gives you evidence to share and fewer assumptions to defend.

1. Listen first. Run short, small-group conversations to find out which fears and objections are present, using the categories above.
2. Say what is known. Share what has been decided about roles and what has not, and when people will hear more.
3. Set simple guidance. Explain what is acceptable, what data should not be used and who to ask.
4. Prepare managers first. Give them practice, a brief and a route for questions they cannot answer.
5. Choose one workflow. Redesign it with the people who do the work, and adjust its measures.
6. Review after a quarter. Look at the behavioural signals above, then decide where to extend.

Successful AI adoption depends on the conditions around the tool at least as much as the tool itself. Where people feel safe, informed and involved, the technology has a fair chance of being used well.