By the time many organisations began writing their first AI policies, employees had already tried a generative AI tool. They had used it to rewrite a difficult email, summarise a document, prepare a presentation or test whether it could complete part of their work.

In many workplaces, people were experimenting before IT had approved a tool, before leaders had decided what responsible use meant and before anyone had considered how the work itself might change.

This does not mean that AI has escaped the usual pattern of technology adoption. The difficult part now appears later, when individual experiments have to become reliable organisational practice.

We have used the S-curve for decades to think about how new technologies spread. Adoption begins with a small number of people who are willing to experiment, gathers pace as the technology becomes more familiar and eventually slows as it reaches most of its potential users.

Everett Rogers’ Diffusion of Innovations describes five groups along that curve: innovators, early adopters, the early majority, the late majority and laggards. The labels are imperfect, but the model helped organisations anticipate that people would adopt a new technology at different speeds.

Geoffrey Moore later introduced the idea of a “chasm” between early adopters and the early majority. Early adopters are often prepared to try an unfinished technology because they are interested in what it might become. The early majority is more cautious. They want to know that it works, that it solves a real problem and that there will be reliable support when something goes wrong. Many technologies attract enthusiastic early users but struggle to make that next step. That gap is the chasm.

Generative AI has not made the chasm disappear. People can move from curiosity to first use within minutes. ChatGPT and similar tools require no installation project, specialist training or organisational approval. The more difficult passage comes later, when scattered experiments have to become reliable working practices.

Three features of generative AI help explain why initial use spread so quickly.

1. It arrived as a consumer tool

ChatGPT was publicly available to anyone with an internet connection. UBS estimated, using Similarweb data, that it reached 100 million monthly active users by January 2023, around two months after its launch.

Comparisons with the adoption of the telephone or the internet are tempting but not particularly useful. They required physical infrastructure, while ChatGPT could spread through infrastructure that already existed. The more relevant point is that millions of people could encounter the technology directly, without waiting for an employer to introduce it.

Many employees tried AI at home before their organisations had bought licences, written policies or decided what responsible use should look like. They arrived at work already curious about what it could do.

2. Employees brought AI into the workplace themselves

Enterprise technology usually travels from the organisation to the employee. A system is selected, approved by IT, tested and then introduced through a formal rollout.

Generative AI often travelled in the other direction.

The 2024 Microsoft and LinkedIn Work Trends Index found that 75% of the knowledge workers surveyed were using generative AI at work. Among those users, 78% said they were bringing their own AI tools to work rather than relying only on tools provided by their employer.

These figures came from a survey of knowledge workers, not the entire workforce, and Microsoft has a commercial interest in workplace AI. Even with those qualifications, the findings describe a real shift. Employees were not simply waiting for a corporate programme. They were testing public tools on their own and deciding for themselves where those tools might be useful.

By the time some organisations began discussing AI adoption, adoption of a sort had already started.

3. Natural language makes it easy to begin

Many workplace technologies require people to learn how the system works before they can do anything useful with it. They may need to understand a programming language, a database or a proprietary interface.

Generative AI lets people begin with ordinary language. They can ask a question, request a summary or upload a document without knowing much about the technology underneath.

The barrier to first use is therefore unusually low. But the simplicity of the interface can be misleading. Producing an answer is easy. Deciding whether that answer is accurate, appropriate, safe or useful takes more judgement.

The learning has not disappeared. It has shifted from operating the tool to evaluating what the tool produces and deciding when it should be used.

Starting quickly is not the same as adopting well

The 2026 Stanford AI Index reports that organisational adoption reached 88% in 2025. That figure shows how widely AI has spread, but it does not tell us how deeply it has been integrated or how much value it has produced.

“Using AI” might mean that a few employees occasionally use publicly available generative AI tools , that a company has purchased licences or that an entire workflow has been redesigned around AI. These are very different levels of adoption.

An organisation can make software available in a matter of days. Deciding how it should affect work takes longer.

Which tasks should AI support? What information can safely be entered into it? When must its output be checked? Who is responsible when an AI-supported decision is wrong? Which parts of a role should remain firmly human? What happens to the expertise people have built over many years?

These questions require technical knowledge, good governance and careful work design. They also affect how people see their work and their own value within it.

The psychological side of AI adoption

AI does not create the same response in everyone. Some people welcome a tool that removes repetitive work. Others worry that the work being removed is precisely what made them valuable. Many are curious and anxious at the same time.

Imagine someone who has spent years becoming exceptionally good at analysing information, writing reports or producing first drafts. They are now told that AI can complete part of that work in seconds. Even when the technology is genuinely helpful, the change can raise an uncomfortable question: if the machine can do this, what am I now here to do?

That is not necessarily resistance to technology. It may involve professional identity, status, confidence and job security.

There is also the strain of repeated adjustment. AI implementation is unlikely to be a single change programme with a clear endpoint. Tools improve, organisational policies change and expectations rise. People may still be learning one way of working when another version appears.

Technical training can explain how to use a tool and what the rules are. It cannot, by itself, help someone decide what the change means for their role or speak honestly about the fear of becoming less relevant.

Nor should every concern be treated as a psychological blockage. Some objections are well founded. Employees may be noticing poor data, unclear accountability, additional workload or risks that leaders have underestimated. Listening to those concerns is part of responsible implementation.

Where coaching could help

My working hypothesis is that coaching can become an important part of the support surrounding AI adoption.

It is not a substitute for governance, technical expertise, communication or work redesign. A coach cannot repair an unreliable system or make a careless implementation responsible. Coaching should not be used to persuade people to accept decisions that deserve to be questioned.

Its contribution lies in giving people time and space to examine what the change means for them and their work.

A confidential coaching conversation allows people to say what they may not be willing to say in a training session or a meeting with their manager. Someone can examine concerns about job security, competence or professional identity without having to present a polished, publicly acceptable response.

Coaching can also help people move from knowing what a tool can do to deciding how they will use it. Writing a prompt is one small part of working well with AI. People also need to decide when to trust the output, when to challenge it, what judgement they should retain and how their own habits may need to change.

For leaders, the questions are different. They may need to make decisions before the technology or its consequences are fully understood. They have to balance pressure for rapid adoption with responsibility for risk, trust and the effect on their teams. Coaching can help them examine their own assumptions, including excessive enthusiasm, avoidance or impatience with people who are less convinced.

Teams may need structured conversations about how responsibilities are changing, where human oversight remains necessary and what they have learnt from their experiments. These conversations are difficult to replace with a generic video or a company-wide town hall.

Research on workplace coaching has found positive effects on learning, performance and affective outcomes. We do not yet have enough evidence to say that coaching solves AI-related anxiety or guarantees better adoption. The argument here is narrower: when people are dealing with uncertainty, changes in professional identity and decisions that require judgement, coaching is relevant.

From an occasional benefit to embedded support

By embedded coaching, I mean coaching that is built into the change process and available when people are making decisions, testing new ways of working or dealing with changes in their roles. It sits closer to the work than a separate development benefit someone might use months later.

Coaching is still often reserved for senior executives or introduced when a specific performance problem has appeared. That may be too limited for a technological change affecting people across functions and levels of an organisation.

Embedded coaching does not mean giving every employee permanent individual coaching. An organisation might offer individual coaching to leaders responsible for AI adoption, team coaching where roles and workflows are changing, or group coaching for employees dealing with similar questions. Managers could also be helped to have better conversations with their teams rather than receiving a set of communication slides and being expected to manage the rest alone.

Scalable coaching does not have to mean replacing human coaches with automated ones. Digital tools and AI-generated prompts may help with reflection, preparation or follow-through. Conversations involving fear, identity, conflict or job security require judgement, trust and genuine human attention.

The right combination will depend on the organisation, the work and the people affected. The important shift is to stop treating coaching only as an executive perk or as a response after something has gone wrong.

AI tools can spread through an organisation faster than people can decide how to use them well. Governance, training and work redesign remain essential, but people have to create and carry out all three while dealing with the effects of the change themselves.

My working hypothesis is that embedded coaching can help them do that work with more thought, honesty and care.