When an AI Tool Goes Unused, Ask About the Task

Picture a small business that buys an AI subscription, gives the team a demonstration and checks usage a month later. Some seats remain unused. The manager considers another training session. Before booking it, there is a more revealing conversation to have: what happened on the last job where someone considered using the tool?

An unused account tells you very little about that moment. Someone may have lacked suitable material, needed an approval, struggled to check the output or chosen an existing method that worked better. Each explanation calls for a different management response. Start by finding the task behind the usage figure.

What the surveys can tell us

The OECD’s Generative AI and the SME Workforce, published on 5 November 2025, draws on a 2024 survey of 5,232 small and medium-sized enterprises across seven countries, including the UK. Respondents were people familiar with their company’s technology who had heard of generative AI. Among non-users, 57% reported that generative AI did not suit their work, while about half reported a lack of employee skills. Concerns about information supplied to models also featured among the reported barriers.

These are companies’ stated reasons, and respondents could agree with several explanations. They are not independent tests proving that a particular task is unsuitable. The results describe the surveyed businesses at that time; they do not establish today’s UK adoption rate or explain any individual colleague’s behaviour.

Permission and preparation are different problems

A PNAS study by Anders Humlum and Emilie Vestergaard, published online on 30 December 2024, studied about 18,000 workers in eleven Danish occupations exposed to generative AI, linking survey responses with administrative records. Workers recognised potential productivity benefits, but the study also identified employer restrictions and workers’ perceived need for training as obstacles to using ChatGPT.

The two studies examine different units: businesses and individual workers. Their findings should not be combined into one adoption figure. My practical inference is narrower: investigate permission, preparation and suitability separately before deciding what help to offer. Neither paper diagnoses the problem in your own team.

Start with the last relevant job

Invite someone to describe a recent, specific task. Avoid beginning with whether they are enthusiastic about AI. Ask what they needed to produce, what they considered doing with the tool, and where the proposed approach stopped making sense.

• Opportunity: Did a relevant task actually arise during the period being reviewed?
• Access and permission: Could the person access the approved tool, and were the proposed materials permitted for that use?
• Execution: Could they prepare the input, assess the output and turn it into finished work?
• Alternative: What method did they use instead, and why did it seem preferable?

Record the explanation in the person’s own terms, alongside anything still uncertain. “I could not find an approved example” is more useful than “low engagement”. Several obstacles may apply to the same job. Leave an unresolved question open until someone has checked it.

One studio, three reasons for non-use

Consider a hypothetical design studio trialling AI-assisted proposal drafting. Three colleagues do not use it on recent jobs. One is unsure whether a confidential client brief may be supplied. Another knows how to request a draft but is unsure how to check its milestones against the proposed work. A third finds that the studio’s existing proposal template is quicker for a routine repeat commission.

The first case needs a decision about permitted material and an approved way to work. The second needs practice in reviewing a complete proposal. The third needs a fair comparison of methods for that kind of commission. Giving everyone the same introductory demonstration would leave those questions unanswered.

The studio could prepare an authorised practice brief, work through a proposal with the second colleague and retain the existing template for the repeat job while assessing other uses. These are possible next steps, not promised productivity gains. The useful result is a clearer account of what needs testing.

Match the response to the obstacle

For a skills question, choose one representative job and practise it from input to checked output. Include the judgement involved in accepting, revising or rejecting a draft. Give the learner time to ask questions and repeat the task. Afterwards, check which difficulty remains instead of treating attendance as proof of competence.

For an approval question, identify who can resolve it. Agree the tool, the material, the intended use and the point requiring further review. The MaryChuks guide to decision rights explains how to make that authority explicit. Encouragement from a manager should come with a usable answer about the work being proposed.

For a task-fit question, compare the AI-assisted approach with the existing method on agreed examples. Include preparation, checking and completion in the comparison. Decide what counts as an acceptable result before starting. A short trial can inform the next decision; it does not establish that one method will always be better.

Review opportunities as well as activity

At the next review, keep a small record of relevant jobs: whether the tool was considered, whether it was tried, what was delivered and why another method was chosen. This gives you a way to distinguish a period with no suitable work from a period in which suitable work repeatedly encountered the same obstacle.

Define the tasks included in the trial before counting success. Show excluded tasks and the reasons for excluding them. Otherwise a team could make its results look stronger simply by removing difficult cases afterwards. Keep the record proportionate: enough detail to support a decision, without turning every working day into an account of chatbot activity.

Make the next decision concrete

End the review with a named action: resolve a permission question, provide practice, test a particular task or retain the existing approach. Give that action an owner and a review date. If no worthwhile use has been identified, record that conclusion and the evidence that would justify reconsidering it.

For an independent creator, the same questions can guide a subscription review. For a team leader, they turn a broad concern about adoption into a manageable piece of work. The next conversation can then begin with what changed for the task, and whether the person can complete it well.

Sources

• OECD, Generative AI and the SME Workforce: New Survey Evidence, 5 November 2025. See the survey methodology and reported barriers to use.
• Humlum and Vestergaard, The unequal adoption of ChatGPT exacerbates existing inequalities among workers, PNAS, published online 30 December 2024.