The paradox is common: the organization has chosen a tool, announced its launch and trained part of the teams. A few months later, usage remains sporadic, scattered or invisible. This gap is not an irrational resistance. It often signals that the project addressed technology before work.
An available tool is not yet a practice. Adoption advances when the organisation connects AI to specific tasks, clarifies the rules, allows time to experiment, organizes mutual support and measures a work result. The right indicator is not the number of accounts opened, but the ability to reproduce a useful and mastered practice.
Do not confuse access to the tool with its adoption.
Purchasing a license is visible, budgetable, and quick. Adoption is more discreet. It requires that a person modify a work sequence they already master, accept a learning period, and know within what limits the new practice is permitted.
The shortcut is tempting: if the tool is easy to use, usage should appear spontaneously. However, the simplicity of the interface says nothing about the real cost of integration into a profession.
Before using AI on a meeting report, you still need to decide:
- what information may be entrusted to it;
- who checks names, decisions and commitments;
- where to store the resulting document;
- how to report AI assistance when necessary;
- what happens when the answer is incomplete.
As long as these decisions remain individual, use remains fragile. One person experiments. Another abstains. A third invents their own rule.
Five barriers lie behind the word 'resistance'.
The use case is too vague
“Saving time with AI” is not a task. “Preparing an initial summary of three meeting reports using our template” is. The more general the project remains, the more each person must bear the cost of defining it alone.
Rules are absent or unreadable
What can be uploaded to the tool? Which models are approved? Which outputs must be checked? When in doubt, some people take risks while others do nothing.
The benefit arrives too late
A new practice initially requires more attention. People need to learn the interface, prepare the context, review and correct. If the first benefit is not quickly perceptible, the old habit returns.
Skills are reduced to prompting
The prompt is only part of the work. Identifying a suitable task, providing the right context, evaluating the response and inserting it into a process requires business, informational and critical skills.
Usage remains socially risky
Some fear being judged as less competent if they use AI. Others worry that their new efficiency will be used to increase workload or reduce headcount. The organization cannot ask for experimentation without explaining the human project that accompanies it.
The OECD notes that the skills gap remains a major barrier to adoption. Its 2025 report also emphasizes the need for practical understanding to identify where and how to use AI, beyond specialized technical skills.
Start from a real task, with the people who perform it.
A good use case is not necessarily spectacular. It is concrete enough to test and frequent enough for the learning to be reused.
This method avoids looking for a problem to justify an already chosen tool. It also helps identify the uses for which standard automation, better documentation, or removing a step would be more appropriate.
Set up a short adoption loop.
Adoption is not a launch event. It is a loop:
| Step | Question |
|---|---|
| Trial | Is the task feasible under our conditions? |
| Back | What helps, hinders or causes concern? |
| Adjustment | Which rule, data point or instruction needs to change? |
| Learning | Can another person reproduce the practice? |
| Review | Does the benefit still hold after several weeks? |
Peer exchanges play a central role here. A commented business example is often worth more than a long library of prompts. It shows the acceptable level of results, possible errors, and the place of human judgment.
A trial must also have an end. Without a decision point, pilots accumulate and trust erodes. Every experiment should lead to one of four decisions: stop, correct, expand or integrate.
Measure a transformation in work, not the platform's activity.
The number of activated accounts, queries, or participants in a training course describes activity. It does not demonstrate useful adoption.
Stronger indicators look at:
- the frequency of a clearly defined practice;
- the total time, including review;
- quality before and after;
- the number of errors or omissions detected;
- the share of the output that needs to be reworked;
- the satisfaction of the person producing it and that of the recipient;
- the ability to pass the practice on to someone else.
The project becomes credible when one can say: on this task, with this data and under this rule, the team achieves this result and knows how to control its limitations.
Available AI creates a possibility. Adoption begins when that possibility connects with real work.
Reference sources
- OECD, The Adoption of Artificial Intelligence in Firms, 2025
- OECD, AI and skills, 2026
- Microsoft Research, Shifting Work Patterns with Generative AI, 2025
FAQ
Why are employees not using the tool made available to them?
The causes may be multiple: lack of a clear use case, lack of time, unclear rules, concern about data, low confidence in the results, or integration cost too high within the workflow.
Is prompt training enough?
It can help teams get started, but it does not resolve task selection, data governance, output validation or changing roles.
Which first use case should we choose?
A frequent task that is sufficiently bounded, reversible, measurable and low-risk. It should produce a visible benefit without depending on a complete process redesign.
How to measure meaningful adoption?
Measure the frequency of a defined usage, the time or quality gained, the rate of human takeover, detected errors, and another person’s ability to reproduce the practice.
