A technically successful AI tool can fail for a very simple reason: nobody uses it. We see it regularly. The assistant is live, it works, the demo convinced management, and yet three months later usage has stalled. Teams kept their habits, or use it half-heartedly, without real conviction.

It is almost never a technology problem. It is a buy-in problem. And buy-in is not decreed in a launch email: it is built from the start of the project. This article describes what really holds teams back, and the practices that work to bring them on board.

Understand the obstacles before trying to remove them

Resistance to AI is rarely irrational. It rests on legitimate concerns, which are better heard than worked around.

  • Fear for one’s job: “if the tool does my work, what am I for?”. This is the most common fear, and often the least openly expressed.
  • Fear of being judged: not knowing how to use it, asking a naive question, or having one’s output compared with the machine’s.
  • A sense of extra work: a new tool to learn, an extra step, answers to check. If the gain is not visible quickly, the tool becomes a burden.
  • Distrust of reliability: one or two wrong answers at the wrong moment are enough to discredit a tool for a long time.
  • Feeling presented with a fait accompli: a tool chosen without the people who will use it starts with a handicap.

Involve users from scoping

The most effective lever comes into play before the tool even exists. The people who will do the work every day know where time is lost, which cases cause trouble, and what would really help. Involving them in choosing the use case, in design workshops and in the first tests changes two things: the tool is better designed, and it is no longer “management’s tool”, but a tool they helped build.

In practice, this can be as simple as a few interviews at the start, a small group of volunteer testers, and a regular check-in to share progress and gather feedback.

People don’t adopt a tool that is imposed on them. They adopt a tool that solves a problem they identified themselves.

Be transparent about what AI will and won’t do

Vagueness feeds fear. Stating clearly what the tool will handle, what stays in the teams’ hands, and how the freed-up time will be used defuses a good part of the concerns. If the goal is to absorb more volume without hiring, to reduce repetitive tasks or to spend more time with customers, say so. If roles must evolve, announce it honestly, with a support plan, rather than letting rumours do it.

You also need to be honest about limits: the tool can be wrong, its answers must be checked in certain cases, and that is normal. A team that has been warned reacts much better to an error than a team that was promised an infallible tool.

Train on real tasks, not on the tool

Generic training on “generative AI” has its uses, but it rarely changes practices. What works is training each team on its own tasks: drafting a customer reply with the assistant, finding a procedure, preparing meeting notes. Short sessions, in small groups, with the department’s real documents and cases.

You also need to plan for afterwards: easy-to-reach support, a few practical guides, and a place to ask questions without fear of being judged.

Rely on internal champions

In every team, a few people are naturally curious and comfortable with new tools. Identifying them, training them a little earlier and a little more deeply, and making them champions is very effective. A colleague showing how they save time on a task everyone knows is far more convincing than an official presentation. Champions also bring field difficulties back up, which lets you adjust the tool quickly.

The decisive role of managers

Front-line managers set the tone. If they use the tool, talk about it, and leave time to learn it, the team follows. If they see it as a gadget, or keep asking for things “the old way”, adoption stops. So bring them on board first, give them the means to support their teams, and avoid a common trap: using the tool’s usage data to monitor or assess individuals. Nothing kills trust faster.

Show results quickly, and measure usage

The first weeks are decisive. A narrow scope that delivers a visible gain quickly is better than a wide rollout that searches for its purpose for months. Sharing these first results, with concrete examples from daily work, builds momentum.

In parallel, track real usage: how many people use the tool, how often, for which tasks, and how satisfied they are. These indicators serve to improve the tool and the support, not to rank teams. A drop in usage is a signal to understand: a missing feature, an unreliable answer, a step that is too heavy.

A concrete example

A team of administrative officers was meant to adopt an assistant to quickly find the right internal procedures. The first launch, announced by email with a link and a short video, had little effect: a few tries, then a return to old habits. Informal feedback was clear: “it doesn’t know our special cases”, “I don’t have time to learn a new tool”.

The relaunch took a different path. Three volunteer officers listed the twenty questions they were asked most often, and the knowledge base was completed accordingly. Each team had a one-hour session on its own cases. The three volunteers became champions for their colleagues, and a monthly check-in surfaced unsatisfactory answers. The department head started using the tool in meetings too. A few weeks later, the assistant had become the first reflex for most routine questions.

Our recommendation

Budget change management as part of the project, not as an option added at the end. Involve users from scoping, be transparent about what AI will and won’t do, train on real tasks, rely on champions and managers, and measure usage to adjust. This part, far more than the choice of model, is often what makes the difference between a project that creates value and a tool that sits idle.