A common outcome of an AI project is a working system that almost nobody uses. The pilot went well, the rollout happened, and three months later usage sits with the four people who were enthusiastic from the start.
AI adoption fails for reasons that have very little to do with the technology, and the reasons are consistent enough to plan around.
It has to be faster than the current way
People are comparing the new tool against what they do now, including all the shortcuts and habits accumulated over years. A tool that is theoretically better and practically slower loses that comparison every time.
Watch someone do the task the old way before deciding the new way is an improvement. The existing process is frequently faster than the process diagram suggests, because the diagram does not include the shortcuts.
The reasons people quietly do not use it
- It sits in a separate system, so using it means leaving the tool they live in.
- They tried it, got a poor answer early, and concluded it does not work.
- They do not trust it and have no way to check the output quickly.
- They think it might replace them, which nobody will say out loud.
- Nobody showed them what good use looks like for their actual job.
The fourth is the one most often present and least often addressed. An unanswered question about job security suppresses adoption regardless of how good the tool is.
If you have not said clearly what this means for people's jobs, they have already decided, and they have assumed the worst.
Ethan Caldwell, CTO & Co-Founder, Engineered With AI
Put it where the work already happens
Every context switch costs adoption. A tool inside the system somebody already has open all day gets used; the same capability behind a separate login does not, however good it is.
This is worth real engineering effort, and it is regularly deprioritised because integration is unglamorous next to capability. The integration is frequently what determines whether the capability is used at all.
Volunteers first, mandates never
Start with people who want to use it, support them properly, and let the results become visible. Colleagues asking how somebody got their afternoon back is a far stronger driver than an instruction from management.
Mandated usage produces compliance behaviour: people use it in the way that satisfies the metric and go back to their own method for the actual work. That looks like adoption in a report and changes nothing.
Make the first experience a good one
First impressions are disproportionate. Rather than a general demonstration, sit with someone and use it on a real task from their queue. One genuinely useful result does more than an hour of training, and one bad early result can end usage permanently.
Give them a way to verify
People will not rely on output they cannot check. Showing sources, showing what the answer was based on, or making it easy to compare against the record turns a leap of faith into a quick confirmation.
This matters more for experienced staff, who have the most to lose from being wrong and the most professional judgement to apply. They are also the people whose adoption most influences everyone else.
Be honest about what changes
If roles will change, say so, specifically, and say what happens to the people in them. Vague reassurance is heard as evasion and it damages trust in everything else you say about the project.
Where the honest answer is that the work shifts rather than disappears, describing what it shifts to is what allows people to engage with it. Adoption follows from that conversation far more reliably than from any feature. The automation hub covers how we phase this in practice.
Measure use, not licences
Seats assigned tells you what you bought. What matters is how many people used it in the last week, on what kind of task, and whether that number is rising.
Break it down by team as well. AI adoption is usually uneven rather than uniformly poor, and finding the team where it is working tells you more about what to change elsewhere than any survey will.
Rollout stalled?
We will find out why the tool is not being used and fix the process around it.