The pattern is remarkably consistent. Licences are purchased, there is a launch announcement and maybe a lunch-and-learn, the first two weeks show real curiosity, and then usage quietly fades. By month two, Copilot is a line item on the Microsoft bill and a slightly awkward topic in leadership meetings.
Nobody calls this a failure, because nothing visibly broke. That is exactly what makes it expensive: a stalled rollout keeps costing money while teaching the organization that "AI does not work here."
Why it happens
The tool was never connected to real tasks
"You now have Copilot" is not an instruction anyone can act on. The people who keep using an assistant past the novelty phase are the ones who found it inside a task they already do weekly: drafting the same category of email, summarizing the same recurring meeting, building the same report skeleton. If nobody mapped the tool to each role's actual repeatable tasks, usage decays to the few natural tinkerers.
The data underneath was not ready
Copilot inherits your information landscape. If files are scattered, duplicated, and inconsistently named, and permissions were never reviewed, answers come back generic at best and wrong at worst. People try it three times, get one bad answer, and quietly conclude it cannot be trusted. In our experience this is the single most common technical cause, and it is fixable, but it is data work, not AI work.
Nobody owns adoption
IT keeps the licences running. Department managers assume IT has it covered. The vendor's job ended at provisioning. When no single person is accountable for whether the tool produces value in real roles, fading usage is nobody's problem, so it fades further. One of the engagement examples on our Results page was exactly this situation: the fix was less about technology and more about giving adoption an owner and a role-by-role plan.
How to restart a stalled rollout
- Pick two or three roles, not the whole company. Choose roles with repetitive, text-heavy work.
- Sit with each role and list their weekly recurring tasks. Find the three where an assistant genuinely helps. Build the prompt or workflow together, in their real files.
- Fix the sharpest data problem first. Usually that means one team's document library or one shared drive, cleaned and permissioned properly, not a company-wide overhaul.
- Name an adoption owner with time to do the job, and give them a simple measure: weekly active use per role, plus one outcome that matters, like time to produce a specific report.
- Review in four weeks against a baseline you wrote down before restarting. Expand to more roles only when the first group would object to losing the tool.
The honest alternative
Sometimes the right call is to park it. If the underlying workflow is broken, or the data foundation needs months of work, restarting adoption theatre helps nobody. Fix the foundation first, then bring the assistant back when it has something solid to stand on. That order of operations is the whole premise of how we work: How It Works lays it out stage by stage.
If your rollout sounds like this article, the two-minute AI Readiness Self-Check will tell you which layer is stalling it, and a direct conversation will get you a restart plan sized to your team.