A company bought Microsoft 365 Copilot licenses for five hundred people. It organized a two-day training. For the first week after training, the usage stats looked impressive — people were trying things out, testing, asking questions. After a month, fifteen people were using Copilot regularly. Fifteen out of five hundred. Three percent.
I hear this story — with different numbers but an identical pattern — from large organizations. Corporations, state-owned companies, firms on major stock indices. The pattern is always the same: training, spike of enthusiasm, silence, drop to a few percent of regular usage.
What Microsoft Doesn't Tell You
Microsoft reports millions of active Copilot users. Sounds impressive. But what does "active" mean? In official definitions: a user who used an AI feature at least once in a given period. Once. They opened the panel, saw a suggestion, closed it. They're in the statistics.
Real, regular usage — the kind that changes how people work — is an entirely different metric. And that metric, from my observations and conversations with clients, hovers around 3-4% of users in large organizations. The rest went back to what they were doing before.
This isn't a Copilot problem. It's a problem with how companies approach the adoption of AI tools.
Training Is an Event. Adoption Is a Process.
Most organizations treat Copilot deployment (or any other GenAI tool) as a technology project with an end date. Buy licenses — check. Run training — check. Send an email with instructions — check. Project closed.
Except people don't change how they work after a two-day training. Not because they're resistant or incompetent. Because nobody showed them how Copilot solves their specific problem in their specific process. The training demonstrated general features. A demo on sample data. "Look what AI can do." People nodded, returned to their desks, and opened Excel the same way as yesterday.
Why People Stop Using It
From conversations with users across several organizations, I see recurring reasons:
"I don't know what it's good for in my job." The training was generic. It showcased tool capabilities, not solutions to specific problems in specific roles. The accountant doesn't know how Copilot helps with account reconciliation. The project manager doesn't know how to use it for status reporting. Everyone saw the demo, nobody got a recipe for their workday.
"I tried it and the result was bad." The first experience with GenAI is crucial. If someone asked Copilot to summarize a document and got a banal, useless text — they won't come back. Nobody taught them how to formulate prompts, how to provide context, how to iterate. One bad experience closes the door for months.
"My process hasn't changed." People work within processes. If the process doesn't account for AI — there's no room for Copilot. If a report still needs to be written in a specific template, in a specific way, and nobody modified that template for AI — the tool is an add-on, not part of the work.
"Nobody asks if I'm using it." No follow-up after training. No check-ins, no asking about barriers, no support. People read this as a signal: "This isn't a priority."
What Works Instead of a One-Time Training
From my practice — AI adoption in an organization requires an approach closer to change management than product training.
Workshops embedded in real tasks. Not "what can Copilot do," but "how Copilot helps you create this specific report you produce every week." Fifteen people, their real data, their real processes. After such a workshop, twelve out of fifteen change how they work — because they saw value in their own context.
Process change, not just tool change. If you want people to use AI for creating reports — change the report template. Add a step in the process that assumes AI usage. Make it the norm, not an option.
Regular check-ins. Once a month: what's working, what isn't, what do you need. Not a survey — a conversation. People have questions, frustrations, ideas. You need to listen and respond.
Measuring adoption, not licenses. How many people changed how they work? How many processes incorporate AI? How much time do people who use the tool regularly save? These are adoption metrics — not the number of activated licenses.
Caring about user needs. Sounds obvious, but it's the rarest practice. Asking people: what do you need? What's getting in your way? How can I help? Adoption doesn't happen despite people — it happens because of them.
Three Percent Is Not a Tool Failure
Three percent regular usage is not a problem with Copilot, Claude, Gemini, or any other GenAI tool. It's feedback about how the organization approached the change. The tool works. People can learn. But between "we have a license" and "people work differently" is an enormous space that must be deliberately cultivated.
Companies that take this seriously work with AI tools in a way that is embedded in organizational realities — not as a one-time project, but as a change in how people work that requires time, attention, and consistency.
If you bought licenses and see that adoption is low — that's not the end of the road. It's the beginning of the real work. Let's talk about how to plan it.