Enterprise software has a long history of tools that passed every technical gate and then quietly died in the field: licenses bought, training delivered once, usage flat within a quarter. AI tools are unusually exposed to this failure mode, for two reasons. First, using them well is a skill — an employee who tries twice, gets a mediocre result, and concludes "it doesn't work for my job" is behaving rationally. Second, the tools arrive wrapped in anxiety: about job security, about being blamed for an AI mistake, about whether using help "counts" as doing the work. A rollout plan that addresses neither is just a deployment, not an adoption.
Start with a workflow, not a tool
"We now have an AI assistant, use it for whatever" is the weakest possible launch. Adoption sticks when the tool is introduced inside a specific workflow with an obvious payoff: draft the response to this ticket category, summarize this meeting into actions and owners, produce the first pass of this weekly report. Pick one or two workflows per team where the before-and-after is visible in minutes, and launch there. Employees who experience a concrete win generalize on their own; employees handed a blank text box mostly do not. This also keeps the change aligned with whatever your acceptable-use policy permits, because the sanctioned paths are the well-lit ones.
Train for judgment, not just clicks
One demo at launch is not training. Effective programs are short, role-specific, and repeated: what this tool is good at in your job, how to write a request that gets a useful answer, and — most important — how to review the output. Every user should leave training able to answer two questions: "what do I check before I trust this?" and "what do I never send to it?" That review habit is not a compliance afterthought; it is the difference between AI-assisted work and AI-generated mistakes with an employee's name on them. Make it explicit that the human remains accountable for the work product, and equally explicit that using the tool is encouraged, not a mark against them.
Champions beat mandates
The strongest adoption signal an employee can receive is a respected peer saying "this saves me an hour a day — here's how." Recruit champions from each team: not the most senior person, but the practical one others already ask for help. Give them early access, a direct line to the build team, and time to share what works — a running channel of good examples ("prompt of the week," before-and-after samples) does more than any memo. Champions also route real feedback back to you: which workflows disappoint, what people are afraid of, what they wish it did. Leadership matters too, but as visible users rather than mandate-issuers; a mandate produces logins, not adoption.
Measure honestly, and expect the dip
Decide the metrics before launch, and keep them honest. Useful ones: weekly returning users by team, share of the target workflow going through the tool, time-to-complete for that workflow measured against a pre-launch baseline, and a periodic one-question pulse ("did this help you this week?"). Ignore vanity numbers like total prompts sent. When a team's usage is flat, treat it as information, not resistance: usually the workflow fit is wrong, the tool is awkward to reach, or the outputs disappoint — all fixable, but only if you ask. Expect an early dip after the novelty fades; the durable signal is where usage settles in month two and three. And close the loop publicly: when feedback changes the tool, say so. Nothing builds trust in a rollout faster than evidence that complaints are read.
Finally, address the job-anxiety question directly rather than letting it fester. Be clear about what the tools are for in your organization and what accountability looks like. Employees who believe the tool exists to help them do better work will explore it; employees who suspect it exists to replace them will quietly ensure it fails.
Where to go next
From shadow AI to sanctioned AI is the companion piece — many of your employees are already using LLMs unofficially, and that energy is convertible. Measuring ROI extends the metrics discussion from adoption to business value.