The most common AI rollout we see is the simplest: buy licences for everyone, send an email, wait. Six months later usage is concentrated in a few enthusiasts, most seats are idle, and the finance director is asking what the money bought. The alternative is not complicated. It is a rollout by role, in phases, with enablement on each function’s own tasks and measurement that shows what changed.
Start with the trial you probably already have
Most businesses have a Copilot trial running in pockets, or a few people on Claude or ChatGPT. That is the pilot group. Find out what they actually use it for, where it helps, where it does not, and what they have worked out about prompting their own work. They become the champions, and their patterns become the first enablement material.
Sequence by role, not by seniority
The order should follow where the recurring, tool-shaped work is and where the data is least sensitive, not the org chart. A typical sequence:
| Phase | Who | Why first |
|---|---|---|
| 1 | Marketing, sales support, admin | High volume of drafting, summarising and reformatting; data mostly internal or public |
| 2 | Operations, procurement, customer service | Reports, minutes, correspondence, knowledge lookups; benefits from the configured assistants |
| 3 | Engineering and R&D | Developer tooling is its own rollout, with its own review gates |
| 4 | Finance, legal, HR | Highest-sensitivity data; last, with the classification rules proven |
Each phase gets its own enablement session, on that function’s tasks, with the custom assistants, GPTs or skills configured for them beforehand.
Gate access with the policy session
Nobody gets a licence before the short acceptable-use session: what may be shared with which tool, how to check output, how to report a problem. It takes an hour, it is the training the EU AI Act asks for, and it is what makes the data classification real rather than a document.
Enable on real tasks
Generic training does not change behaviour because nobody’s job is generic. Sample ten to fifteen recurring tasks in the function, the monthly report, the supplier email, the meeting notes to actions, the spreadsheet reformatted into slides, and run the session on those, with people doing their own work in the tool. They leave with working patterns, not a slide deck.
Measure what changed
Before each phase, time a sample of the function’s recurring tasks. After enablement, time them again with the tools. That number, with usage by role from the platform’s admin reporting and the champions’ feedback, tells you whether the rollout is working, where the next session should go, and what the licences are worth. It is also the number that silences the finance director’s question.
Keep the list current
An approval route for new tools, reviewed by the AI working group or whoever owns the policy, so that the next useful tool is added to the sanctioned list rather than adopted around it. Governed access only stays governed if it is easier to ask than to bypass.
Timeline
Selection two weeks, if the tool is not yet chosen. Rollout design two to three weeks. Pilot phase live within six weeks of the start. One function every two to three weeks after that, so a five-function business is enabled within a quarter.
The rollout is described under AI tools and platform rollout, and the enablement under team enablement.