Last verified 23 September 2026. Microsoft 365 Copilot Business costs $21 a user a month on a yearly plan, for companies up to 300 users, and needs a Microsoft 365 business licence. Copilot sees whatever each user can already open. Details for experts at the end of this book.

The idea in one page

A rollout is a change to how a department works, not a licence order. Settle five things before the first seat, in this order.

Seats follow tasks, not people. Start from the pilot's tasks. A seat goes to a role that does those tasks often. Everyone else waits for the next wave, or gets the free chat that comes with their existing account. Unused seats are reclaimed after 30 days and reassigned; the usage report's last-activity date shows which ones.

Cost the time, not only the licence. A rollout's full cost is the licences plus IT's setup time, cleaning up permissions, training hours for every participant, and support hours each week. Divide it by accepted outcomes, the tasks done with the tool that met the same quality check as before. Logins and prompts don't count.

Access through a group. IT creates one group per wave in the directory, the central list of accounts, and assigns the licence to the group. Adding someone to the group gives them access; removing them takes it away. That's least privilege at rollout scale: nobody has access by accident, and the undo is one step.

The assistant sees what each person can see. Assistants built into office software, like Microsoft's, search the files, mail and chats each user already has permission to open. That's also their risk. A folder shared with "everyone in the company" in 2019 was harmless because nobody found it. The assistant finds it. So check and fix sharing before the switch is flipped.

Agree what's measured before you measure it. Most enterprise assistants report usage per person: prompts, active days, last use. In Germany, introducing a technical system that can monitor employees' behaviour or performance needs the works council's agreement, and at public institutes the staff council's. Agree in writing which usage figures are looked at, at what level (the team, not the person) and by whom. The DPO checks the contract with the provider, where data is processed, and whether the data protection impact assessment, the written risk check the GDPR requires for risky processing, needs an update.

Objections are information. The respected sceptic usually has a reason: quality, job security, or a rollout that failed before. Give objections a named route to someone who can act on them, and answer each one in writing. Don't make use compulsory to improve the numbers.

Worked example: two rollouts, four quadrants

Both versions follow the four quadrants from T01-L01 · From Idea to Working Result: The Four-Quadrant Method. The company one is told in full. Names, numbers and dates are illustrations.

Company: Microsoft 365 Copilot for sales and service

Structure it. Sven, the sales operations lead, takes it on after the pilot. "Give the AI assistant to the people whose work the pilot showed it helps, with their files as safe as before, agreed with the works council, and with a way to stop. A useful result: after six weeks, more quote replies go out on time at the same quality, every seat is used, and nobody's performance is judged by a usage report."

Understand it. He starts with the work, not the tool. From the pilot's results note and a 15-minute conversation with each team lead he maps the department's 40 people:

• 18 sales reps write quotes, customer emails and meeting summaries all day in Outlook, Word and Teams. The pilot's tasks.
• 12 service agents work almost entirely in the ticket system of the CRM, the customer database, which Copilot doesn't reach. The CRM's own AI feature, which the pilot used, covers them.
• 10 back-office staff do orders and invoices in the ERP. Occasional writing, no pilot task.

Then he asks IT for one test seat and signs in as a pilot user, with the DPO watching, and asks the assistant three questions any curious colleague might ask: "What bonuses did sales get last year?", "Show me salaries" and "What's in the restructuring plan?" The first answer summarises a commission spreadsheet from a SharePoint site created for a 2019 sales event and shared with "everyone except external users". The second finds nothing. The third finds a slide deck from a management offsite, shared by link in a Teams chat that included two sales reps.

He also reads the usage report IT can see, in the Microsoft 365 admin center under Reports, Usage, Microsoft Copilot. It lists, per person, prompts submitted, active days and the date of last use. Names are hidden by default; an administrator can show them with one setting in the admin center's organisation settings for reports.

The permission findings change his problem. He had been thinking about seats and training. The bigger risk was the department's old sharing. Back to Structure it: "Before any seat: fix what's overshared on the sites and in the chats this department can reach. Seats for the 18 reps and the 4 team leads, not 40. Usage is looked at per team, never per person, and the works council signs that before anything starts."

Choose tools.

• Microsoft 365 Copilot Business, because the company already runs Microsoft 365 and the reps work in Outlook, Word and Teams, where the assistant sits. The alternative is ChatGPT Business, which is stronger as a standalone assistant but would put the reps' work in a second place, outside the files they already use.
• Group-based licensing in Microsoft Entra ID, the company's directory, where IT already manages accounts: two groups, Copilot-Wave1 and Copilot-Wave2. IT assigns the Copilot licence to each group in the Microsoft 365 admin center, under Billing, Licenses, Assign licenses.
• SharePoint's oversharing reports, part of SharePoint Advanced Management, which comes with the first Copilot licence. IT finds them in the SharePoint admin center under Reports, Data access governance. They list sites shared with everyone in the company and files shared by open link. From a report, IT can send each site's owner, the person responsible for that site, a site access review: a request to check who has access and fix it.
• The company's existing sign-in with multi-factor authentication, so nothing new to manage.
• A Teams channel for questions, and two colleagues from the pilot as first helpers, called champions.

Build it: the plan, on two pages. The cost first, for the first year, in a small table:

At an internal rate of about €60 an hour, the 140 hours of time cost more than the licences. Sven writes that into the plan, so nobody is surprised when the champions' managers ask.

Then the stop rule, agreed with the head of sales before the start: pause the rollout if confidential material turns up for someone who shouldn't see it, if the quote-reply quality check from the pilot fails more often than before, or if the works council asks. Expand to wave 2 only if wave 1's accepted outcomes rise and no oversharing incident is open.

Build it: fix what the assistant can see. IT runs SharePoint's report of sites and files shared with everyone in the company, filtered to what the 40 people can reach. It lists 31 sites. For each, the site owner gets a site access review with one question: keep, restrict to a named group, or archive. Sven chases them for two weeks. The 2019 event site is archived. The commission spreadsheet moves to the finance team's site. The offsite deck's link is removed. Four site owners have left the company; their sites go to their managers. Then Sven repeats the three test questions with the test seat. None of them finds anything now.

Build it: agree it. Two meetings, one page each.

The DPO checks that Copilot runs under the company's existing Microsoft data processing agreement, the contract that says what Microsoft may do with the company's personal data, and that Microsoft doesn't train its models on the company's prompts or files. She notes one setting. Copilot can also offer Anthropic's models, and those are processed outside the EU Data Boundary, Microsoft's commitment to keep EU customers' data in the EU. IT leaves them switched off. She updates the data protection impact assessment and signs.

The works council gets the plan, the usage report and a draft agreement. The core lines:

Usage figures are looked at only per team of at least five people. User names stay hidden in all usage reports. Nobody's use or non-use of Copilot is used to assess their performance, and nobody is required to use it. Copilot's own features for summarising meetings are used only when everyone in the meeting has been told at the start. The agreement is reviewed after six months.

The council asks for one change: the respected senior rep, who has objected openly, should be in wave 1, if he wants to be. Sven asks him. He agrees, on the condition that he gets to write down what doesn't work.

Build it: wave 1. First IT proves the way back: it adds the test account to Copilot-Wave1, removes it again, and checks on the account's licence page that Copilot is gone. Then IT adds 8 people to the group: four reps including the sceptic, two team leads, the two pilot colleagues as champions. Each gets 90 minutes of training on their own tasks, as in T13-L02 · Show a Colleague AI on Their Own Task, in 30 Minutes, and the team's one-page working rule from T13-L03 · Write a One-Page AI Rule Your Team Can Use Without Asking You, updated with one new line: "Copilot finds what you can open. If it shows you something you shouldn't see, tell Sven the same day and don't forward it."

Understand it: what wave 1 shows (4 weeks). Sven measures what the pilot measured: time to first reply on quote requests and the price check. Median time to first reply drops from 6.5 to 3 working hours. No wrong price goes out. Two seats are almost unused, and the reason is simple: both reps are on the road most days, and they use their phones, where the setup hadn't been explained. One incident: a rep's summary of a Teams channel includes a salary figure from a file someone posted there. It's his own team's channel, and he could have read it anyway, but it shows the clean-up missed Teams files. IT extends the report to Teams, and three more files are moved.

The sceptic's list has five entries. Three are fixable (mobile setup, a template, German spelling settings). Two are about quality: Copilot's email drafts are too long and too friendly for long-standing customers. He's right, and his rule of thumb, "never send a draft to a customer you've known for more than a year without rewriting the first sentence", goes into the team's rule. Every entry gets a written answer in the channel.

Structure it: decide wave 2. Accepted outcomes are up, no oversharing incident is open, the works council has had no complaints. The two unused seats are reassigned. Wave 2 adds the other 14 people in two weeks. The service agents stay with the CRM's AI feature, and the back office gets the free Copilot Chat that comes with their existing licence, with the working rule. The seat list, the cost table and the objection log go to the head of sales, who signs the plan for the rest of the year.

What it taught him: the licences were a purchase order. The permission clean-up and the works agreement took three weeks, and they're what let the rollout go on after wave 1.

Lab: the institute's hosted assistant for eight groups

Structure it. Dr Paulsen, the institute's research coordinator, writes: "All eight groups use the institute's hosted assistant, Chat AI from GWDG, the data centre of the University of Göttingen and the Max Planck Society, instead of personal accounts, with one rule for what may go in, by January. No licence costs; the cost is time."

Understand it. Sign-in is the institute login through Academic Cloud, the portal GWDG runs for member institutions, so access is already there for everyone. The questions are elsewhere. Chat AI's own models store no chats, but the external OpenAI and Anthropic models in the same menu cost money and send data outside. Its document feature keeps uploaded files by design. Two groups hold clinical data. And the staff council points out that asking people to stop using personal accounts, and checking whether they do, touches how their work is monitored.

Choose tools and Build it. The same five steps without seats. The rule, adapted from T13-L03 · Write a One-Page AI Rule Your Team Can Use Without Asking You and T15-L04 · Set Up AI for a Whole Research Group: self-hosted models only for unpublished material; external models only for published or public material, and only after a PI, the group's lead researcher, signs for their group's budget; nothing from the clinical groups' restricted folders. The document feature gets its own line: published papers only. The DPO signs; the staff council agrees on one condition, that nobody's use is checked individually. Wave 1 is the pilot group plus two volunteers' groups, with a 60-minute session per group and one champion each. The objection route is a named person in the coordination office. After four weeks, one group asks for external models for manuscript editing; its PI signs, and IT confirms the budget line. Wave 2 takes the remaining five groups.

In both cases the purchase, or the lack of one, was the small decision. What made the rollout hold was fixing what the tool would see, agreeing what would be measured, and giving the sceptic a job.

What goes wrong

Seats for everyone. Symptom: 40 licences, 15 used after two months. Fix: seats follow the pilot's tasks and roles. Reclaim seats unused for 30 days and reassign them.

Oversharing found by users. Symptom: someone gets a summary of a file they shouldn't see. Fix: run the sharing report and ask three awkward test questions with a test seat before wave 1. Include chat files, not only sites. Name the person to tell, and make telling them normal.

The works council hears last. Symptom: a stop in week two, and trust lost. Fix: bring the plan, the usage report and a draft agreement before any seat is switched on. Agree what's measured, at what level and by whom.

Measuring people instead of work. Symptom: a ranking of who uses it most. Fix: measure the pilot's outcomes per team, keep names hidden in usage reports, and write down that non-use has no consequences.

Cost counted as licences only. Symptom: champions burn out and their managers ask why. Fix: put setup, clean-up, training and support hours into the plan, and give the champions time for it.

Everyone at once. Symptom: forty people with questions in week one and nobody to answer them. Fix: two waves, champions from the pilot, a channel and a stop rule written before the start.

No way back. Symptom: a problem appears and nobody knows how to switch it off for one team. Fix: licences through one group per wave; removing the group is the undo, and IT has tested it once.

Do it yourself (5 hours plus the rollout)

1. Structure it (20 min). Write who needs the tool for which tasks, what a useful result looks like after six weeks, and what would make you stop.
2. Understand it (90 min). Map the department's roles against the pilot's tasks. With a test seat and the DPO, ask three awkward questions and note what the assistant finds. Read the usage report. Rewrite your problem sentence.
3. Choose tools (20 min). Decide the seats, the groups, the support route and the objection route. Name one alternative tool and why you didn't pick it.
4. Build it: plan and cost (40 min). Write the two-page plan: seat list by role, the cost table with time, the stop rule and the wave plan.
5. Build it: clean-up (60 min, plus site owners' time). Have IT run the sharing reports under Data access governance and send each site owner a site access review. Repeat the test questions until they find nothing.
6. Build it: agreements (60 min, plus waiting). Take the plan to the DPO and the works or staff council. Agree in writing what's measured, at what level, and by whom.
7. Build it and Understand it: wave 1 (the rollout). Have IT add and remove a test account once to prove the undo. Add wave 1 to its group, train on their own tasks, run the channel, log every objection with its answer, measure the pilot's outcomes, and decide wave 2 by the rule.

Exit check

Written finding: a rollout plan and the results of its first wave, with:

• the seat list, each seat tied to a role and a task, and the cost table including people's time;
• the oversharing check: what the test questions found, what was fixed, and the repeat that found nothing;
• the signed page from the DPO and the agreement with the works or staff council;
• the access groups, and IT's note that removing a group removes access;
• after wave 1: accepted outcomes against the pilot's measure, unused seats and what happened to them, every objection with its written answer, and the wave 2 decision by the stop rule.

It passes when each of these exists and a colleague who wasn't involved can say, from the plan alone, who has access, why, and how to take it away.

Rule to remember

Fix what the assistant will see, and agree what you'll measure, before the first seat is switched on.

Details for experts

Checked on 23 September 2026. You don't need this to follow the book; it's for setting things up or answering a colleague.

Further reading and tools

Next books

• T13-L05 · Keep Everyone's AI Skills Current, and Show It When Asked: keeping training records for everyone the rollout reaches.
• T17-L04 · Run a Pilot with Real Users and Real Data: the pilot and results note a rollout starts from.
• T13-L03 · Write a One-Page AI Rule Your Team Can Use Without Asking You: the working rule each team gets.
• T04-L04 · Connect an Assistant to Real Documents Without Widening Anyone's Access: why an assistant's search follows permissions, in depth.
• T16-L05 · Keep Your Toolset Useful: retiring seats and tools that no longer earn their cost.

Tools

• Microsoft Copilot: Taught. The assistant in the company rollout.
• ChatGPT: Catalogued. The standalone alternative.
• Langdock: Catalogued. A managed workspace some organisations approve instead.