Why Copilot rollouts stall (and what does work)

In briefWhat this article is about

The conversation we have most often is not about whether Copilot works. It is about why it delivers something for the organisation down the road and nothing here. By that point the organisation has paid, rolled out and explained, and is left with an invoice that runs for a hundred per cent of its people while the usage sits with a handful. This article describes the five places that goes wrong, in the order in which they occur.

The thread running through all five is that none of them is a technical problem. They are about who gets the licence, about what sits in the shared folders, about what is being measured, about what happens after the first session, and about what the tool is not meant for. At the end there is a ninety-day timeline and a list of numbers that make the conversation with your board factual rather than anecdotal.

01What "stalling" looks like in numbers

The picture organisations have of their own rollout is usually gloomier than they dare say out loud, and the research backs them up. In The State of AI in the Enterprise, published in January 2026, Deloitte reports that only a quarter of the companies surveyed have moved at least forty per cent of their own AI pilots into production. In the same study, 37 per cent say they use AI largely on the surface, without changing the way work is done, against 30 per cent that genuinely adapt core processes around it.

For the Netherlands the picture is sharper. In State of Integration & AI 2026, carried out by Sapio Research among 611 IT and business decision makers across six European countries, 21 per cent of Dutch AI projects produce a measurable result. That was the lowest score among the countries surveyed. At the same time 67 per cent of Dutch companies say they use AI, up from 34 per cent in 2023. So a great deal is being introduced and little is being harvested.

Two figures from that same report explain much of the gap, and they are not the figures that get the most attention: 38 per cent of organisations run into internal resistance, and 34 per cent of those using AI name scaling as the sticking point. The sticking point sits with people and with the step from pilot to whole organisation. The choice of tool does not appear in those figures at all.

Do not read these figures as a forecast for your own organisation. Read them as evidence that a rollout which gets stuck is neither exceptional nor a sign of incompetence. It is the ordinary outcome of an ordinary approach. What follows are the five places that approach loses.

02Stall 1: the licence goes to whoever asks for it

The first decision is almost always the same, and almost always wrong. A budget appears for a number of seats, the question goes out broadly, and the licences land with the people who shouted loudest or sit highest in the organisation. That is a fine way to reward enthusiasm and a poor way to find out where AI saves work.

Microsoft supplies a report for exactly this purpose that is rarely opened. The readiness report in the Microsoft 365 admin centre shows two things at once: who is technically eligible for a licence, and how actively those people use the Office applications Copilot works inside. The second half is the interesting one. Someone who opens Word twice a month and otherwise lives in a line-of-business application will notice very little from a Copilot licence, however much they want one.

Microsoft's own rollout guidance explicitly recommends a phased approach: start with a small group, gather what happens there, and expand on what you learn. That order is not about saving money. It is about finding the people who later become your internal points of contact. Without that first group, a broad rollout leaves you with nobody who can explain from experience why something does or does not work.

If you are considering starting without licences at all, that is wiser more often than it sounds. Copilot Chat is included with almost every business Microsoft 365 subscription and costs nothing extra. Four weeks of watching who picks it up unprompted produces a better distribution list than any inventory round. What Copilot Chat does and does not do is covered in our Copilot Chat training; what the paid add-on costs and which items sit alongside it is set out in what Microsoft 365 Copilot really costs.

03Stall 2: there is no baseline, so there is no conversation

Ask how things are going six months after the rollout and you will get stories. Someone produced a quote in twenty minutes instead of two hours. Someone else found the meeting summary unusable. Both are true, neither is an answer, and the board that has to sign the renewal has nothing to go on.

The missing piece is almost never a dashboard. It is a number from before the start. Microsoft offers three measurement sources and all three are available as soon as you have licences:

• The readiness report in the admin centre: who is eligible, and how active those people are in the Office applications. Available within 72 hours, with up to 72 hours of latency on the usage data.
• The usage report: adoption, retention and engagement per application, plus which agents are being used. Data for a given day is generally visible within 48 hours.
• The Copilot Dashboard in Viva Insights: adoption and impact pages per group, with the option to delegate access to whoever does the analysis. Organisations with at least one assigned licence can now also see the Copilot Chat insights here.

Of those three, the word retention matters most. The number of people who have ever opened Copilot always rises after a rollout; it only tells you about curiosity. The number who used it this week and last week is the figure to follow. It is also the first figure to drop once the support stops.

A baseline takes an afternoon and determines whether, three months from now, you are having a conversation about results or about impressions. Do it before the first licence is assigned, and record what you see, not only in the dashboard but also in a short survey among the target group. Two questions will do: which part of your work costs you the most time at the moment, and where would you want help. You ask those same two questions again ninety days later.

04Stall 3: Copilot shows what was already shared too widely

This is the stall that genuinely halts rollouts, and it rarely appears in the plan. Copilot gives a user access to information that user was already permitted to see: no permissions are added and none are bypassed. What differs from before is the way of asking. To find a salary overview through search, you had to know it existed and roughly what it was called. To find it through Copilot, you only have to ask what the board spends on staff this year.

What surfaces then is not a fault in Copilot but ten years of accumulated practice: sites once shared with the entire organisation, folders where someone broke permission inheritance to give one colleague access, sites without an owner that nobody looks at any more. Microsoft now has an extensive toolkit for this:

• Assessments up front. The Content Management Assessment in SharePoint Advanced Management flags sites with oversized audiences, broken inheritance, inappropriate sharing, and sites that are inactive or ownerless. In Microsoft Purview, risk assessments run automatically each week across the hundred most used sites.
• Interim shielding. Restricted Content Discovery keeps designated sites out of Copilot's answers while you clean up. The same applies to policies that exclude sensitive material from grounding an answer.
• Structural measures. Sensitivity labels on sites and files, automatic labelling of what already exists, and policies that block external sharing of anything marked confidential.

The practical lesson is about sequence. A rollout that starts with licences and then hits permissions comes to a halt exactly when the organisation's attention is at its peak. A rollout that starts with an assessment of the shared environments loses two weeks and keeps its pace afterwards. Pull one more question forward here too: who takes the AI administrator role. That is a distinct role in Microsoft Entra, with access to the settings, the health and usage reporting, and the approval of agents.

05Stall 4: one session, and then silence

The fourth stall is the best known and the most often repeated. There is an hour-long kick-off, a series of possibilities is demonstrated, everyone is impressed, and then everyone goes back to their own work. That is where the gap sits: the session was about Copilot, and on the desk lies a grant application, a set of annual accounts or a rostering problem. The translation between the two is precisely what did not happen.

There is also a psychological mechanism underneath that most plans underestimate. One answer that is wrong weighs heavier than ten that are right. Someone who finds an invented figure in a summary in week one concludes the tool is unreliable and carries on the old way. That judgement is not unreasonable, it is simply made too early. Correcting it takes someone who can explain what went wrong and how the question could have been put differently.

What demonstrably does work is neither new nor specific to AI, and comes from the change management trade: a network of colleagues who are already doing it. Microsoft's own adoption material notes that learning from co-workers is the most effective form of learning in a professional setting, and that internal points of contact deliver two things at once, namely less pressure on IT and honest feedback about what goes wrong in practice. A few rules of thumb that go with it:

• Pick the points of contact inside the team and not in IT. The questions colleagues ask are about their own material and how to tackle it.
• Make the role visible and allow time for it. Someone doing it in the evenings on top of their job lasts two months.
• Schedule a second session three to four weeks after the first. By then the questions are real and they are about the person's own documents.
• Collect what went well and share it. One colleague showing how they now produce their weekly report convinces more than any vendor presentation.

The practising itself is half the work. Someone who brings their own draft letter, their own spreadsheet or their own meeting into the session learns more in two hours than in a day of generic examples. That is how our Microsoft 365 Copilot training is set up, and for those who want to sharpen how they phrase instructions, our prompt engineering training. The background on the features themselves is in the complete guide to Microsoft 365 Copilot.

06Stall 5: nobody wrote down what it is not for

The fifth is the quietest. Almost every rollout plan states what Copilot can do. Almost none states where it stops, and that difference decides how much trust survives.

Two kinds of boundary deserve to be written down, and they are separate. The first is about content: which work is it unsuitable for. Calculations where the exact figure matters, legal wording that has to be precise, a summary that will pass as an official record, a decision that has to be traceable. The second is organisational: which data does not go in, who checks the output before it leaves the building, and what should someone do who receives an answer they were not supposed to see.

That second list is exactly what the Deloitte governance figure points at: 21 per cent of companies have a mature model for it, while three quarters intend to work with agents that act on their own within two years. The gap widens as soon as agents stop merely answering and start performing actions. An organisation that already knows who may publish which agent, and who supervises it, will not be chasing the facts later.

In practice this is one side of A4 per team, written by the team itself. Three headings are enough: this is what we use it for, this is what we do not use it for, and this is what we always check. That sheet is more useful than any policy document, because it is written in the language of the work and because the people who have to follow it wrote it themselves.

07What does work: the same steps, a different order

What stands out about rollouts that succeed is not that different things happen. They are the same components, in a different order and with measurement around them. Four steps, which come down to: know where you stand, make it measurable, take it to the workplace, and embed it.

These four steps are also the structure of our AI adoption training, in which assessment, measurement, change management and governance are treated as one coherent route and you leave with a working adoption plan. If your first question is rather which training suits which group, the full overview is on the page with AI and Copilot training.

08Ninety days, concretely

Below is a timeline of the kind we use in practice. It assumes an organisation running Microsoft 365 that wants to introduce Copilot, with a first group of twenty to fifty people. Larger organisations run the same steps with more people per step, not with more steps.

Weeks 1 and 2: look before you buy

Read the readiness report and record who is technically eligible and who is active in the applications. Switch on Copilot Chat for everyone entitled to it and let it run for four weeks. Start an assessment of the shared environments: sites with oversized audiences, broken inheritance, sites without an owner. Run the short survey among the target group with the two questions from section 03. Appoint an AI administrator.

Weeks 3 and 4: clean up and choose

Tackle the ten to twenty highest-risk sites: strip excessive access, assign ownership, and where needed keep them temporarily out of Copilot's answers. Meanwhile, watch who picks up Copilot Chat unprompted. Use that to compose the first group, concentrated in two or three teams. Record the baseline.

Weeks 5 to 8: practise with real work

Assign licences to the first group. Run training sessions with people's own documents, organised per role rather than per application. Appoint points of contact and make their role visible. Draw up the one-page team agreement on what it is and is not used for. Schedule the second session for three to four weeks later.

Weeks 9 to 12: measure and expand

Hold the second session, now with the questions that came out of daily practice. Put the usage report and the Copilot Dashboard alongside the baseline, with the emphasis on repeat use rather than one-off opening. Repeat the survey. Use the result to compose the second group, and put the points of contact from group one to work supporting group two.

09The item that rarely gets budgeted

The budget for a Copilot track lists the seats. What it usually leaves out is the time of the people carrying the introduction: the permission clean-up, the hours of the points of contact, the second session, the analysis after ninety days. That item is modest next to the licence costs, and it determines what those licence costs return.

The arithmetic is uncomfortably simple. An organisation with two hundred licences pays a substantial sum for them each year, whether or not anyone works with them. If usage settles with a quarter of the people, you are paying four times the price per active user. The support that would have prevented that costs a fraction of the difference. What the licences cost exactly, and which items sit alongside them, is worked out in what Microsoft 365 Copilot really costs. If you are still weighing up the tool itself, Copilot, ChatGPT or Claude puts the four options side by side.

In closingA sober conclusion

A rollout that stalls rarely stalls on the wrong choice of tool. It stalls because the licence was the first thing that happened, when it should have been the fourth step. Everything in this article comes down to bringing forward three things that normally sit at the back: knowing who does the work the tool suits, knowing what sits in the shared folders, and knowing what you are going to measure.

That is no great art and it does not take a quarter of preparation. It takes two weeks of looking before you buy, and it produces a conversation in which you can substantiate what happened. In a period where one in five Dutch AI projects shows a measurable result, that substantiation is itself an advantage.

If your own rollout is stuck at one of these five places, the first step is usually not a training course but a diagnosis: where exactly is it stuck, and what is the cheapest intervention there. That is where our AI adoption training begins, and it is also where a no-obligation conversation begins.

SourcesVerified on 12 September 2026

• deloitte.com – State of AI in the Enterprise 2026, press release of 21 January 2026: 25 per cent with at least 40 per cent of pilots in production, 37 per cent surface-level use, 21 per cent with a mature governance model
• ictmagazine.nl – Only 21 per cent of AI projects deliver results in the Netherlands, on State of Integration & AI 2026 by Frends, carried out by Sapio Research among 611 decision makers, April 2026
• learn.microsoft.com – Rollout Microsoft Copilot to your organization, the recommendation to start phased and the AI administrator role
• learn.microsoft.com – Microsoft Copilot readiness report, who is eligible, application usage, and availability within 72 hours
• learn.microsoft.com – Microsoft Copilot usage report, adoption, retention and engagement, generally available within 48 hours
• learn.microsoft.com – The Copilot Dashboard in Viva Insights, adoption and impact pages, delegated access, and the under-counted email metrics between 1 June 2025 and 18 February 2026
• learn.microsoft.com – Configure a secure and governed foundation for Microsoft Copilot, remediating excessive access, interim protections and structural measures
• learn.microsoft.com – Restricted SharePoint Search, explicitly not a security boundary and intended as a temporary measure
• learn.microsoft.com – Data risk assessments in Microsoft Purview, the weekly assessment of the hundred most used sites
• learn.microsoft.com – FastTrack for Microsoft Copilot, what falls inside and outside the service, including end-user training and the adoption and change management strategy
• learn.microsoft.com – Establish champions for user adoption, learning from co-workers as the most effective form and what champions deliver