M365 Copilot a year on: admin reality vs the sales pitch

Microsoft 365 Copilot reached general availability for enterprise customers in November 2023 at $30/user/month, a significant additional cost on top of existing M365 licensing. We're now past the initial wave of excitement, past the early adopter period, and into the reality of what it's actually like to deploy and manage this thing.

I've been involved in several Copilot deployments and spoken with colleagues and customers across a range of sizes and sectors. Here's what I've found.

What the sales pitch says

The demo scenario: meeting finishes, Copilot summarises it, drafts action items, creates a follow-up email, builds a PowerPoint. User types a natural language question about company data, gets an instant synthesised answer. AI-generated first drafts, reduced meeting overhead, faster document creation.

These things are real. They work in demo. They work for some users in production.

What the admin reality looks like

Data governance is suddenly critical. Copilot has access to everything the user has access to in Microsoft 365: SharePoint, Teams, emails, documents. If your data governance is poor (overly permissive sharing, sensitive data without appropriate access controls, documents without sensitivity labels), Copilot will surface that data to users who probably shouldn't have it.

This isn't a Copilot bug. It's working as designed. If Sarah in HR has access to a shared SharePoint site with salary information because someone added her to the wrong group three years ago, Copilot will include that information in Sarah's Copilot responses. Data governance problems that were invisible become visible through Copilot.

User adoption is variable. The users who love Copilot tend to be high-volume document and email workers: executives, knowledge workers, people who spend their days in meetings. For these users, the productivity gains are real and they'll tell you enthusiastically.

For users who do more structured, task-based work (field service, process-heavy roles, frontline workers), Copilot is less naturally applicable. Deploying Copilot licences broadly without thinking about use case applicability results in expensive licences that aren't being used.

Prompting skills matter. Copilot's output quality is highly dependent on input quality. Users who write good prompts get useful output. Users who write vague prompts get vague output. The expectation that it "just works" without any user training leads to disappointment. A structured Copilot adoption programme (helping users understand what it's good for and how to prompt effectively) makes a measurable difference to adoption and satisfaction.

The ROI question

At $30/user/month, Copilot needs to justify itself. For power users (the executive who spends four hours a day in meetings and needs everything summarised, the consultant who generates proposals and documentation constantly), the ROI case is real. Time savings of 30-60 minutes per day at senior employee cost rates makes the maths work.

For everyone else, it's harder to demonstrate. "Broadly useful" doesn't translate to a specific time saving that justifies the licence cost.

My recommendation: start with a defined group of high-value users where the use case is clear, build evidence of value, then expand based on demonstrated ROI rather than broad deployment.

Where things actually stand

M365 Copilot is a real product that provides real value for the right users and use cases. It's not the universal productivity transformation it was marketed as. The data governance requirement is more demanding than most organisations anticipated. The adoption work is substantial.

If you went in expecting magic and got something useful-but-complicated, that's about right for where we are. The product is improving. Each update adds capability. The organisations that are getting genuine value are the ones who invested in the foundations: data governance, user adoption, defined use cases.

That's also, frankly, how most significant technology deployments work. This one's no different.