Microsoft 365 Copilot: The Numbers Don't Lie

... But They Don't Tell the Whole Story Either

There's a lot of noise around Microsoft 365 Copilot right now. Microsoft's marketing machine is in full swing, and the headlines look impressive. But if you peel back the numbers, the reality is a bit more nuanced — and frankly, more interesting.

The headline figures

Microsoft 365 now has more than 450 million commercial seats worldwide, yet Microsoft 365 Copilot has reached roughly 15 million paid seats — that's around 3.3% conversion. Seventy percent of Fortune 500 companies have adopted Microsoft 365 Copilot, but for most, adoption means pilots and phased rollouts rather than enterprise-wide deployment. 

The market signal is clear: Copilot adoption is broad but shallow. Enterprises across industries are testing the waters, but few have fully committed.

The biggest blocker isn't cost or capability — it's governance. Since Copilot can access data from across the M365 tenant — SharePoint, Teams, OneDrive, and Exchange — data governance concerns are the biggest roadblock to adoption. Legal, compliance, and data security teams are all expressing apprehension about oversharing. Regulators in the EU and UK are scrutinising Copilot closely. The Dutch government commissioned a data protection review which identified transparency, retention, and accuracy gaps.  For UK organisations in particular, this is a real consideration before hitting that deploy button.

There's also a competitive picture worth noting. Among paid AI subscribers, ChatGPT holds 55.2% paid subscriber market share, Gemini 15.7%, and Copilot 11.5% — down from 18.8% in July 2025, a 39% contraction in just six months.  Distribution and adoption, it turns out, are very different things.

The OpenAI vs Anthropic question

Here's where it gets interesting, and something I've been watching closely. Many people assume Copilot is OpenAI. It was, until recently. On 24 September 2025, Microsoft integrated Anthropic's Claude models into Microsoft 365 Copilot, allowing enterprise customers to choose Anthropic in addition to OpenAI for specific tasks. OpenAI models remained the default.

Then things moved quickly. From January 2026, Anthropic's Claude models were enabled by default in Microsoft 365 Copilot for most commercial tenants worldwide.  Worth noting: for customers in the EU, EFTA, and United Kingdom, the Anthropic toggle defaults to Off — so UK admins need to actively opt in if they want Claude features.

GitHub Copilot paid users now primarily rely on Claude Sonnet 4 when using Visual Studio Code's automatic model selection. That's a significant shift that most IT managers haven't noticed yet.

Why does this matter?

Microsoft's move to a multi-model strategy isn't just a technical footnote. Anthropic agreed to buy over $40 billion of capacity on Microsoft's Azure cloud as it repositioned Claude to be delivered to Microsoft's Copilot customers.  This is a strategic realignment of enormous scale — Microsoft is deliberately reducing its dependency on any single AI provider.

Copilot will continue to be powered by OpenAI's latest models, and customers now have the flexibility to use Anthropic models too — starting in the Researcher agent or when building agents in Microsoft Copilot Studio.  The pitch from Microsoft is "best model for the task" rather than "one model to rule them all."

So my question has to be added to companies security thoughts and readiness of their own tenants for M365 Copilot, does the multi-model approach just add confusion to the end user, Copilot with OpenAI, Copilot Cowork with Claude, is this not leading customers to go straight to their provider of choice.

Only time will tell and the revenues that get reported on the next quarterly earnings call.

The practical takeaway

If your organisation has M365 licences and you haven't seriously evaluated Copilot yet, the governance question needs to come first. Who has access to what data? Is your SharePoint tidy enough that you'd be comfortable with an AI summarising it for anyone in the business?

For those already on the Copilot journey, the model landscape is shifting underneath you. The AI powering your meeting summaries today may not be the same one doing it next quarter — and that has implications for data processing agreements, especially for UK and EU organisations.

This isn't a reason not to adopt. It's a reason to go in with your eyes open.