Microsoft Quietly Swaps OpenAI and Anthropic for Its Own AI in Excel and Outlook

Microsoft just started quietly pulling the plug on OpenAI and Anthropic inside two of its most-used apps. Excel and Outlook are now routing tens of thousands of AI prompts every week to Microsoft’s own MAI models instead, according to a Bloomberg report from 7 July 2026. For UK businesses paying for Microsoft 365 Copilot, the AI doing the work behind the scenes just changed — and most users won’t have noticed.

What’s Actually Changing

Microsoft has built its own family of AI models, branded MAI (Microsoft AI), and started shifting everyday Copilot tasks in Excel and Outlook onto them. Think the bread-and-butter stuff: summarising a long email thread, drafting a quick reply, formatting a spreadsheet, cleaning up a data table.

This isn’t a full break-up. OpenAI’s frontier models still handle the harder, more demanding tasks where raw reasoning capability matters. Anthropic’s models remain embedded in specific Office applications for particular use cases too. Microsoft’s own language makes clear this is a routing decision, not a divorce — cheaper in-house models for routine work, expensive frontier models kept for the jobs that actually need them.

Mustafa Suleyman, who leads Microsoft AI, put the motivation bluntly: “We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost.” No ambiguity there. This is about margin, not model quality.

Why This Matters More Than It Sounds

Microsoft has been OpenAI’s single largest backer, pouring in an estimated $13 billion-plus and building its entire Copilot strategy on top of GPT models for the past three years. Watching Microsoft actively route paid customer traffic away from that partnership, even partially, is a genuine shift in the AI industry’s power dynamics.

UK investors keep asking whether this signals OpenAI’s position is weakening. Falls apart fast if you take that reading too literally — Microsoft still uses OpenAI extensively, and the OpenAI-Microsoft relationship includes deep infrastructure ties that aren’t unwinding overnight. What’s actually happening is more mundane and more important: the AI cost model that made sense in 2023, when nobody had viable in-house alternatives, doesn’t make sense anymore once you’ve spent three years and billions of dollars building your own.

The Bigger Pattern: Every Major Player Building In-House

Microsoft isn’t alone in this move. Amazon has poured resources into its own Nova models while still hosting Anthropic’s Claude prominently through Bedrock. Google, obviously, has always run predominantly on its own Gemini models. The pattern across the industry is consistent: use frontier third-party models to prove the product works, then build in-house alternatives to capture the margin on routine usage once you understand exactly what your customers actually need.

This is a maturing market behaving like every tech market before it. Cloud computing went through an identical phase — companies rented compute from AWS, then built private infrastructure once usage patterns became predictable enough to justify the capital investment. AI inference is following the same curve, just faster.

What This Means for the Companies Being Displaced

Neither OpenAI nor Anthropic loses Microsoft’s business outright, and both companies have diversified their customer bases significantly beyond any single platform. Anthropic in particular has leaned hard into enterprise API customers and Claude Code adoption among developers, a market that doesn’t route through Microsoft’s Office suite at all.

Still, losing “tens of thousands of prompts weekly” from two of the world’s most-used productivity apps is a real revenue hit, even if it’s a small percentage of either company’s total business. I’ve seen this pattern before in enterprise software: a big platform partner quietly reduces reliance over eighteen months, then the relationship gets renegotiated on very different terms two years later.

Nine times out of ten, this is how platform dependency actually erodes — not with a dramatic breakup announcement, but with a slow, deliberate rerouting that only becomes obvious in hindsight.

UK Businesses Using Microsoft 365 Copilot

If your organisation pays for Microsoft 365 Copilot, this change happens entirely behind the scenes — no action required, no interface change, no new setting to configure. The model handling your Excel formula suggestion or Outlook email draft may simply be different than it was last month.

Whether that matters practically depends on what you’re using it for. Routine formatting and summarisation tasks are unlikely to show a noticeable quality difference between MAI and GPT-based models — that’s precisely why Microsoft chose these tasks first to migrate. Complex analysis or nuanced writing tasks should, in theory, still route to the more capable frontier models.

UK data protection teams should note: Microsoft’s own MAI models processing your Office data changes the data flow compared to routing through OpenAI or Anthropic’s infrastructure, even if the practical UK GDPR compliance posture stays broadly similar under Microsoft’s existing enterprise agreements.