Microsoft’s quiet rerouting of Copilot—and why it matters
Microsoft’s new Copilot strategy is a shift in how it runs AI inside Excel and Outlook, routing a growing share of prompts to its own Microsoft AI (MAI) models instead of relying mainly on OpenAI and Anthropic, with the clear goal of cutting recurring inference costs while preserving productivity features for enterprise users. Tens of thousands of AI prompts in Excel and Outlook are now answered each week by MAI rather than external models, the first sign that Microsoft’s in-house AI stack has moved from lab experiment to live production workload. This is not a headline feature for users; it is a backend financial move with strategic teeth. In a world where every AI answer burns tokens, Microsoft is trying to turn Copilot from an impressive demo into a sustainable business.
Cost, not curiosity: the real driver behind in-house AI models
The logic behind this shift is blunt: Microsoft Copilot costs are becoming too heavy to leave in external hands. Every Copilot reply carries an inference cost, the compute bill incurred each time a model processes a prompt and returns a response, and in high-volume apps like Excel and Outlook even tiny per-prompt charges add up fast across email, spreadsheets, and workflows. Microsoft’s AI chief Mustafa Suleyman put it plainly when he said, “We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost.” At the same time, the broader industry is pulling back on AI token spending as budgets blow up; even large tech firms have burned through annual AI allocations in months as token usage spirals. In that context, MAI is less a vanity project and more a survival tactic.
Protecting Copilot quality while swapping the engine
The risky part of this move is simple: users do not care about Microsoft’s inference bill; they care whether Excel AI features and Outlook AI integration stay fast, accurate, and useful. For everyday Microsoft 365 customers, the model name behind a Copilot response matters less than whether a spreadsheet explanation, email draft, or meeting summary remains clear and timely. Microsoft is treating the MAI rollout as a workload-routing test rather than a full reset, shifting predictable, narrow tasks—like routine spreadsheet questions or email drafts—onto cheaper first-party models while keeping frontier systems for more demanding queries. If MAI produces slower, weaker, or less reliable responses, users will feel the downgrade quickly and the savings will not be worth the reputational hit. But if they cannot tell the difference, Microsoft gains a powerful argument for routing far more Microsoft 365 traffic through its own stack.
What MAI unlocks for Microsoft—and for enterprise buyers
Under the hood, MAI is growing into a full internal model family, with systems for reasoning, coding, images, voice, and transcription, including MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Voice-2, and MAI-Transcribe-1.5. One MAI model tuned for a consulting client reportedly delivers ten times better cost efficiency than OpenAI’s GPT-5.5, while a coding model matches Anthropic’s Opus 4.6 capabilities at lower cost. Microsoft’s goal is not to abandon OpenAI or Anthropic but to turn Copilot and Azure AI into multi-model platforms that pick the best model per task while giving Microsoft more control over its feature roadmap, pricing, and performance for enterprise AI spending. With hundreds of millions of Office and Teams seats, even shifting a small slice of workloads in-house moves real money—and gives Microsoft more room to adjust Copilot pricing without depending on external token economics.
The next phase: beyond Excel and Outlook to a new AI margin playbook
Microsoft’s experiment in Excel and Outlook is only the opening move. MAI already handles tens of thousands of prompts each week in those apps while still representing a small share of total AI usage, but Teams is an obvious next step, especially with MAI-Transcribe-1.5 promising multilingual transcription tuned for specialized workflows. A Microsoft-built transcription model is slated to roll out to Teams and other products in the coming months, further shrinking the bill for external AI. The most important signal now is not only whether Microsoft expands MAI across more workloads, but how transparent it becomes about which models power which tasks. For enterprises, this shift is a warning and an opportunity: AI will not stay a blank-check line item forever. Vendors are racing to reclaim margins, and in-house AI models are becoming the main lever. Microsoft’s rerouting of Copilot shows that the new AI battle is as much about cost discipline as capability.
