Microsoft Replaces OpenAI and Anthropic Models With Its Own MAI in Excel and Outlook

Microsoft is systematically reducing its reliance on OpenAI and Anthropic by replacing their models with internally-built MAI (Microsoft AI) models in Excel and Outlook, according to Bloomberg. Tens of thousands of AI-assisted prompts now execute weekly on Microsoft’s proprietary models, a significant shift from the company’s previous dependence on third-party frontier labs.

The Scope of the Swap

The replacement is happening inside two of Microsoft’s most widely deployed productivity applications. Excel and Outlook together reach hundreds of millions of enterprise users. Moving AI functionality from OpenAI and Anthropic to internal models in these products is not a pilot: it represents a production-scale commitment to Microsoft’s own model capabilities.

Microsoft Chief AI Officer Mustafa Suleyman said in June that the company aims to reduce Anthropic spending by deploying more MAI models, according to Bloomberg’s reporting. The statement signals this is a deliberate strategic initiative, not an incremental cost-saving measure.

Revenue Implications for Frontier Labs

For OpenAI and Anthropic, Microsoft’s move reduces embedded revenue inside one of the world’s largest software suites. Microsoft has been one of both companies’ most significant customers through its Azure partnership with OpenAI and its Anthropic usage across enterprise products.

The calculus is straightforward. Every prompt that moves from an OpenAI or Anthropic API call to an internal MAI model is revenue that leaves the frontier labs and stays on Microsoft’s balance sheet. At tens of thousands of weekly prompts and growing, the volume is material.

The Cloud Provider De-risking Pattern

Microsoft’s shift fits a broader pattern among large cloud providers building internal AI capabilities to reduce dependency on frontier labs. AWS has invested in its own foundation models. Google uses Gemini across its product suite rather than licensing external models.

The pattern makes economic sense. Cloud providers pay per-token to frontier labs while simultaneously competing with them for enterprise customers. Building internal models that can handle routine tasks (the kind of work Excel and Outlook AI features perform) eliminates that per-token cost while keeping customers locked into the provider’s ecosystem.

What This Changes for Agent Infrastructure

For enterprises building agent workflows on top of Microsoft’s stack, the MAI substitution raises practical questions. Model behavior, latency, and capabilities differ between providers. Organizations that built prompts and agent logic tuned for GPT-4o or Claude Sonnet may see different results when those calls route to MAI instead.

Microsoft has not disclosed which specific MAI models are powering the Excel and Outlook replacements, or whether enterprise customers can choose which model handles their requests. That transparency gap matters for organizations that need reproducible AI behavior for compliance or audit purposes.