Microsoft has begun replacing some of the large language models from OpenAI and Anthropic that power its Microsoft 365 Copilot features with its own in-house MAI models, Bloomberg reported on July 7, 2026. The shift, already rolling out in apps like Excel and Outlook, marks a significant step in Microsoft’s plan to reduce its reliance on external AI providers while optimizing performance and cost for its productivity suite.

What’s actually changing under the hood

Copilot in Microsoft 365 apps such as Excel and Outlook now routes certain AI tasks to Microsoft’s proprietary MAI models instead of exclusively using OpenAI’s GPT-4 or Anthropic’s Claude. This is a server-side change—users won’t see a toggle and don’t need to change any settings. The MAI models handle a growing share of Copilot’s intelligence, from generating formulas and analyzing data in Excel to summarizing email threads and drafting replies in Outlook.

The transition is gradual, and some features still rely on partner models. Microsoft hasn’t disclosed the exact architecture or training data behind MAI, but the models are understood to be fine-tuned for productivity scenarios, potentially leveraging Microsoft Graph to better understand user context. Early anecdotal reports suggest that Copilot responses may now arrive with slightly lower latency, though Microsoft has not published official benchmarks.

The change does not affect the user interface or the core Copilot experience—your prompts and the results still look the same. However, the AI reasoning and text generation are increasingly handled by models built entirely by Microsoft.

What this means for you

For everyday users

If you use Copilot in Excel or Outlook, you probably won’t notice a dramatic difference. Some routine tasks might feel snappier, while others could produce slightly different phrasing or suggestions. The underlying model switch is designed to be seamless. No action is required on your part; just keep using Copilot as you normally would.

For IT administrators and compliance teams

This shift could have significant implications for data governance. When Microsoft processes Copilot prompts through its own models, the data still resides within the Microsoft 365 service boundary, but the training and inference pipelines may differ. Admins should check Microsoft’s documentation on data residency and handling for MAI models once it becomes available. Key considerations include:
- Data processing location: Are MAI inference runs confined to your tenant’s geographic region?
- Model transparency: What safety measures and human oversight apply to Microsoft’s own models?
- Audit and compliance: How do MAI outputs align with your organization’s AI usage policies?

Microsoft typically updates its compliance certifications for new AI services, so watch for entries in the Microsoft 365 admin center or the Service Trust Portal.

For developers integrating with Copilot

Developers who build add-ins or workflows on top of Copilot extensions should monitor the Microsoft 365 Developer Blog for any changes in response formatting or API behavior. While the high-level APIs are unlikely to break, subtle shifts in model behavior could affect automated tests or user-facing outputs. If your solution relies heavily on Copilot in Excel or Outlook, consider running a regression suite to spot any unexpected variations.

How we got here

Microsoft’s road to MAI is paved with both ambition and economic pragmatism. Here’s a quick timeline of the events leading up to this switch:

  • 2019–2023: Microsoft invests $13 billion in OpenAI, gaining exclusive rights to run models like GPT-4 on Azure and embedding them into products—first Bing, then GitHub Copilot, and eventually the Microsoft 365 Copilot.
  • November 2023: Microsoft announces its own custom AI accelerator chip, Maia 100, signaling a desire to reduce dependency on Nvidia and control its AI infrastructure.
  • Early 2024: The company launches small language models (SLMs) under the “Phi” family, capable of running on devices like Copilot+ PCs. While not directly labeled MAI, these models demonstrate internal AI development muscle.
  • Late 2024–2025: Reports surface that Microsoft is testing its own foundational models for conversational AI, with internal teams exploring ways to cut the high per-query costs of OpenAI APIs. Bing Chat starts blending multiple models, including some homegrown ones.
  • July 2026: Bloomberg reports that Microsoft has begun replacing some OpenAI and Anthropic models with MAI in Microsoft 365 apps like Excel and Outlook.

This progression mirrors industry-wide cost optimization efforts. Running large models in the cloud is expensive, and subscription-based Copilot licenses don’t cover that overhead indefinitely. By switching to its own models, Microsoft can better tune performance for Office scenarios—where a deep understanding of spreadsheets and emails is more critical than broad conversational ability—and plausibly keep costs in check for the long term.

What to do now

For most people: Keep using Copilot as before. The transition is occurring automatically, and Microsoft intends for it to be invisible.

For enterprises with strict compliance requirements, take these steps:
1. Watch for official guidance: Microsoft typically communicates AI model updates through Message Center posts and the Microsoft 365 Roadmap. Bookmark the admin center and check for entries related to “Copilot model updates” or “MAI.”
2. Review data handling policies: When Microsoft publishes documentation on how MAI models process data, compare it against your organization’s data handling and security standards.
3. Test workflows: If your team relies on Copilot for critical tasks in Excel or Outlook, designate a few power users to compare outputs before and after the rollout. Document any discrepancies in analysis or language.
4. Engage with Microsoft support: Enterprise customers with premier support can open a case to ask about model switching, governance, and any impact on service-level agreements.

For developers: No immediate code changes are required, but stay tuned to the developer blog for any new APIs or deprecations. You might also start planning for a future where MAI models offer customization options—perhaps fine-tuning on your organization’s data.

What’s next

All eyes will be on Microsoft’s official confirmation and any benchmarks that compare MAI with the previous OpenAI and Anthropic models. Performance, accuracy, and cost savings will determine how fast the transition accelerates. Expect the company to expand MAI to other Copilot experiences—Word, PowerPoint, Teams—over the next months.

We’ll also be watching for new Copilot capabilities that only MAI can enable, such as tighter integration with Microsoft Graph or more advanced reasoning on structured data inside spreadsheets. If successful, this move could pressure other cloud AI providers to double down on proprietary models, reshaping the enterprise AI landscape in the process.