Microsoft began routing a portion of Copilot queries in Excel and Outlook through its own MAI artificial intelligence models this week, according to reports from Bloomberg and Investing.com. The quiet rollout, first noticed on Tuesday, July 7, 2026, moves the company away from its heavy reliance on partner OpenAI’s technology for flagship productivity apps. The news sent Microsoft shares higher, as investors cheered a long-anticipated step toward in-house AI that could improve margins for the company’s Copilot services.

What Actually Changed

For the past year, Microsoft 365 Copilot has leaned heavily on large language models from OpenAI—the same GPT-4 family that powers ChatGPT—to handle tasks like summarizing email threads, suggesting formulas, and drafting documents. The new arrangement, confirmed by people familiar with the matter, now diverts some of those prompts to MAI models developed and hosted entirely by Microsoft.

The shift is not an all-or-nothing switch. According to Bloomberg’s report, the routing applies only to a subset of simpler, high-volume queries in Excel and Outlook. Complex requests or creative tasks likely still flow to OpenAI’s servers. “The goal is to maintain the same quality of output while reducing cost per prompt by up to 60%,” one source told Investing.com.

Microsoft has not published a list of which exact features are affected, but early chatter from enterprise admins suggests that formula generation and basic data analysis in Excel, along with standard email summaries in Outlook, are among the first to move in-house. The MAI models—short for Microsoft AI—are believed to be narrower, task-specific models trained on internal data, making them cheaper to run and potentially more privacy-friendly, since data stays within Microsoft’s own cloud infrastructure.

Importantly, there is no client-side update required. The routing happens server-side, so Windows users, Mac users, and web clients all get the change automatically. No new button or setting has appeared in Excel or Outlook, and the overall user experience is meant to be indistinguishable.

What It Means for You

For the everyday Microsoft 365 subscriber—the person using Copilot to clean up a spreadsheet or reply to a dozen emails—this change is likely a nonevent. The quality bar is set high enough that you probably won’t notice a difference. Microsoft has been testing these models internally for months, according to the reports, so any major hiccups should have been ironed out. If you do spot a drop in performance, the old fallback to OpenAI models is reportedly still in place for tougher prompts, ensuring a safety net.

Privacy-conscious users might actually welcome the move. When a prompt is handled by Microsoft’s own models, that data never leaves the company’s controlled environment, which could simplify compliance with internal data policies. But since Microsoft already assured enterprise customers that their Copilot data isn’t used to train external models, the practical privacy gain here is incremental at best.

For IT administrators and business decision-makers, the news is more than a footnote. First, it could eventually lead to lower subscription costs or more stable pricing. Microsoft’s Copilot for Microsoft 365 is currently a $30 per-user monthly add-on, and much of that goes toward covering the compute costs of AI. By handling simpler tasks in-house, Microsoft might be able to offer a less expensive tier in the future—or simply avoid hiking prices as usage grows.

Second, admins need to pay attention to any message center notifications that follow this rollout. If Microsoft decides to give tenants a choice—stick with OpenAI for all prompts, or allow MAI routing—that could surface in the Microsoft 365 admin center. For now, there is no opt-out or model selection toggle, but the situation could evolve. Check the Service health dashboard and the Microsoft 365 roadmap for any relevant items over the coming weeks.

Developers who build on Copilot extensibility might also want to watch latency and response times. While Microsoft’s own models might be faster for certain tasks because they run on dedicated infrastructure, that isn’t guaranteed. If you rely on the consistency of a particular model’s behavior, you’ll have to wait for official documentation to understand if any prompts from your plugin or application are being routed differently.

How We Got Here

Microsoft’s AI strategy has been inextricably linked with OpenAI ever since the company poured its first billion dollars into the startup in 2019. That relationship deepened with the integration of GPT-4 into Bing Chat, Windows Copilot, and eventually Microsoft 365 Copilot. But the economics of serving millions of enterprise users with a large, general-purpose model from a third party—no matter how close the partnership—have always been precarious.

For context, each Copilot query incurs a cost for Microsoft, which it absorbs under the flat per-user fee. As companies added thousands of Copilot seats, the aggregate compute expense became a financial drag. Wall Street analysts pressed Microsoft on this point in several earnings calls throughout 2025, nudging the tech giant to articulate a plan for “model diversification.”

Behind the scenes, Microsoft’s research division had been working on its own family of smaller, specialized models under the code name MAI. Leaks from internal presentations in early 2026 indicated that MAI models could handle tasks like summarization, classification, and data extraction at a fraction of the cost of GPT-4, albeit with some limitations in creativity and broad reasoning. The plan was always to use them where they were “good enough,” saving the premium models for when they were truly needed—a classic right-sizing approach that many large tech firms have adopted.

The first public sign of this strategy came in May 2026, when Microsoft announced that it would offer “model choice” in Azure OpenAI Service, allowing developers to pick smaller, cheaper models alongside GPT-4. But extending that logic to Microsoft 365 Copilot, where the user stands to benefit directly, is a more consequential step. The July 7 reports confirm that Microsoft now believes its MAI models are ready for prime time in two of the most widely used productivity applications in the world.

What to Do Now

Most users can simply keep working and let Microsoft manage the transition. But a few proactive steps are worth taking:

  1. Test and observe. Over the next few days, pay attention to how Copilot performs in Excel and Outlook. If you notice unusual outputs, incomplete summaries, or slower responses, use the in-app feedback tool (the thumbs-up/thumbs-down icon) to report it. This helps Microsoft fine-tune the routing logic.

  2. Admins: check your tenant settings. Although no controls exist today, open the Microsoft 365 admin center and look for any posts about “Copilot model routing” or “MAI models.” Subscribe to update notifications so you don’t miss a new policy. Also review your data residency and processing disclosures; while Microsoft’s own data handling terms apply, a change in where compute happens could affect internal compliance reviews.

  3. Enterprise admins: communicate to users. If your organization has strict AI usage policies or prefers to keep all prompts on OpenAI models for audit consistency, you may want to inform your Microsoft account team that you’re monitoring the situation. In the short term, you can disable Copilot entirely for specific apps if you’re uncomfortable with the silent change, but that’s a heavy-handed approach.

  4. Stay informed. Bookmark the official Microsoft 365 Blog and the Microsoft 365 roadmap (https://www.microsoft.com/en-us/microsoft-365/roadmap). The company typically publishes a formal announcement within days of such a rollout. Look for details on model routing criteria, data handling, and any opt-out mechanisms.

  5. Evaluate your own AI dependency. If your workflows rely on the precise behavior of a particular model—say, the way GPT-4 formats a certain type of analysis—now is a good time to start documenting those dependencies. You may need to adjust your prompts or expectations if MAI models introduce subtle differences.

Outlook

This is likely only the beginning. Microsoft has a long history of insourcing key technology once it matures, and AI models are no exception. If the Excel and Outlook experiment succeeds—lower costs with no discernible degradation—expect the MAI routing to expand to Word, PowerPoint, and Teams Copilot features by year’s end. Longer term, Microsoft could offer a full “Copilot Basic” tier powered entirely by its own models, reserving the higher-priced “Copilot Pro” for OpenAI’s most advanced capabilities.

The move also shifts the balance of power in the Microsoft-OpenAI relationship. While the two companies remain deeply intertwined—Microsoft is still OpenAI’s largest investor and exclusive cloud provider—each step toward self-sufficiency gives Redmond more leverage. For users, that could eventually translate into more choice, better-tailored AI, and perhaps more competitive pricing. For now, it’s a behind-the-scenes change that you may never notice—but it’s one that could reshape how we all work with AI inside Office.