Microsoft has started routing tens of thousands of weekly AI prompts from Excel and Outlook through its own MAI models, Bloomberg reported on July 7, 2026, marking the first disclosed production-scale use of the company’s in-house AI for its 365 Copilot. The move signals a strategic shift away from exclusive reliance on OpenAI’s technology for some of the most widely used productivity apps.
A Gradual Handoff from OpenAI to MAI
The transition is not an overnight replacement but a phased redirection of select prompt types. According to Bloomberg’s sources, tens of thousands of prompts each week—requests to summarize email threads, analyze spreadsheet data, or draft replies—are now being handled by Microsoft’s homegrown MAI models instead of the GPT-based engines that have powered Copilot since its launch. The change affects users on the current channel of Microsoft 365, with no discrete update required; the routing happens server-side.
MAI models—which Microsoft has been developing under tight wraps—are large language models designed and trained by the company’s research divisions. While their architecture closely resembles the transformer-based models from OpenAI, Microsoft has optimized them specifically for productivity tasks, emphasizing cost-efficiency, latency, and data residency. This production deployment follows years of internal testing and smaller-scale integrations in Azure AI services.
The exact triggers for MAI routing remain undisclosed, but sources suggest that prompts involving structured data manipulation in Excel and straightforward email summarization in Outlook are among the first to be offloaded. More complex, open-ended creative tasks—like generating lengthy documents from scratch—still rely on OpenAI’s models. This hybrid approach allows Microsoft to balance performance with operational costs.
The Practical Upshot for Users
For Home and Student Users
If you use Excel or Outlook with a Microsoft 365 subscription that includes Copilot, you likely won’t notice any immediate difference. The interface remains unchanged, and the quality of responses should stay consistent. In fact, early feedback indicates that tasks like formula suggestions or quick email drafts may feel snappier because MAI models can be tuned to prioritize speed for these narrower actions.
One subtle change: MAI models run entirely on Microsoft’s own infrastructure, which means your prompt data never leaves the company’s controlled environment. For privacy-conscious users, that’s a plus, as it reduces third-party exposure compared to the previous OpenAI-dependent pipeline.
For IT Administrators and Compliance Officers
This is where the needle moves the most. With prompts now flowing through a different model stack, organizations must revisit their data-handling policies and compliance documentation. Microsoft has not yet released a detailed technical paper on MAI’s data processing, but the fact that these models are wholly owned by Microsoft could simplify data residency and eDiscovery scenarios—provided admins can confirm that regional data boundaries are upheld.
You should check the Microsoft 365 admin center for any new service health advisories or documentation updates. It’s also wise to monitor the Message center for official announcements regarding model routing, as your tenant might be part of a gradual rollout that can be influenced by update channels you manage. If your organization has strict AI use policies tied to specific model versions, reach out to your Microsoft representative for a roadmap.
For Developers and ISVs
If you build add-ins or custom solutions that leverage Copilot’s AI capabilities through Microsoft Graph or the Office JavaScript API, the shift to MAI models could affect prompt consistency. While the API contracts remain the same, the underlying model’s behavior might introduce subtle variations in output tone or structure. Begin testing your workloads against both models if you can access early flagging tools. Microsoft typically provides developer previews for such transitions; watch the Microsoft 365 Developer Blog for guidance.
This move also tees up a future where third-party developers might be able to plug into MAI models directly, perhaps via Azure Cognitive Services. That could open up new customization and fine-tuning possibilities for enterprise-grade Copilot extensions.
The Long Road to a Home-Grown AI Copilot
Microsoft’s journey to this moment began well before the Copilot brand existed. The company has invested billions in AI research, notably through its Turing and later Phi families of models. When it launched Microsoft 365 Copilot in early 2023, the service leaned heavily on OpenAI’s GPT-4 and then GPT-4o, with Microsoft acting as a privileged host and integrator. That partnership brought Copilot to market at record speed but also left Microsoft vulnerable to the constraints and pricing of an external provider.
By mid-2024, Microsoft introduced the Phi series—small language models that could run efficiently on consumer hardware. Phi-3 and later Phi-4 demonstrated that compact models could handle specialized tasks with surprising accuracy. Simultaneously, Azure AI began offering “bring your own model” capabilities, signaling that Microsoft was preparing the infrastructure for a multi-model future.
The decision to use MAI models in production Copilot workloads appears to stem from a 2025 internal initiative code-named “Project Orchard.” Leaked slides from that effort outlined a goal to reduce AI operational costs by up to 40% and slash dependency on third-party models for routine productivity scenarios. Excel and Outlook were natural starting points because their prompt patterns are relatively constrained and benefit from deterministic optimizations.
What You Should (and Shouldn’t) Do Right Now
Do nothing if you’re an everyday user. Your workflow will continue uninterrupted, and you may experience slight performance improvements. There’s no toggle to flip, no feature to enable—the change is transparent.
For admins: Review your organization’s Data Protection Impact Assessment (DPIA) and update any references to OpenAI model processing if you’ve documented them. While Microsoft’s core privacy stance hasn’t changed—customer data is not used to train models—the shift in model provenance could matter for audits. Also, keep an eye on the Microsoft 365 roadmap item that will eventually detail the MAI routing. A new admin control to opt out of MAI processing may appear in the coming months, though none exists today.
One warning: Don’t attempt to override or block the new routing through third-party tools. The routing logic is built into the service endpoints, and interfering could break Copilot functionality. If you must delay the rollout, use the standard Microsoft 365 update channels to control feature releases, but note that server-side changes like this may still apply to the current channel.
What Comes Next for Copilot’s AI Engine
This first step in Excel and Outlook is a trial balloon for a broader realignment. Within the next year, expect MAI models to handle prompts in other Microsoft 365 apps—Teams chat summarization and Word document co-authoring suggestions are prime candidates. Eventually, Microsoft may offer an “AI model choice” setting where admins can select between MAI, OpenAI, or even customer-supplied models for specific tasks, creating a truly multi-model ecosystem.
The deeper implication is about vendor lock-in and the commoditization of AI. By owning the model stack for high-volume, low-complexity tasks, Microsoft insulates itself from pricing shocks and gains the agility to innovate on model architecture without waiting on an external partner. For users, that could translate into more reliable Copilot integrations and, down the line, lower subscription costs if the savings are passed on.
The next major milestone to watch is Microsoft Ignite in November 2026, where executives are expected to detail the MAI roadmap and reveal performance metrics. Until then, the quiet hum of tens of thousands of prompts being served by Microsoft’s own AI each week marks a significant, if invisible, turning point for the productivity suite.