If you’ve used Copilot in Excel or Outlook recently, you might have interacted with a model you didn’t expect—one built by Microsoft itself, not OpenAI or Anthropic. That’s the key revelation from a Bloomberg report on July 7, which disclosed that Microsoft has already moved tens of thousands of weekly AI prompts in those two apps onto internally developed “MAI” models. The change isn’t visible to the user, and it doesn’t mean GPT is gone for good. But it does signal that Copilot is no longer a one-model product, and that Microsoft is willing to quietly route your requests behind the scenes to whichever AI engine it thinks works best—or costs least.

The Model Swap Is Real, but It’s Not What You Think

According to Bloomberg’s source, the migration is targeted: high-volume, routine tasks in Excel and Outlook—like summarizing email threads, extracting action items, or formatting spreadsheet content—are increasingly being handled by Microsoft’s own models instead of frontier systems from OpenAI or Anthropic. The report did not suggest a full cutover. Instead, it described a strategic decision to use a cheaper, more integrated model for workloads that don’t require the most expensive horsepower.

Microsoft’s sales teams are being trained to hammer this home. A July 16 report from Whalesbook detailed internal meetings where the company encouraged salespeople to position Copilot and its MAI models as faster and less expensive than competing offers from OpenAI and Anthropic—even though Microsoft continues to resell and host both partners’ technology.

So what’s actually happening? Think of Copilot less as a static brain and more as a smart router. When you ask Excel to generate a complex formula, it might still call on GPT-4. But when you ask it to summarize a column of data, a smaller, internally optimized model might do the job more economically, with lower latency, and with the added benefit of being trained on Microsoft’s own telemetry about how people use Office.

Who Wins and Who Loses: Users, Admins, and Partners

For the everyday user of Excel or Outlook Copilot, the immediate impact is subtle—maybe even invisible. If the MAI model does its job well, you might notice slightly faster responses. If it doesn’t, you could spot a drop in nuance or accuracy. The real risk is that Microsoft’s in-house models might not keep pace with rapid improvements from labs like OpenAI. But for simple tasks, the bar is lower, and cost savings are enormous at scale. Microsoft recorded $64.6 billion in infrastructure additions in fiscal 2025, and every penny saved on inference helps justify that spend.

For IT administrators, the shift is a canary in the coal mine. They’re about to see model choice become a compliance and governance headache. On July 24, unless admins intervene, all eligible commercial tenants will automatically have access to OpenAI-operated models as a subprocessor for Copilot. That means some Copilot queries could travel outside Microsoft’s own data centers to OpenAI’s servers—a move that excludes government clouds (GCC, GCC High, DoD) and lacks key compliance attestations like FedRAMP High, PCI DSS, HITRUST, or SOC 1 Type 2. Meanwhile, the Excel/Outlook model swap shows that Microsoft is already using different models for different jobs. Admins may soon need policies to decide which model families are acceptable for which data types.

For OpenAI and Anthropic, the move is a double-edged sword. Microsoft remains a major customer and cloud host, but its newfound willingness to swap models signals that the AI supplier relationship is no longer exclusive or sacred. However, Microsoft also added OpenAI as a direct subprocessor on June 23, giving customers the option to use OpenAI models explicitly. So the partnership isn’t ending—it’s becoming more transactional.

From Exclusive to Flexible: The End of a Monogamous Partnership

This strategic pivot didn’t come out of nowhere. On April 27, 2026, Microsoft and OpenAI announced a reworked deal: Microsoft’s license to OpenAI intellectual property became non-exclusive, and the two companies stopped sharing revenue on certain Microsoft resales. Microsoft retains IP rights through 2032, and OpenAI continues to pay Microsoft for cloud services through 2030. But the revised terms gave Microsoft the freedom to develop and deploy its own AI without fear of undercutting a partner.

Since then, Microsoft has moved quickly:

  • March 2026: Axios reported that Microsoft 365 Copilot Researcher gained a “Critique” layer using Anthropic’s Claude to review answers generated by OpenAI. Microsoft also introduced a “Council” mode to compare outputs from multiple models.
  • June 23: OpenAI was added to Microsoft’s Online Services subprocessor list, making it an official data processor for Copilot.
  • July 7: Bloomberg broke the news of MAI models quietly replacing OpenAI in Excel and Outlook.
  • July 9: The OpenAI subprocessor option became available for admins to configure.
  • July 16: Whalesbook detailed the shifting sales narrative.
  • July 24: Deadline for automatic enablement of OpenAI subprocessor unless admins opt out.

This timeline paints a picture not of a breakup, but of a more grown-up, less monogamous relationship. Microsoft executive vice president Charles Lamanna told Axios in March that customers wanted tools that could “change the models operating underneath” as labs leapfrog each other. The company’s multi-model “Critique” approach improved scores on the DRACO deep-research benchmark—but at the cost of higher latency. That trade-off is the heart of the new model strategy: sometimes a cheaper, faster model is good enough; sometimes you need the heavy artillery.

How to Prepare: Settings, Deadlines, and Compliance

If you’re an IT admin, circle July 24 on your calendar. That’s the date Microsoft will automatically enable OpenAI as a subprocessor for Copilot in commercial tenants—unless you actively turn it off. To opt out, go to the Microsoft 365 admin center, navigate to Copilot settings, and adjust the toggle for “OpenAI as subprocessor.” You can scope the setting by user or Entra ID group. For government or highly regulated industries, double-check that your tenant is already excluded (GCC, GCC High, DoD, and sovereign clouds are not eligible). Also review the compliance documentation: many common certifications are missing for OpenAI-operated models, so if you need FedRAMP, PCI, HITRUST, or SOC 1 Type 2, you’ll want to keep those workloads on Microsoft’s hosted models.

For anyone else, now is the time to pay attention to Copilot’s output quality. If you notice a dip, report it through in-app feedback—Microsoft tracks that closely. And if your organization is evaluating whether to stick with Copilot or use direct API access to OpenAI or Anthropic, factor in the governance benefits of the Copilot platform: unified identity via Entra, data controls through Purview, and application context from Microsoft Graph. A slightly less advanced model might be worth it if it keeps your data inside your compliance boundary and reduces shadow AI.

The Outlook: A Model-Orchestrated Future

Microsoft isn’t abandoning OpenAI; it’s building a future where Copilot can pick from a palette of models depending on the task, cost, and compliance needs. The Excel and Outlook swap is just the first visible step. In the coming months, expect to see more MAI models creep into Teams, SharePoint, and other M365 surfaces. The real test will be whether Microsoft’s in-house models can stay competitive on quality while the company trains its sales force to sell “good enough” as “better integrated.” For now, the message is clear: the AI behind your favorite Office apps is getting a corporate reorg. And you might not even notice—until you do.