Microsoft has quietly started processing select Copilot prompts in Excel and Outlook using its own artificial intelligence models, turning away from long-time partner OpenAI and competitor Anthropic for at least some of its productivity AI tasks. The change, reported by Bloomberg on July 7, 2024, marks the first known instance of the company routing live Copilot requests through its homegrown MAI models inside widely used Microsoft 365 apps.

Inside the shift: what’s actually different

Copilot users in Excel and Outlook won’t notice a splash screen or new button. The change is entirely on the backend. When you type a request—“Create a pivot table from this data” or “Draft a reply to this email chain”—some of those prompts now travel to Microsoft’s own large language models, branded internally as MAI, rather than to OpenAI’s GPT-4 family or Anthropic’s Claude. Microsoft hasn’t published a list of which trigger words or task types are being redirected, but Bloomberg’s sources suggest the split is based on the complexity and domain of the request. Simple data manipulation in Excel or quick email drafts in Outlook appear to be early candidates.

MAI models are not new. Microsoft has been developing them in parallel with its multi-billion-dollar OpenAI partnership. They power features inside Azure, Bing, and even the Copilot sidebar in Windows, but had not been publicly acknowledged as handling live interactions inside the core Office apps until now.

What it means for you

For everyday users

If you rely on Excel’s “Add formula column” or Outlook’s “Coaching by Copilot,” you may see subtle differences in response style. Early internal testing, according to Bloomberg, shows MAI models can be slightly faster for structured tasks like formula generation but may handle tone and nuance differently when composing emails. Microsoft’s goal is parity, but users should expect some variation as the models are tuned.

Data handling may also shift. OpenAI processes prompts through servers that may reside outside Microsoft’s tenant boundaries, while MAI requests should stay within Microsoft’s own infrastructure. That could affect enterprise compliance postures, though Microsoft has not confirmed whether MAI-routed prompts are ringfenced to the customer’s geographic region.

For IT admins and compliance officers

This is the segment that needs to pay closest attention. If your organization has data-loss-prevention (DLP) policies or AI usage guidelines tied to OpenAI’s terms, you’ll want to ask Microsoft whether MAI processing falls under the same Copilot licensing agreements. Microsoft has repeatedly stated that all Copilot requests in Microsoft 365 are governed by existing enterprise data protection commitments, but the addition of a new model pipeline invites scrutiny.

Check the Microsoft 365 admin center for any updated documentation on “Copilot model selection.” As of this writing, no new toggle exists to force or block MAI usage. The routing is determined automatically by Microsoft’s servers. For regulated industries, this lack of granular control may be a sticking point.

For power users and developers

If you build prompts or custom Copilot integrations, test your workflows. MAI models may interpret instructions differently. The shift could also preview a future where Copilot supports bring-your-own-model (BYOM) scenarios, allowing enterprises to plug in a model that complies with internal policies. That’s speculation, but Microsoft has hinted at such flexibility in recent Ignite sessions.

How we got here

Microsoft’s relationship with OpenAI is deep. The company has invested an estimated $13 billion in the startup and baked GPT-4 into virtually every productivity surface under the Copilot brand. Yet simultaneously, Microsoft’s own AI research division has been building what it calls “small” and “specialized” models that are cheaper to run and easier to fine-tune for specific tasks. The MAI family, first mentioned publicly in late 2023, was originally described as a tool for Azure customers to create custom AI workloads. Its arrival in Excel and Outlook signals that Microsoft sees enough performance and cost benefit to start swapping out third-party models in production.

This isn’t a sudden divorce. Over the past year, Microsoft has gradually introduced its own models in lower-stakes copilot experiences—think Bing Chat’s creative mode or the Shopping Assistant in Edge. July’s move into Excel and Outlook, apps used by hundreds of millions of people daily, represents a new level of confidence in MAI.

Cost is likely a driver. Running prompts through OpenAI’s largest models carries a per-token fee that Microsoft must absorb to offer Copilot as a fixed-price add-on. MAI models, which run on Microsoft’s own cloud infrastructure, could slash those costs significantly. Equally important is latency: a model that runs closer to the user’s data and doesn’t traverse an external API can respond faster, a critical metric for real-time editing in Excel.

What to do now

No immediate action is required for most users. Copilot will continue to work as before; the transition is designed to be transparent. However, there are a few steps worth taking:

  • Test critical workflows. If you rely on Copilot for business-critical Excel reports or compliance-sensitive email drafts, compare outputs over the next few weeks. If you notice quality regressions, file feedback through the Copilot interface—Microsoft’s product teams actively use that data to tune the models.
  • Review your Copilot license terms. While unlikely to change overnight, the addition of a new model should be on your risk team’s radar. Reach out to your Microsoft account manager for a written statement on data residency and processing guarantees for MAI-routed prompts.
  • Educate employees. Users may perceive the shift as a downgrade if responses vary. A brief internal note clarifying that Copilot is being optimized and any changes are expected can prevent help-desk churn.
  • Watch for admin center controls. Microsoft typically follows announcement-worthy changes with configuration options. An “AI model selection” policy could appear in the Microsoft 365 Apps admin center or via PowerShell cmdlets in the coming months.

What to watch next

This is likely the first wave. Expect more Copilot endpoints—Word and PowerPoint are obvious candidates—to adopt MAI models for at least a portion of their prompts. Microsoft may also offer a “model preference” setting that lets users or admins choose between OpenAI and MAI, similar to how Bing allowed toggling between creative, balanced, and precise modes. Competitors like Google, which uses its own Gemini models across Workspace, will be watching to see if MAI-driven Copilot gains or loses credibility among enterprise users. For now, the AI power struggle inside Microsoft’s own ecosystem has officially begun.