Microsoft has started swapping out third-party AI models that power Copilot in some Microsoft 365 apps, replacing them with its own in-house models, according to a Bloomberg report. The change affects core productivity tools like Excel and Outlook, marking a significant shift in how the company delivers AI features to millions of users.

The backend swap nobody saw coming

Behind the scenes, Microsoft is reportedly substituting certain OpenAI and Anthropic models with its own Microsoft AI (MAI) family of models. The transition is not a wholesale rip-and-replace but a gradual integration—specific tasks and applications are being moved to the new models while others remain on existing infrastructure for now. Bloomberg’s sources, who requested anonymity because the plans are not yet public, confirm that Excel and Outlook are among the first applications to receive the new MAI-powered Copilot experience, with potential expansion to Word, PowerPoint, and Teams on the horizon.

The exact list of features affected remains under wraps. Copilot in Excel relies on AI for formula suggestions, data analysis, and chart generation; Outlook’s Copilot summarizes email threads, drafts replies, and surfaces priority messages. It’s likely that only a subset of these capabilities is currently running on MAI. Microsoft has not released build numbers, version identifiers, or a precise rollout timeline, so end users cannot easily confirm which model is handling their requests at any given moment. The user interface remains unchanged—the Copilot button stays the same—but the intelligence behind it is quietly morphing.

What this means for your daily work

For the average information worker, the immediate effect should be invisible. If you ask Copilot to summarize an email chain or suggest a formula, it will still deliver a result. However, regular users—especially power users who craft complex, multi-step prompts—may notice subtle shifts in response style, accuracy, or even speed. MAI models are trained to perform well on specific Office tasks, but they may interpret ambiguous requests differently than the GPT-4 class models they replace.

IT administrators and compliance officers should pay closer attention. When a third-party model like OpenAI’s processes data, certain data handling and residency agreements apply. An in-house model raises new questions: where does the inference run, and how is training data managed? Microsoft has not yet published updated privacy or data processing documentation specific to MAI. Until that happens, enterprises in regulated industries should treat the change as a potential shift in data flow and seek precise guidance from their Microsoft account teams. The Microsoft 365 admin center and Message center are the primary channels to watch for official disclosures; admins should also review the Data Protection Impact Assessment (DPIA) if applicable.

Developers who build on top of Copilot’s APIs, or who extend Office with custom AI plugins, might eventually need to adjust. If Microsoft gradually exposes MAI-exclusive endpoints or deprecates certain OpenAI-backed features, integrations could break. For now, the full Copilot API stack remains unchanged, but the underlying model transition signals a future where Microsoft may offer its own model fine-tuning and customization options—and possibly restrict third-party model choices.

The long road to homegrown AI

Microsoft’s Copilot story began with a heavy bet on OpenAI. The company’s 2019 investment, followed by billions more, gave it early access to the GPT family, which powers everything from Bing Chat to GitHub Copilot. But running OpenAI’s large models at scale for hundreds of millions of Office users is expensive. Each query passes through cloud infrastructure that incurs per-token costs, and as Copilot adoption grows, so does the bill.

The development of MAI (Microsoft AI) models has been an open secret. Over the past year, Microsoft has released smaller, task-specific models like Phi-3, optimized for on-device and low-latency scenarios. The broader MAI family—reportedly including models trained on internal data and tailored for productivity tasks—now appears mature enough to handle mainstream Office workloads. By swapping in its own models, Microsoft gains what every cloud provider wants: vertical integration that controls cost, performance, and data governance.

This isn’t Microsoft’s first pivot away from pure dependence on OpenAI. Earlier versions of Copilot already use fine-tuned versions of GPT models and sometimes route simpler requests to smaller, cheaper models. The Bloomberg report underscores that cost control is a primary motivator; running MAI could slash per-query expenses significantly, especially for high-frequency, low-complexity actions like email categorization or basic chart generation.

The move also mirrors industry dynamics. Rivals like Google embed their own Gemini models across Workspace, while Apple is developing on-device AI that stays entirely local. By reducing reliance on external partners, Microsoft positions itself to offer more predictable pricing, tighter privacy controls, and faster innovation cycles—provided its models perform well.

What you should do now (and nothing to panic about)

For the vast majority of Office users, there is no action item. Keep your Microsoft 365 apps updated through the usual channels—the MAI transition will arrive as a server-side change, not a client patch. If you want to stay ahead, you can:

  • Monitor Microsoft 365 Roadmap and Message center. Administrators should subscribe to update notifications; Microsoft typically posts major AI model changes as roadmap IDs or admin announcements. Search for “Microsoft AI” or “Copilot model update.”
  • Test before trusting. If your job depends on precise Copilot output—for legal document review, financial formulas, or high-stakes summarization—run a few side-by-side tests with previous results to see if quality drifts. Use the in-app feedback buttons (thumbs up/down) liberally; Microsoft uses that signal to tune MAI performance.
  • Ask the right compliance questions. If your organization has strict data handling rules, send a formal inquiry to your Microsoft representative asking whether MAI processing occurs within your tenant’s existing data boundary, whether prompts are logged differently, and if the models have been audited for bias or factual consistency. Expect answers to trickle out slowly.
  • Experiment with prompt specificity. Early MAI models may respond better to highly structured prompts than the more conversational style that GPT-4 favors. If you notice a drop in output quality, try adding more context or breaking the request into smaller steps.

On the user side, you might eventually see subtle UI cues—a different loading indicator right before a feature update, or a discrete note in the Copilot settings pane—but for now, the change is entirely under the hood.

What’s next: a more Microsoft-shaped AI future

Expect this to be the opening move, not a one-off event. Microsoft will likely expand MAI to Word, PowerPoint, and Teams throughout 2024 and 2025, targeting features where quick, cost-efficient inference matters most. The company’s annual Ignite conference (typically held in the fall) will almost certainly include sessions detailing the MAI roadmap and perhaps early benchmarks comparing in-house models to OpenAI’s.

The strategic question looming over the partnership: how will OpenAI react? Microsoft retains a financial stake and ongoing collaboration, but if MAI proves up to the task, the days of exclusive GPT integration inside Office could be numbered. For consumers and enterprises, that could mean a Copilot that feels faster, cheaper, and more tightly woven into the Microsoft 365 ecosystem—but also one that is less connected to the cutting-edge advances coming from OpenAI’s research lab. As always, the proof will be in the spreadsheet.