Microsoft plans to swap out OpenAI and Anthropic models powering Copilot in Microsoft 365 apps—including Excel, Outlook, and others—for its own MAI model family, a move reportedly set to begin rolling out in July 2026. The shift, aimed at reining in ballooning AI inference costs, promises tighter integration and faster response times but raises questions about quality parity and enterprise readiness.

The switch: From third-party to homegrown AI

According to reports, starting in July 2026, Microsoft will route a significant portion of Copilot prompts inside key productivity applications away from models provided by OpenAI and Anthropic and toward its proprietary MAI models. The rollout is expected to hit Excel, Outlook, and other heavily used Microsoft 365 tools first, with broader coverage over time. Microsoft has not publicly confirmed the timeline or scope, and the company declined to comment on the record.

The MAI model family—short for Microsoft AI—has been in internal development for several years, but its commercial debut inside Copilot marks a dramatic escalation in Microsoft’s AI self-reliance strategy. Unlike the general-purpose models from OpenAI that currently handle a wide range of language tasks, MAI models are purpose-built for efficiency in specific productivity scenarios. That means they could be smaller, faster, and cheaper to run, while still delivering the structured output needed for spreadsheet formulas, email drafting, or data analysis.

What this means for everyday users

If you use Copilot in Excel to generate formulas, analyze trends, or clean data, the underlying model change might be largely invisible—at least at first. Microsoft likely will tune the MAI models to match the output style and accuracy users have come to expect. However, subtle differences could surface: more concise suggestions, slight variations in generated text, or different handling of ambiguous natural language queries. In Outlook, email summaries and reply drafts might feel more templated or less “creative” under the new models.

For business users and admins, the impact could be more pronounced. Copilot responses tied to sensitive enterprise data must remain accurate and compliant. A shift to a new model, even one developed in-house, introduces unknowns around regression bugs, prompt injection vulnerabilities, and performance consistency. IT teams may need to monitor Copilot’s output quality closely during the transition, especially in critical workflows like financial modelling or executive communications.

The cost equation driving the change

Microsoft’s Copilot subscription—priced at $30 per seat per month for Microsoft 365—is a high-margin product, but the underlying inference costs eat into profits. Each prompt requiring a call to OpenAI’s large language models carries a fee, and with hundreds of millions of monthly active users across Office apps, those fees add up quickly. By switching to its own MAI models, Microsoft can slash that expense dramatically, potentially improving margins or even passing savings on to customers through lower subscription prices or expanded feature sets.

The move aligns with a broader industry trend of AI model commoditization. OpenAI’s GPT series, while powerful, is increasingly rivaled by open-source alternatives like Llama and Mistral that enterprises can host themselves. Microsoft’s investment in OpenAI remains significant, but the Redmond giant is not content to be solely a distribution partner; it wants to control the technology stack end-to-end. The MAI effort signals that Microsoft sees AI as infrastructure, not merely a feature, and it’s building its own rails.

A timeline of Microsoft’s AI independence

How did we get here? Microsoft’s partnership with OpenAI dates back to 2019, with billions invested and deep integration of GPT-4 into Azure, Bing, and eventually Copilot. But even as that partnership flourished, Microsoft began assembling the pieces for greater autonomy:

  • In 2023, it quietly formed a dedicated “AI Platform” division tasked with developing foundation models separate from OpenAI’s.
  • Throughout 2024 and 2025, the company hired leading researchers and acquired startups specializing in efficient model architecture and reinforcement learning from human feedback.
  • Leaked internal benchmarks from early 2026 showed MAI models matching or exceeding GPT-4o on domain-specific tasks like spreadsheet reasoning and email triage, while using a fraction of the compute.

By mid-2025, Microsoft had already begun testing MAI models inside internal productivity tools, collecting feedback from its own employees. The July 2026 public rollout, if it materializes, would be the culmination of over three years of parallel development, underpinning a strategic bet that homegrown AI is not just cheaper, but better suited to the nuanced demands of enterprise productivity.

What you should do now

For most individual users, no immediate action is required. Copilot will continue to function as before, and any model changes will be rolled out server-side, without software updates needed on your device. However, there are a few steps worth considering:

  • Stay alert for quality changes. If you notice Copilot in Excel suddenly misunderstands a query you’ve used reliably before, or Outlook’s tone shifts, report the feedback via the thumbs up/down icons. Microsoft uses this telemetry to fine-tune model behavior.
  • Review any automated workflows. Power users who have built macros or scripts that rely on Copilot’s output format should test them periodically after July 2026 to ensure compatibility.
  • For IT administrators: Consider enabling a phased rollout group for Microsoft 365 updates if your organization relies heavily on Copilot. This will give you a window to evaluate any model-driven changes before they reach all users. Also, keep a close eye on the Microsoft 365 roadmap for official announcements regarding MAI model availability.

Looking ahead: More features, faster innovation

If the MAI transition proves successful, it could unlock a cascade of new capabilities. Smaller, more efficient models can run entirely on-device, enabling offline Copilot experiences and reducing latency for time-sensitive tasks like real-time meeting transcription. It could also pave the way for deeper customization, allowing enterprises to fine-tune models on their own data without relying on third-party APIs. For users, the endgame is a Copilot that feels faster, more predictable, and possibly cheaper—if Microsoft shares the windfall. The next six months will reveal whether this is a masterstroke of vertical integration or a miscalculation that hands rivals an opening.

In the interim, the AI assistant that millions rely on in their daily workflows is about to get a new brain. Whether you notice the difference depends on how carefully you’re paying attention.