On July 10, 2026, veteran Microsoft analyst Paul Thurrott published a stark assessment of the company’s AI strategy. In his regular “Ask Paul” column, Thurrott argued that Microsoft’s multi-year push to develop its own in-house AI models—a project known internally as MAI—risks being overtaken by the very technology it aims to serve. By the time these models are ready, AI assistants may have already transformed Microsoft 365 into a background service, rendering the traditional Office suite obsolete.

This isn’t the first warning about Microsoft’s AI timeline, but Thurrott’s analysis lands at a pivotal moment. The company has spent billions on its OpenAI partnership while simultaneously building a proprietary model family designed to reduce that dependence. Now, with AI assistants growing rapidly in capability, the window for MAI to influence the next generation of productivity may be closing.

The MAI models: Microsoft’s plan B for AI

Microsoft’s MAI effort has been an open secret in tech circles for at least two years. First reported by The Information in 2024 and later confirmed through job listings and internal reorganizations, the project aims to create large language models that Microsoft can fine-tune and deploy across its entire ecosystem—from Windows and Bing to Copilot and Microsoft 365. The goal is twofold: mitigate the cost and risk of relying exclusively on OpenAI’s technology, and enable deeper, more specialized integrations that third-party models might not support.

According to Thurrott, the MAI models have been in active development throughout 2025 and early 2026, with some prototypes already powering experimental features in Microsoft 365 Copilot. However, he stresses that a production-ready version capable of replacing or supplementing OpenAI’s GPT models across the suite is still at least a year away—likely not until late 2027 or early 2028.

That timeline matters because of what’s happening in parallel: the rapid evolution of AI assistants.

The assistant revolution: Office as a back-end service

The core of Thurrott’s argument is that AI assistants are quickly making the traditional Microsoft 365 apps—Word, Excel, PowerPoint, Outlook—feel like legacy interfaces. Today, users can already ask Copilot to draft a document, analyze a spreadsheet, or summarize an email thread without directly opening those applications. As these assistants become more autonomous, the need for a user-facing Office suite could diminish.

Microsoft itself is driving this shift. At Build 2026, the company demonstrated “Copilot Agents” that can proactively manage tasks across calendars, documents, and communications. In this model, Microsoft 365 becomes a set of APIs and data stores, while the AI assistant becomes the primary interface. Thurrott’s concern is that if MAI models aren’t ready to power these agents, Microsoft may have to lean even harder on OpenAI—or cede ground to competitors like Google, which has tightly integrated its Gemini models with Workspace.

For users, this means that the tools they depend on for work could change dramatically. An executive who today spends hours in Excel may soon interact only with a natural language prompt that requests a financial model. The underlying engine—whether it’s Excel or a custom data service—becomes invisible. If Microsoft’s own AI can’t drive that transformation, someone else’s will.

What the delay means for enterprises

Thurrott’s warning has immediate implications for IT decision-makers. Organizations that have standardized on Microsoft 365 are currently evaluating Copilot licensing and preparing for broader AI adoption. A delayed MAI model could introduce uncertainty in several areas:

  • Licensing costs: Microsoft currently pays OpenAI for each Copilot query that relies on GPT models. Bringing MAI online could lower those costs, potentially reducing the price of Copilot for enterprises. A delay means continued exposure to OpenAI’s pricing—and whatever market forces dictate.
  • Feature depth: In-house models allow for tighter optimization. MAI, for example, might be tailored to understand SharePoint metadata or Excel calculation chains in ways a generic model cannot. Without MAI, Copilot features may remain more superficial, limited to text generation and summarization rather than deep data manipulation.
  • Compliance and data residency: Many enterprises demand that AI processing happen within their own cloud tenant or geographic region. A Microsoft-owned model makes that far easier to guarantee than one from a third party. MAI’s absence prolongs reliance on a vendor whose infrastructure and data handling policies Microsoft doesn’t control.

For IT admins, the practical takeaway is to keep Copilot deployments focused on areas where the current GPT-4 models excel—email triage, meeting summaries, basic content creation—and to be cautious about investing too deeply in AI-driven workflows that might change substantially once MAI (or a competitor’s equivalent) arrives.

How we got here: Microsoft’s two-track AI strategy

Microsoft’s dual approach—partner deeply with OpenAI while secretly building a replacement—has been an exercise in hedging. The relationship with OpenAI, which dates to 2019, gave Microsoft early access to transformative technology and allowed it to leapfrog competitors in embedding AI into Office. But that partnership has also been fraught with tension. As The Wall Street Journal reported in 2025, internal friction over model sharing, compute allocation, and the direction of AGI research led Microsoft to accelerate its MAI program.

The project reportedly gained momentum after the departure of several key OpenAI researchers in late 2024, some of whom joined Microsoft directly. By mid-2025, Microsoft had assembled a team of hundreds focused specifically on training models that excel at structured data, multilingual business communication, and security—areas where generic models lag. Early benchmarks leaked to the press showed MAI outperforming GPT-4 on tasks like VLOOKUP reconstruction and legal document annotation.

Despite that progress, Thurrott notes that the project has hit predictable bottlenecks: data acquisition, safety testing, and the sheer complexity of integrating a new model family into Microsoft’s sprawling cloud services. Historical precedent isn’t encouraging—Microsoft’s previous attempts to build foundational technologies in-house, from the Windows Phone OS to the EdgeHTML engine, often struggled to keep pace with dedicated competitors.

What to do now

For most Windows and Microsoft 365 users, no immediate action is required. The transition Thurrott describes will unfold over years, not months. However, several steps can put you ahead:

  • Stay informed about Copilot roadmap announcements. Microsoft’s annual Inspire conference (usually in July) and Ignite (November) are likely venues for MAI-related news. If Microsoft introduces a “Copilot Pro” tier powered by MAI, it will signal the model’s readiness and the company’s strategy.
  • Audit your AI dependencies. If your organization has built custom workflows around the current Copilot (e.g., via Power Automate or Azure OpenAI Service), document those integrations. A shift to MAI might require updates to APIs or prompt engineering.
  • Invest in AI literacy, not tool-specific training. The skills that matter—writing good prompts, understanding a model’s limitations, designing human-in-the-loop processes—transcend any single model. Ensure your team is learning those fundamentals rather than memorizing how a particular version of Copilot works.
  • Watch the competition. Google’s Gemini and Apple’s Intelligence are also evolving quickly. If MAI slips further, these alternatives could become more attractive, especially for organizations that aren’t deeply tied to Office.

Outlook

Thurrott’s warning is not a prediction of doom, but a reminder that timing matters in technology. Microsoft’s MAI models could still reshape the productivity landscape—but only if they arrive before the landscape reshapes itself. The company’s vast enterprise install base buys some time, yet history shows that user habits can change fast once a sufficiently better alternative appears.

Keep an eye on Microsoft’s developer conferences and insider builds. The moment MAI moves from rumor to release candidate, the productivity game will change—one way or another.