Microsoft’s MAI Pivot: What It Is and Why It Matters
Microsoft’s shift to internal AI models in Excel, Outlook, and other Microsoft 365 apps is a strategic move to route Copilot prompts through its own MAI systems instead of relying mainly on OpenAI and Anthropic, reshaping how enterprise AI is paid for, controlled, and priced over the long term.
To cut costs, Microsoft is pushing to have more Microsoft 365 AI prompts answered by its internal MAI models rather than by Anthropic and OpenAI models. Tens of thousands of prompts in Excel and Outlook are already completed each week with these in-house systems, where the products previously leaned more heavily on external models. This is not a lab experiment; it is the first disclosed production-scale shift of Copilot traffic towards Microsoft internal AI models inside core productivity apps. In other words, the company is testing a new economic engine for Copilot right in your spreadsheets and inbox.
The Cost Story Behind Copilot: Tokens, Budgets, and MAI
The move is driven by a stark reality: Copilot burns through huge volumes of AI tokens, and those tokens are expensive when bought from someone else. For now, Microsoft gets much of that technology at a discount through its long-standing partnership with OpenAI, but that arrangement will not last forever, and the company does not want its margins dictated by another lab’s pricing. Quote-worthy or not, the logic is obvious—enterprise AI spending cannot remain an open bar.
Microsoft’s CEO of AI, Mustafa Suleyman, has been unusually blunt. “We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost,” he said at its Build conference. This thinking extends beyond Copilot: the company has already moved GitHub Copilot users to token-based billing and pulled employees’ access to Claude Fable over data-retention concerns, clear signs that AI usage is being reined in and metered more tightly. Microsoft pushing everyone toward its own models makes Copilot cost reduction easier to manage, even if the return on investment remains hard to measure.
Performance Trade-Offs: Will Copilot Feel Different?
Users care less about token economics and more about whether Copilot gets their work done. Microsoft insists the experience inside Office is unlikely to change much, since the model behind a feature can swap out while the same Copilot surface stays put. In practice, this means your Excel formula explanation or Outlook email rewrite might already be powered by MAI without you noticing. For now, Microsoft is routing only a slice of Excel and Outlook traffic—tens of thousands of prompts per week—to these internal models.
Still, shifting brains under the hood is not risk-free. Feature parity and performance across Microsoft 365 apps will depend on how quickly Microsoft internal AI models close the gap with frontier systems. At Build, the company launched seven MAI models, including an agentic coder and a text-to-image generator, and claims one can match the coding ability of a popular prior-generation Anthropic model at lower cost. MAI-Code-1-Flash, its coding assistant, is pitched as a competitor to Claude Code and OpenAI’s Codex, and these MAI models are already available in Copilot for Business and Enterprise users. The bet is clear: “good enough” intelligence that is cheap and controlled beats “best in class” that is rented.
A New Microsoft 365 AI Strategy: Owning the Stack
For years, Microsoft advertised that large parts of Office 365 ran on models from both OpenAI and Anthropic. That story is changing. Although it still relies on these third-party systems, it has increasingly worked to stand up its own agents and is now funneling at least some prompts from Excel and Outlook to its own AI models. This is the clearest sign yet of a Microsoft 365 AI strategy built around owning the stack—from infrastructure to models to the Copilot interface—rather than being a permanent reseller of external intelligence.
The OpenAI deal, renegotiated to let Microsoft build competing models while keeping a license through the decade, freed it to walk this path. Microsoft has already made a big push for its new internal models and has announced plans to release its own Teams-based transcription model. Combined with earlier moves—reducing workers’ access to external tools like Claude Code after costs skyrocketed—this shows a pattern: third-party models are now a bridge, not a destination. Enterprise buyers should read this as a signal that vendor lock-in in AI will increasingly be controlled by platform owners, not model labs.
What It Signals for Enterprise AI Spending and Copilot’s Future
Microsoft’s belt-tightening is not an isolated act; it is part of a wider trend in enterprise AI spending. After a burst of heavy token consumption earlier in the year, large firms including Amazon, Uber, Meta, and Accenture have reportedly moved to curb spending as token costs spiral. Uber reportedly spent its entire annual AI budget in the first three months of this year, and an undisclosed company spent half a billion dollars on AI tokens in a single month after leaving engineer usage uncapped. AI tokenmaxxing leaderboards going out of fashion is no coincidence; budgets are biting back.
In this context, Microsoft internal AI models are a defensive and offensive play at once: they restrain Copilot cost growth and tighten Microsoft’s grip on its own AI destiny. For Copilot users, the near-term impact is subtle—same interface, evolving engines—but the long-term implications are big. If Microsoft succeeds in routing most Microsoft 365 AI prompts through MAI, it will not only shrink payments to labs like Anthropic, as Suleyman openly aims to “reduce and ultimately eliminate” those costs, but also redefine how AI features are priced and bundled in enterprise software. The message to customers is clear: the Copilot you see today is the front end of a deeper shift in who builds, owns, and profits from your AI workflows.
