Microsoft wants to reduce its dependence on the most widely used artificial intelligence models. In June, the American giant officially launched MAI, a family of proprietary, state-of-the-art AI systems. The goal is to widen its competitive edge and avoid a slow, gradual defeat at the hands of OpenAI and Anthropic.

Following an initial presentation, Bloomberg now reports that Microsoft is quietly rolling MAI into Excel and Outlook. The mechanism: user prompts are routed to Microsoft’s own models rather than to GPT. According to early data, tens of thousands of queries a week are already being processed on Microsoft’s in-house technology instead of OpenAI’s.

As Money.it International has covered before, this fits a broader pattern of AI giants racing to control both the technology and its distribution — see our earlier piece on how Sam Altman keeps trying to sell AI even without a clear market.

What we know about Microsoft MAI

At the Build 2026 event, Microsoft showed off seven MAI models for the first time. Among them is the company’s first reasoning system, MAI-Thinking-1, which is already said to match Anthropic’s Claude Opus 4.6 on coding tasks.

Does this mean OpenAI and Anthropic will disappear from Microsoft’s products? Not for now. The strategy is a three-way hedge: Microsoft keeps its stake in OpenAI, and Claude remains integrated inside Copilot. But wherever possible, the MAI models step in, letting Microsoft capture greater cost savings without users noticing any drop in capability.

That’s not the only motive. This new setup also shields Microsoft from a looming deadline. OpenAI currently gives Microsoft discounted pricing, but that agreement expires in 2032 (a detail that matters more than it sounds: having a credible in-house alternative hands Microsoft real leverage in any future renegotiation, reducing how much it will ever have to pay OpenAI or Anthropic to keep Copilot running).

The risks for Microsoft

The MAI project is still in its infancy, and it isn’t without risk. The transition remains gradual, and the company has not published any timeline for further rollout. Most Copilot workloads, for now, still run on external models.

Doubts have already been raised about quality. MAI has been presented as a set of models comparable to the previous generation of OpenAI and Anthropic systems — not to today’s frontier LLMs. If the models underperform, the reputational cost could end up outweighing the direct savings Microsoft is chasing.

On the other side of the ledger, OpenAI and Anthropic have reason to be uneasy too. Both companies rely on Microsoft as their main partner for enterprise distribution, and that same partner has now also become a competitor. This isn’t an immediate financial threat, but no one — including Microsoft’s rivals — knows how far the MAI models will go, or how quickly.

Microsoft is also racing against the clock on infrastructure: Nvidia recently pushed back its next-generation AI “supersystem” to 2028, a delay that could either slow the whole industry’s roadmap or give in-house efforts like MAI more time to mature before facing the newest chips.

Editor’s note

This article was originally published in Italian on money.it by Pasquale Conte on July 13, 2026 as «Come funzionano MAI, i nuovi modelli AI di Microsoft che sostituiscono GPT». It has been translated and adapted for an international audience by the Money.it International desk.