.jpg?w=1024)
Microsoft is giving some Office AI tasks to its own models.
The company is now using internally built MAI models to complete some prompts in Excel and Outlook, according to Bloomberg, reducing its reliance on outside systems from OpenAI and Anthropic. The shift is still small compared with Microsoft’s broader AI workloads, but it points to a larger priority: making enterprise AI cheaper to run at scale.
For Microsoft customers, the change may be nearly invisible. For the AI industry, it shows how even the biggest model buyers are trying to own more of the stack.
Cutting AI costs without abandoning partners
The move comes as the cost of running large AI systems continues to climb across the technology industry.
Microsoft has invested heavily in OpenAI and continues to benefit from discounted access to its models through the companies' long-running partnership. However, that pricing arrangement is not expected to last indefinitely, creating an incentive for Microsoft to reduce its reliance on outside AI providers.
At the company's Build developer conference in June, Microsoft introduced seven new MAI models covering areas such as coding, image generation, reasoning, and transcription. During the event, Microsoft's AI chief, Mustafa Suleyman, said, "We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost," according to Bloomberg.
Microsoft has already expanded MAI beyond Office applications. The company's models are available in GitHub Copilot, while a Microsoft-built transcription model is expected to arrive in Teams and other products in the coming months.
Part of a wider industry trend
Microsoft is not abandoning OpenAI or Anthropic. Instead, the company appears to be adopting a mixed-model approach, using its own models where they offer a better balance of cost and performance while continuing to rely on external models for other workloads.
The move also reflects a trend across the AI industry as companies search for ways to reduce soaring inference costs. TechCrunch noted that several major technology firms, including Amazon, Meta, Uber, and Accenture, have also taken steps to rein in AI spending as operating large AI systems becomes increasingly expensive.
At the same time, Microsoft continues expanding its enterprise AI ambitions. Last week, the company unveiled Microsoft Frontier Company, backed by a $2.5 billion investment to help businesses deploy AI systems and measure real business outcomes rather than simply experiment with the technology.
What changes for Microsoft customers
For Microsoft customers, the change may happen almost entirely behind the scenes.
Excel, Outlook, and other Microsoft 365 applications will continue offering AI features even if the underlying model changes. Owning more of its AI stack could give the company greater flexibility over pricing, performance, and product development while reducing exposure to rising model costs.
For OpenAI and Anthropic, the shift highlights a new phase of competition in which major customers are increasingly developing their own AI models alongside external partnerships rather than relying on a single provider.
AI costs are shaping the next phase
Microsoft's latest move suggests the AI race is becoming as much about economics as raw model performance.
Building cutting-edge AI remains expensive, but so does paying for third-party models at massive scale. By routing more everyday Office tasks to its own MAI models while keeping external models available where needed, Microsoft is pursuing a strategy that balances cost, control, and performance.
Whether that approach expands further will likely depend on how well Microsoft's in-house models perform compared with the industry's leading AI systems, particularly as enterprise customers continue demanding strong accuracy, reliability, and measurable returns on their AI investments.
Check out these free AI training courses from Microsoft, Google, AWS, IBM, DeepLearning.AI, and Kaggle, which can help workers build practical AI skills without paying for expensive programs.