Microsoft is increasingly directing AI requests within its Microsoft 365 suite from third-party platforms such as OpenAI’s ChatGPT and Anthropic’s Claude to its proprietary AI models. This strategic shift aims to lower the mounting costs linked to external AI usage while maintaining AI integrations across Office applications like Excel and Outlook.

The company has started funneling a small yet growing portion of AI-generated responses through its internal offerings, known collectively as Microsoft AI (MAI) models. These models are already integrated into several Microsoft products, including Copilot for Business and Enterprise, and feature specialized tools like MAI-Code-1-Flash, a coding assistant positioned against third-party tools like Claude Code and OpenAI’s Codex.

This move aligns with Microsoft’s broader plan to reduce its financial reliance on external AI providers. Mustafa Suleyman, Microsoft’s CEO of AI, outlined at the recent Build conference that the goal is to “reduce and ultimately eliminate” payments to Anthropic, marking a clear intention to reclaim control over AI-related expenditures.

Microsoft’s adoption of in-house AI solutions extends beyond traditional Office apps. The company has announced intentions to deploy its own transcription model for Teams, underscoring a commitment to internalize AI capabilities where possible. The ongoing transition comes amid growing concerns in the tech industry over spiraling costs tied to AI token usage, which represents the computational expense associated with generating AI responses.

This recalibration follows earlier efforts by Microsoft to limit employee access to cost-intensive external AI tools and move GitHub Copilot users to token-based billing to better manage expenses. Such cost-control measures come against a backdrop where other major companies have exhausted their AI budgets prematurely due to unrestricted usage, highlighting the financial risks of external AI dependency.

By steering Microsoft 365’s AI operations to its own models, the company hopes to better predict and manage costs. While it remains difficult to precisely measure the return on investment for these AI integrations, controlling token spending represents a pragmatic step in an industry rapidly reassessing its AI strategies amid soaring consumption and expenses.