Microsoft shifts Office AI prompts to in-house MAI models to cut costs

Microsoft is joining the AI cost-cutting trend by leaning on its own models. According to Bloomberg, the company has started diverting selected Excel, Word and Outlook AI features from OpenAI and Anthropic to its in-house MAI family, driven by token costs and data-residency considerations. OpenAI and Anthropic still handle the majority of Copilot traffic, so this is an incremental hedge rather than a divorce — but the direction is unmistakable.

The economics are the story. Microsoft says one MAI variant customized for McKinsey achieved a ten-fold cost-efficiency improvement over OpenAI's GPT-5.5, and cheaper inference is the lever that makes free or bundled AI features viable at Office's billion-seat scale. Routing high-volume, low-complexity prompts (summarize this email, fix this formula) to a smaller owned model, while reserving frontier models for hard tasks, is the emerging playbook across the industry.

The uncomfortable backdrop: Windows Latest reported Microsoft 365 Copilot adoption is still under 4.5% after three years, with only about 1% using it weekly — even as prices rose. That undercuts the narrative that enterprise AI is an obvious ROI win and raises the question of whether cheaper models help adoption or simply protect margins on a product few use.

This fits the week's dominant theme — cost pressure reshaping the market. US firms are turning to cheaper Chinese open-weight models, DeepSeek is building its own chip, and now the largest AI reseller is partially insourcing. Watch whether quality complaints emerge as owned models handle more user-facing prompts, and how OpenAI and Anthropic respond to their biggest distribution partner becoming a competitor at the margin.