Microsoft routes Excel and Outlook AI prompts to in-house MAI

Microsoft has quietly begun routing chunks of the Copilot workload inside Excel and Outlook to its own MAI models rather than to OpenAI or Anthropic, Bloomberg reported, and the in-house models are already handling tens of thousands of AI prompts a week across the two apps. The stock ticked up 1.75% on the day, which tells you the market read this first as a margin story rather than a product story.

The reason it matters is that Copilot inside the productivity suite is Microsoft's single highest-volume AI surface, and until now most of that traffic paid a bill to a third party. Moving even a slice of it to models Microsoft owns end to end changes the unit economics of running Copilot at scale, and it does so on the exact apps where the company has spent the last two years trying to make AI cheap enough to bundle. The framing in the reporting is explicit that slashing infrastructure costs is the driver.

It also gives the OpenAI relationship its first visible off-ramp inside shipping productivity apps. Excel and Outlook were previously described as relying more heavily on OpenAI and Anthropic; today's account is the first time the in-house alternative shows up as a live production replacement rather than a research bet or a supplementary option.

The honest caveat is that the reporting is narrow. It does not tell you which Excel and Outlook tasks got switched, how MAI's answer quality compares to the models it displaced on the same prompts, or whether enterprise Copilot customers are being told which model is now handling their traffic. Take the tens-of-thousands-per-week number as a scale hint rather than a settled share of the workload, because it is still small next to the full Copilot footprint.

What is worth watching from here is how far Microsoft pushes the routing. If MAI holds up on the easy, high-volume prompts, the natural next step is to widen the switch until only the genuinely hard requests fall through to a frontier model. That is the version of the story that would matter most to OpenAI's revenue line and to anyone selling API capacity into other hyperscalers.