- What Changed: Microsoft is reportedly routing selected Excel and Outlook AI prompts to its own Microsoft AI (MAI) models instead of relying only on OpenAI and Anthropic systems.
- Why It Matters: The move targets inference costs, the recurring compute expense created every time Copilot answers a prompt inside high-volume Microsoft 365 apps.
- Scope: The shift appears limited. Microsoft has not confirmed the Excel and Outlook change, and MAI still represents only a small share of Microsoft’s overall AI usage.
- User Test: The strategy only works if cheaper in-house models keep routine Copilot responses fast, accurate, and reliable.
Microsoft is reportedly beginning to route selected AI prompts in Excel and Outlook to its own Microsoft AI (MAI) models, a limited but significant cost-cutting move inside Microsoft 365. The change moves some routine Office tasks away from outside models supplied by OpenAI and Anthropic, according to a Bloomberg report.
The reported shift does not mean Microsoft is abandoning OpenAI or Anthropic across Copilot. Bloomberg’s account describes a narrower change: MAI is handling tens of thousands of prompts per week in Excel and Outlook, while Microsoft’s in-house models still account for only a small share of the company’s overall AI usage. Microsoft declined to comment on the specific Excel and Outlook routing change.
The strategic question is whether Microsoft can move predictable productivity tasks to cheaper first-party models without weakening the answers users receive from Copilot. For everyday Microsoft 365 customers, the model behind a response matters less than whether a spreadsheet explanation, email draft, meeting summary, or transcription remains accurate and fast.
Why Microsoft Is Moving Routine Office Prompts In-House
Every AI answer has a compute cost. That cost, known as inference, is paid each time a model processes a user prompt and generates a response. In a high-volume productivity suite such as Microsoft 365, even small per-prompt costs can become material when repeated across email, documents, spreadsheets, meetings, and enterprise workflows.
That makes Excel and Outlook plausible early targets for model routing. Many spreadsheet and email requests are narrower than open-ended chatbot conversations: they are tied to a file, a mailbox, a thread, a table, or a specific business context. Those constraints can make them better candidates for smaller, tuned models than for the most expensive general-purpose frontier systems.
The move also fits a broader industry push to reduce AI inference spending as usage grows. Microsoft can lower its own costs if internal MAI models handle routine prompts at acceptable quality, latency, and reliability.
Microsoft AI CEO Mustafa Suleyman has also framed third-party model costs as a pressure point. In a recent Bloomberg interview, Suleyman said:
“Anthropic is extremely expensive and I think many people are urgently looking for alternatives,”
“We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost.”
Mustafa Suleyman, Microsoft AI CEO, via Bloomberg
Those comments explain the cost logic, but they should not be read as proof that Microsoft is removing Anthropic or OpenAI from all Copilot experiences. Microsoft’s broader Copilot strategy still includes outside frontier models where they provide better quality, broader reasoning, or customer choice. Microsoft previously added Anthropic’s Claude as an OpenAI alternative in parts of Microsoft 365 Copilot, underscoring that the company is pursuing model choice rather than a single-model architecture.
What Changes for Users — and What Does Not
For users, the practical impact depends on quality. A lower inference bill helps Microsoft, but it only helps customers if common Copilot responses remain useful. A cheaper model that drafts weaker emails, misreads spreadsheet context, or produces less reliable summaries would turn a back-end cost saving into a front-end product problem.
The current reporting also leaves important details unresolved. Microsoft has not publicly confirmed which Excel and Outlook prompt types are being routed to MAI, whether enterprise administrators can see which model handled a request, or whether customers will receive more direct controls over model selection. It is also unclear how MAI performance compares with OpenAI and Anthropic models on real-world Office workloads outside Microsoft’s own testing.
That is why the shift is best understood as a workload-routing test, not a full platform reset. Microsoft appears to be asking a narrower question first: which repetitive Office tasks can be handled by smaller in-house models without reducing the quality users expect from Copilot?
Where MAI Fits in Microsoft’s Broader AI Stack
The Excel and Outlook report comes as Microsoft has been expanding the role of MAI across its products and developer platforms. Earlier Microsoft 365 work broadened Copilot access across Office apps, giving Microsoft more places to route app-specific AI tasks. A prior three-model MAI rollout preceded the June 2 Foundry rollout, which expanded the company’s first-party model lineup for developers.
Microsoft’s official announcement of seven in-house MAI models describes systems for image, voice, transcription, coding, and reasoning tasks. That range matters because model routing works best when a company can match a specific task to a specialized model instead of sending every request to the same large general-purpose system.
MAI is also moving into visible Microsoft products. MAI-Image-2.5 is already live in PowerPoint and rolling out to OneDrive, while MAI-Transcribe-1.5 is being integrated into Copilot, Teams, GitHub, and Dynamics 365 Contact Centre. Microsoft’s Foundry model listing also places MAI models across other text, image, voice, and speech use cases.
The company’s earlier MAI image-generation work and broader AI self-sufficiency push point in the same direction. Microsoft wants more control over the models behind its products, especially where it can tune a model to a specific task and avoid paying another provider for every request.
Microsoft has also said that self-sufficiency depends on training reasoning models from scratch and avoiding distillation from other labs. That matters strategically, but it should be separated from the narrower Excel and Outlook report: the existence of more MAI models makes first-party routing more plausible, but it does not by itself prove that MAI can replace frontier models across all Copilot tasks.
The Key Test: Cost Savings Without Quality Loss
The strongest Office-specific claim comes from Microsoft itself. In its Build materials, the company said an Excel-tuned MAI model could deliver output comparable to GPT-5.4 while being up to 10 times more efficient. That should be treated as Microsoft’s own benchmark or projection, not independent proof that every customer workflow will perform the same way.
If Microsoft can achieve that kind of efficiency on routine Office tasks, the business logic is clear. Smaller in-house models could handle bounded prompts in Excel, Outlook, Teams, or other Microsoft 365 apps, while OpenAI and Anthropic models remain available for broader reasoning, harder tasks, or cases where model choice matters to enterprise customers.
But the product risk is just as clear. Users do not experience inference costs directly; they experience answers. If MAI produces slower, weaker, or less reliable Copilot responses, the cost-saving strategy will be visible in the worst possible way. If users cannot tell the difference, Microsoft gains a stronger case for routing more Microsoft 365 workloads to its own models.
What to Watch Next
Microsoft Teams is the next obvious test case. MAI-Transcribe-1.5 supports 43 languages and includes biasing features for specialized transcription workflows. Meeting transcription is repetitive, high-volume, and tightly scoped, making it another natural candidate for first-party model routing if quality holds up.
The most important signals now are not only whether Microsoft expands MAI beyond Excel and Outlook, but how transparent that expansion becomes. Enterprise customers will want to know which model handled a request, whether administrators can control routing, how Microsoft measures quality against outside models, and whether lower inference costs eventually affect Copilot pricing or mainly improve Microsoft’s margins.
