That is becoming more obvious.

Microsoft has started replacing some OpenAI and Anthropic model usage in Office apps with its own MAI models, including selected workloads in Excel and Outlook.

This is a major signal.

For the last two years, many customers buying Microsoft 365 Copilot probably assumed the value was:

Microsoft 365 Copilot = Microsoft Office + ChatGPT-style intelligence.

That was never technically precise.

Copilot is an orchestration layer.

It takes the user prompt, grounds it in Microsoft 365 data, applies security and compliance controls, sends the task to an LLM, then returns the answer or action inside the Office app.

The LLM underneath can change.

That is the point.

The future of Copilot is not “which single model is best?”

It is:

  • Which model is good enough for this task?
  • Which model is fastest?
  • Which model is cheapest to run?
  • Which model has the right data-processing terms?
  • Which model should be reserved for genuinely hard reasoning?

A simple Outlook draft reply does not need the same model as a complex multi-document legal, financial or commercial analysis.

A simple Excel formula explanation does not need the same model as deep workbook reasoning.

This is where MAI matters.

If Microsoft can use its own models for routine Office tasks, Microsoft gets:

  • lower inference cost,
  • lower dependency on third-party model providers,
  • more control over latency,
  • more control over the product roadmap,
  • and potentially cleaner first-party data processing.

For customers, the question changes.

It is no longer enough to ask:

“Do we have Microsoft 365 Copilot?”

The better questions are:

  • What models are being used?
  • Can we control third-party AI providers?
  • When does Copilot use OpenAI, Anthropic, or Microsoft-hosted models?
  • Do different models have different data-processing or retention terms?
  • Will quality change when routine Office workloads move from GPT or Claude to MAI?

This is especially important because Microsoft’s model picker is already becoming more abstract.

Users may see terms like Auto, Quick response, Think deeper, or Claude Opus.

They may not see the exact model version.

They may not know whether the answer came from OpenAI, Anthropic, or MAI.

That is not accidental.

Copilot is becoming a model router, not a single-model product.

This is also where Microsoft’s economics become visible.

At enterprise scale, AI cost is not just the licence price.

It is also inference.

  • Tokens.
  • GPU time.
  • Memory.
  • Storage.
  • Safety systems.
  • Routing.

If Microsoft can move millions of routine prompts away from expensive frontier models without users noticing a quality drop, the margin profile of Copilot changes.

That does not mean MAI is automatically worse.

It also does not mean OpenAI or Anthropic are disappearing from Copilot.

It means the expensive frontier model will increasingly be reserved for work that actually needs it.

The commercial issue for customers is simple:

You are not just buying “Copilot”.

You are buying a changing mix of models, routing logic, data terms, quality levels and cost economics, hidden behind one product name.

That is a very different procurement and governance conversation.