The short version: Microsoft has begun routing some of the AI prompts inside Excel and Outlook Copilot away from OpenAI and Anthropic and toward its own in-house MAI models, according to a Bloomberg report from July 7. The switch is small so far, tens of thousands of weekly prompts against a Copilot total in the millions, but it targets the routine, high-volume work: drafting an email reply, summarizing a thread, generating a spreadsheet formula. Microsoft has not published which requests go where, so if you use Excel or Outlook with a Copilot subscription, some of your AI answers are already coming from a model you did not choose and were never told about. The interface has not changed. The reason behind it has: cost.
What did Microsoft actually change in Copilot?
Microsoft 365 Copilot has always leaned on outside AI labs, primarily OpenAI, with Anthropic’s Claude added for some tasks last year. Bloomberg’s reporting says Microsoft has now started completing a portion of Excel and Outlook prompts with its own MAI models instead. The company is not replacing OpenAI and Anthropic outright. It is running what amounts to a routing layer: first-party MAI models handle work that is good enough and cheaper to run in-house, while OpenAI and Anthropic models stay in place for tasks that still need them.
The tasks moving first are the commodity layer of office work, the kind most small businesses run through Copilot every day without thinking about it:
- Drafting a reply to a routine email
- Summarizing a long thread or meeting note
- Generating a straightforward spreadsheet formula
Microsoft has not disclosed a feature-by-feature list of what runs on which model, and a company spokesperson declined to comment on the record when Bloomberg asked. That means there is no setting to check and no notification to expect. The swap happens on Microsoft’s side, invisibly, one prompt at a time.
Why is Microsoft doing this now?
Cost. Microsoft AI chief Mustafa Suleyman said it plainly in a Bloomberg interview: “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.”
At Microsoft’s Build conference in June, the company unveiled seven in-house MAI models spanning reasoning, coding, image generation, and speech, including one built to match Claude Opus-tier coding performance at a lower price. Running Copilot at the scale Microsoft runs it, across hundreds of millions of Microsoft 365 seats, means every fraction of a cent per prompt paid to an outside AI lab adds up fast. Building a model that is merely good enough for routine tasks, and owning it outright, is cheaper than renting frontier intelligence for work that never needed it.
Does this affect the AI in my business’s Excel or Outlook right now?
Possibly, and you likely cannot tell either way. Independent testing cited in early coverage suggests MAI models are already matching or beating GPT-4o-class performance on narrow, domain-specific tasks like spreadsheet reasoning and email triage, the exact jobs Microsoft is routing to them first. For most everyday drafting and summarizing, the practical output should look the same.
There is one upside worth naming: MAI models run entirely on Microsoft’s own infrastructure, so a prompt handled by MAI never leaves Microsoft’s environment the way a prompt sent to a third-party lab does. If your business handles sensitive client data inside Excel or Outlook, that is a genuine, if quiet, improvement, not a downgrade to worry about.
What should small business owners actually do about this?
Nothing urgent, but it is worth filing under the same lesson the Cursor and SpaceX deal taught small businesses a few weeks ago: the AI vendor stack underneath the tools you already pay for is not fixed. It gets renegotiated, re-routed, and swapped, usually without asking you first.
Three practical takeaways for a business running Microsoft 365 with Copilot:
- Do not assume the model behind a feature today is the model behind it next quarter. Vendors optimize for cost as much as capability, and that tradeoff happens on their timeline, not yours.
- Judge Copilot by output, not by brand name. If a drafted email or a spreadsheet formula is right, it does not matter whether OpenAI, Anthropic, or Microsoft’s own MAI produced it. Spot-check important work the way you always should, regardless of which model wrote it.
- Watch for this pattern everywhere, not just at Microsoft. As the enterprise AI reckoning showed, every major AI vendor is under pressure to cut inference costs. Expect more quiet model swaps like this one across the tools your business relies on, from your CRM’s AI assistant to your website’s chat widget.
None of this requires action today. It is a reason to stay a little skeptical of any AI feature’s consistency and to keep spot-checking output, the same habit worth building regardless of which company’s model happens to be running underneath.
Frequently Asked Questions
Is Microsoft removing OpenAI and Anthropic from Copilot entirely?
No. Reporting describes a partial, gradual shift. OpenAI and Anthropic still handle the majority of Copilot’s AI traffic. Microsoft is routing a growing slice of routine, high-volume tasks to its own MAI models, not replacing the outside labs wholesale.
Will my Copilot subscription cost change because of this?
There is no indication of a price change tied to this shift. The cost savings are on Microsoft’s side of the ledger, from paying less for inference, not necessarily passed on to subscribers in the near term.
How do I know if a Copilot response came from MAI or from OpenAI/Anthropic?
You cannot, currently. Microsoft has not built a way to see which model handled a given prompt, and it declined to publish a list of which features have moved.
Should I be worried about response quality dropping?
Early benchmarks suggest MAI models are competitive with GPT-4o-class models on the specific tasks Microsoft is routing to them, like spreadsheet formulas and email drafting. If you notice a Copilot response that feels off, treat it the same way you would treat any AI output: check it before you rely on it.
If the AI running your everyday business tools can change without you knowing, what would make you actually notice?
