If you’ve been using Copilot in Teams, Word or Outlook for months, I get why the recent headlines about “OpenAI in Copilot” feel confusing. OpenAI’s models have powered parts of Copilot since it launched, so that bit isn’t news.
What’s actually changed is quieter than that, and it’s the sort of change that’s easy to miss until someone asks about it in a board meeting or an audit.
Here’s the plain English version of what’s going on, why it matters, and what I’d check in your tenant this week.
The bit that isn’t new
Since Copilot launched, Microsoft has run OpenAI’s models inside its own cloud, through something called Azure OpenAI. Think of it like Microsoft hosting the model on its own servers, under its own security certifications, with Microsoft in full control of the plumbing.
When you’ve asked Copilot to summarise a document or draft an email, that request has almost always been handled this way. GPT-5 arrived in Copilot last August, and GPT-5.6 became the preferred model in Word, Excel, PowerPoint, Chat and Cowork from early July this year. All of that ran through Microsoft’s own infrastructure.
So no, you’re not imagining it. You have been using OpenAI-powered Copilot for a good while now.
The bit that is new
What’s changed is who’s actually running the model behind the scenes for some Copilot tasks, and how that’s contracted.
From 9 July 2026, Microsoft started offering a second path: OpenAI-operated models, delivered directly by OpenAI rather than hosted inside Microsoft’s own cloud. Microsoft still oversees this through contract terms and its Data Protection Addendum, but the processing itself now happens on OpenAI’s side of the fence for some GPT experiences in Copilot, not Microsoft’s.
Then, on 24 July 2026, Microsoft switched this on by default for eligible commercial customers. If your organisation’s admin hadn’t already gone in and set it to “No users,” it’s very likely sitting on “on” right now, without anyone in your business having made an active decision about it.
That’s the real story here.
Not that OpenAI showed up in Copilot for the first time, but that Microsoft quietly added a second way for OpenAI to process your data, and made it the default.
Why this is worth ten minutes of your time
A few reasons this matters, even if nothing about your day to day Copilot experience looks different.
First, the two paths aren’t identical on paper. OpenAI-operated models sitting outside Microsoft’s own cloud currently miss a few compliance credentials that Azure OpenAI has, including a PCI DSS attestation, a HITRUST certification and a SOC 1 Type 2 report.
They’re also excluded from in-country processing commitments where that matters, and they’re not available in government or sovereign cloud environments. If your business handles card payment data, health records or anything with strict residency requirements, that’s a genuine difference, not a technicality.
Second, this happened by default. Most changes to how your business handles data go through some kind of approval process. This one didn’t ask first.
It’s a good reminder that Microsoft’s admin settings deserve a regular look, not just a one-off setup during onboarding, something I’ve said before when building an AI operating model for Microsoft 365.
Third, this is part of a bigger pattern. Copilot has quietly become a multi-vendor product. Anthropic’s Claude models sit alongside OpenAI’s inside Copilot Chat and Copilot Studio too.
Every one of these additions is a new name that could end up in your data protection paperwork, your vendor register or a client’s due diligence questionnaire, whether you’ve thought about it yet or not.
What to actually check in your tenant
Here’s where I’d start, in order.
Find the setting. Go to the Microsoft 365 admin centre, then Copilot, then Settings, then View all.
Look for “AI providers operating as Microsoft subprocessors.” Click into OpenAI and you’ll see who currently has access: all users, specific users and groups, or nobody.
You’ll need AI Administrator or Global Administrator rights to see and change this, so if you’re not sure who holds that access in your business, that’s worth sorting out too.
Decide deliberately, rather than by default. There’s nothing wrong with leaving OpenAI-operated models enabled. For most businesses, the difference won’t matter day to day.
But “we left it on because we didn’t know it was there” is a different answer to “we reviewed it and decided it suited us,” and only one of those holds up well if someone asks.
Scope it properly if you need to. You don’t have to choose between “everyone” and “no one.” Microsoft lets you assign this by Entra ID security group, so you could keep OpenAI-operated models available for your general staff while excluding a finance team handling card data, or a health team working with clinical records.
Update your paperwork. If your business runs a vendor register, an AI inventory or privacy impact assessments as part of how it manages AI use, OpenAI’s subprocessor status belongs in there now, alongside Anthropic and Microsoft’s own models.
I’ve covered the basics of getting this kind of AI governance right if you’re starting from scratch, and it’s a lighter lift than most people expect once the first version exists.
The bigger lesson here
None of this means Copilot’s suddenly less safe, or that you need to slow down your AI plans. Microsoft’s own contract terms and data protection commitments still apply, and for the vast majority of everyday tasks, most businesses won’t notice a practical difference between the two paths.
What it does mean is that “we use Microsoft 365 Copilot” is no longer a complete answer if someone asks who processes your data through AI. The honest answer now includes Microsoft, OpenAI, and depending on your settings, Anthropic too.
That’s not a bad thing. It’s actually a sign that Copilot’s maturing into a proper multi-vendor platform rather than a single-model tool bolted onto Office.
But it does mean the governance conversation needs to keep pace, the same one I raised when multiple AI tools started showing up across one business at once.
If you’re the person who signs off on Copilot licensing in your business, this is a good moment to pair that decision with a quick look at what’s actually running underneath it, something I touched on when breaking down what Copilot licensing really covers.
Where to from here
My honest advice: don’t panic, but don’t skip it either.
Set aside ten minutes this week, find that toggle, and make a proper decision about it, rather than letting Microsoft’s default make it for you. If you’ve got regulated data anywhere in the mix, spend a bit longer and loop in whoever owns your compliance obligations before you decide.
If you’d like a second set of eyes on your Copilot admin settings, your AI vendor register, or how ready your business is for this kind of change, that’s exactly the sort of thing we help clients work through.
Have a chat with our team about a quick Microsoft 365 Copilot review, and we’ll help you turn “we think it’s fine” into “we checked, and here’s why.”
About the Author
Carlos Garcia is the Founder and Managing Director of CG TECH, where he leads enterprise digital transformation projects across Australia.
With deep experience in business process automation, Microsoft 365, and AI-powered workplace solutions, Carlos has helped businesses in government, healthcare, and enterprise sectors streamline workflows and improve efficiency.
He holds Microsoft certifications in Power Platform and Azure and regularly shares practical guidance on Copilot readiness, data strategy, and AI adoption.
Sources
- OpenAI as a subprocessor in Microsoft Online Services — Microsoft Learn, last updated 5 August 2026
- GPT-5.6 is now the preferred model in Microsoft 365 Copilot — OpenAI, 9 July 2026
- Available today: GPT-5 in Microsoft 365 Copilot — Microsoft 365 Blog, 6 August 2025
- Expanding model choice in Microsoft 365 Copilot — Microsoft 365 Blog, 24 September 2025
- Microsoft Cloud and AI strength fuels fourth quarter results — Microsoft Source, 29 July 2026
- Microsoft 365 Enterprise Update August 2026 — Empowering.Cloud, 5 August 2026