The model running your Copilot demo might not be OpenAI

Anthropic filed for an IPO earlier this month at a valuation nudging a trillion dollars. That's a significant number. What's more significant for anyone in the Microsoft channel is that since May, Anthropic's Claude has been the default model inside Microsoft 365 Copilot for Excel and PowerPoint. Word follows this summer. Claude is also selectable in Copilot Chat and now embedded in Copilot Studio's multi-model lineup.

So the product you've been selling as "Microsoft AI" now includes a model from a company that isn't Microsoft, which is itself in something of a rivalry with OpenAI, the company Microsoft has invested $13 billion in.

If that makes your head hurt slightly, that's probably the right reaction.

What's actually changed

For the past two years, Copilot has been positioned in the channel as a Microsoft product with an OpenAI engine. The Microsoft-OpenAI relationship was marketed so heavily that most partners couldn't tell you which model was running under the hood, and didn't need to. You sold the M365 licence. The AI was Microsoft's problem.

That framing is getting harder to sustain. Microsoft 365 Copilot is increasingly a platform that routes requests to different models depending on the task, the workload, and customer preference. OpenAI, Anthropic, and Microsoft's own models are all part of the stack now. Which one runs when someone clicks "summarise this email" isn't necessarily visible to the user and isn't necessarily the same answer it was six months ago.

The pitch has changed

The old version was essentially: you're getting the best AI, built into the tools your people already use, with the security and compliance you'd expect from Microsoft. That pitch still works. But the "best AI" part is now a moving target. Microsoft will route to whichever model performs best for the task, or whichever commercial arrangement is currently in play. Those aren't always the same thing.

For customers who don't care about the internals, this is fine. For customers who do, particularly those in regulated industries with data residency requirements or a board that has an opinion about which AI vendors they're comfortable with, the conversation is more complicated.

Twenty-plus years in the channel teaches you that customers rarely ask the right questions until something goes wrong. The ones who are going to ask "which AI is actually processing our data?" will ask it at the worst possible moment: mid-implementation, after the purchase order, when legal decides to get involved.

Getting ahead of that question now is worth ten minutes of pre-sales time.

What to do with it

Microsoft's documentation on Copilot model selection is improving, and Copilot Studio gives organisations real controls over which models are used in which agents. That's actually useful, and worth knowing how to configure before a customer asks.

The other shift worth making is to the pitch itself. You're not selling a Microsoft AI product. You're selling a platform that gives customers access to best-in-class AI capability, within the Microsoft trust boundary, with the controls to configure which models are used where. That's a stronger story, not a weaker one, if you can tell it with confidence.

We have been selling wrappers around technology in the channel for thirty years and made a pretty decent living while doing it. The wrapper has just got more interesting. And slightly harder to explain at a board meeting. Other than that, it's business as usual. As you were.