Same model, a fifth product: Claude inside Microsoft

In the four-jobs lesson I said the same Claude model answers differently in chat, Cowork, Claude Code and Slack, because each app changes three things before you type a word. There’s a fifth place your team can reach Claude, and most businesses never chose it: Microsoft 365 Copilot.

If your business holds Copilot licences, Claude has probably been running inside Copilot since January. It comes with the Copilot licence: nobody had to buy a Claude seat or switch anything on. Anthropic became a Microsoft subprocessor on 7 January 2026, and since then Anthropic models have been on by default for commercial tenants everywhere except the EU, EFTA and the UK. Australian tenants included.

If you also pay for Claude seats, your people now reach the same models two ways. They won’t get the same answers.

Same model, someone else’s product

Inside Copilot, Microsoft sets all three of the things the app changes.

• What it’s told. Microsoft writes the standing instructions, not Anthropic. Expect different defaults for length, format and caution.
• What it can see. Copilot adds emails, files, chats and meetings from your tenant to the prompt before Claude reads it, limited to what that person can already open. Your Claude seats see what you give them, plus the systems you’ve connected.
• What it can do, and where the answer lands. Copilot puts the result inside the Word document or the Teams chat. Claude’s own apps hand it back in Claude, or change the files on your computer.

Then Microsoft adds a fourth: which model. The Claude versions differ from one Copilot surface to the next and often trail Anthropic’s own releases, and some surfaces have an Auto setting that may pick a model that isn’t Claude at all.

So when someone on your team says Claude in Copilot is worse, or better, they’re usually comparing two products, not two models. Then they stop using one of them. That’s how a rollout stalls, one prompt at a time.

Six places it turns up

Each has its own settings: Copilot Chat (in the model picker), Researcher, the editing agents in Word, Excel and PowerPoint, Copilot Cowork (off by default, and each task billed in Copilot Credits on top of the licence), Copilot Studio, and GitHub Copilot.

It’s not only the answers that differ

Inside Copilot, Anthropic works under Microsoft’s data protection terms. But Microsoft’s own documentation excludes Anthropic models from the EU Data Boundary and from its in-country processing commitments.

If your board approved Copilot because Microsoft keeps the data in Australia, that approval may not cover the Claude inside it. That’s a question for your admin and your policy, not for the people using it.

One thing to try this week

Take a real question from your own work, one that leans on a document you’d normally open. Ask it in Copilot Chat with Claude picked, then in Claude’s own chat with the same document attached. Put the two answers side by side and note what changed: the length, the format, what each one drew on and where the answer ended up.

Then give Claude in Copilot its own line on your “what to do where” page, separate from your Claude seats: what it works from, what it’s for, who it’s for. If you haven’t written that page yet, the four-jobs lesson shows ours, and it belongs beside your one-page rules.

Before anyone relies on it, ask your Microsoft 365 admin three questions:

1. Is Claude switched on in our Copilot? It’s the Anthropic setting in the Microsoft 365 admin centre.
2. Is Copilot Cowork switched on, and who can use it?
3. Does our AI policy say anything about Claude inside Copilot?