Claude Meets Copilot Studio

This is the one I've been building toward.

Over the last few posts we deployed Claude in Microsoft Foundry, said hello to it from code, and taught it to turn a messy support ticket into clean, structured data. All useful, all a bit back-end. Today it comes together into something you can actually hand to a colleague: a working agent, built in Copilot Studio, the low-code tool you already know, powered by Claude.

This is the moment the two halves of a Power Platform person's brain finally meet. The business side that knows what a good customer answer looks like, and the AI side that can produce one at scale. And the nicest part is how little new there is to learn, because most of it happens in a tool you've probably already got open in another tab.

The pleasant surprise

I spent the first few posts carefully wiring Claude up by hand in Foundry, so I braced myself for a fiddly connection step here. There isn't one.

Claude is already in Copilot Studio. Microsoft opened the tool up to more than one model, and Anthropic's models sit right there in the picker alongside the OpenAI ones. You choose Claude from a dropdown, the same way you'd choose a connector or a data source. No endpoint to paste, no key to manage, no code.

If you've spent years in the Power Platform, sit with that for a second. A frontier model, available as a dropdown, next to the tools you already use every day. That is the whole promise of this series in one small menu.

One thing to check before you start

There's a catch, and it's worth knowing up front so you don't waste an hour wondering why Claude isn't in your dropdown.

An admin has to turn Anthropic's models on first, in the Microsoft 365 Admin Center. Once that switch is flipped, the models are on by default in the Power Platform Admin Center, where an admin can decide which environments makers can use them in. So this is a tenant-level decision, not just a personal one.

The sneaky part: if Anthropic models are switched off, an agent you built on Claude doesn't error. It quietly falls back to a different model. So if your agent suddenly sounds different one morning, this is the first place to look. If you run your own tenant, flip the switch and move on. If you don't, this is the one short conversation to have with whoever does before you build.

Build the agent, and point it at Claude

Now the fun part, and it's genuinely about ten minutes.

1. Go to copilotstudio.microsoft.com and select Agents > New agent. You can let Copilot interview you, but I like Skip to configure so I can fill in the fields myself.
2. Give it a name and a short description. Mine is Harbor Help, "the first line of support for Harbor Supply Co. customers."
3. In the Instructions, tell it who it is and how to behave. This is the system prompt, and it's where your business judgement lives. Something like: You are Harbor Supply Co.'s support agent. Be warm and brief. Confirm order numbers back to the customer. If a request involves damaged goods or a missed deadline, treat it as urgent and offer to escalate to a human.
4. Open the agent's settings and set its primary model to Claude. That's the whole integration. From now on, the reasoning behind every reply is Claude's.

That last step is the one that would have been a week of work a couple of years ago. Today it's a dropdown. If you'd rather mix models, you can also pick a model per prompt, so a cheap task and a hard one can use different models in the same agent, but for today, one good model everywhere is plenty.

Give it a bit of Harbor's world

Right now the agent is clever but empty. Ask it about a specific order and it will cheerfully make something up, because it doesn't know anything about Harbor beyond what you put in the instructions.

For today, a little goes a long way. Add one knowledge source, a short document with Harbor's returns and replacement policy, so the agent has real facts to lean on. In Copilot Studio that's the Knowledge section: point it at a file or a site and it will use it when answering. Even a one-page policy is enough to stop the worst of the guessing.

I'm keeping this deliberately light, because grounding is a proper topic in its own right, and it's next time's whole post. Today the goal is just to get the agent answering in Harbor's voice, with enough real context that it isn't inventing the company from scratch.

Watch it answer

Here's the difference the whole exercise buys you. Same question, Dawn from the last post, asking about her bent racking.

A raw, empty chatbot gives you the beige non-answer we've all seen:

Dawn: 2 of the uprights on order HS-44821 came bent, I need them replaced before the weekend.

Generic bot: I'm sorry to hear about your experience. Please contact our support team and we will be happy to assist you.

That helps no one. Now the Harbor Help agent, same question, with its instructions and that one policy document behind it:

Dawn: 2 of the uprights on order HS-44821 came bent, I need them replaced before the weekend.

Harbor Help: Thanks Dawn, I've got order HS-44821. Damaged uprights are covered by our replacement policy, and because this is blocking a weekend move I'm flagging it as urgent. I'm connecting you to a person now who can get replacements out to you. Is the delivery address on the order still correct?

Same model, same ten minutes of setup. The second answer knows the order, applies the real policy, spots the urgency, and does the one thing the first one refused to: it actually moves things forward.

So why did we bother with Foundry?

Fair question. If Claude is just a dropdown in Copilot Studio, why did we spend three posts deploying it in Foundry and calling it from code?

Because they're two doors to the same room, for two different jobs.

The Copilot Studio picker is the fast door. Low-code, no plumbing, perfect for an agent a business team will own and tweak. You trade some control for speed, and most of the time that's exactly the right trade.

The Foundry route from the earlier posts is the deep door. You call Claude from your own code, shape the request precisely, enforce a strict schema like we did with the tickets, and wire it into systems Copilot Studio never touches. More control, more responsibility.

The quiet superpower of this series is that you don't have to choose a camp. A Copilot Studio agent can call a custom prompt or an action that runs your Foundry code underneath, so the low-code front and the pro-code engine live in the same solution. You reach for whichever door the job in front of you needs.

The two halves, finally in one place

Step back and look at what you've got. An agent that reasons with a frontier model, answers in your company's voice, knows when to pull in a human, and lives in the same low-code tool you've been using for years. The business-process brain and the AI brain, in one place, on your turf.

That's the whole thesis of this series, and you just built it in an afternoon.

The agent still has one honest weakness: it only knows the little we told it. Ask something outside that one policy document and it's back to guessing. So next time we fix that properly. We ground the agent in Harbor's real world, its policies, its catalog, its orders, so its answers come from facts instead of vibes. That's the difference between a demo and something you'd actually let talk to a customer.

If you build your own version, I'd genuinely love to see it. Show me what your agent said back.

See you next time. 😊

Sources

• Anthropic joins the multi-model lineup in Microsoft Copilot Studio — Microsoft
• Bring your own model for your prompts — Microsoft Learn