Should you let employees choose the model in Copilot Cowork? Yes, but teach them what each choice costs and put a limit in place first. Let people use it without guardrails and they will quickly push you over your spend limit.
That is the answer I give CIOs who ask me about Claude in Copilot Cowork. I see it as good news: your people get access to Anthropic's Claude models inside the Microsoft tool you already own, govern and pay for. But a model picker is also a new decision for every employee, and every decision has a cost. This post is my practical take for IT directors and CIOs on getting the benefit without the bill shock.
The short version
- Copilot Cowork now runs on Claude and GPT models, with Auto as the default. My rule for employees: use Auto, and switch to Opus for complex work.
- Model choice and reasoning effort change your Copilot Credits spend. A model guideline is a cost control, not just a usability tip.
- Employees feel more confident when they get a clear, simple guideline. Without one, they guess.
- Not every task is worth delegating to Cowork. Low value work like inbox clean up can burn credits for little return.
- If Claude is already inside Copilot, the case for buying Claude Enterprise or ChatGPT on top gets much weaker. Unless there is a clear use case, stick to one tool.
Does Copilot Cowork use Claude?
Yes. When Copilot Cowork reached general availability in June 2026, Microsoft said it "runs on Anthropic models, including Opus 4.8 and Sonnet 4.6". Back in March, Microsoft explained that it had brought "the technology that powers Claude Cowork into Microsoft 365 Copilot" by working closely with Anthropic.
Today Cowork is multi model. Microsoft's Cowork model guide lists Auto as the default, alongside OpenAI models such as GPT 6 Astra and Claude models such as Opus 5, Claude Sonnet 5 and Claude Fable 5.1. Each response shows a small model badge, so people can see which model did the work.
Two things confuse almost every group I work with. First, Copilot Cowork is not Claude Cowork. Claude Cowork is Anthropic's own product; Copilot Cowork is Microsoft's agent, working inside your Microsoft 365 permissions and admin controls. Second, Microsoft has renamed Microsoft 365 Copilot to Microsoft Copilot. Expect your employees to use both names for a while.
Is your data protected with Claude in Copilot Cowork?
For the Claude models most people will use, yes. Microsoft says they are covered by the same data protection commitments as the rest of Copilot, so your data is handled the same way whichever model does the work.
Cowork also keeps your existing access model. It inherits each user's permissions, can only reach files and email that person can already open, and shows sensitivity labels in the session. Switching the model does not change what Cowork can see.
There are two exceptions worth knowing:
| Situation | What it means for you |
|---|---|
| Most Claude models in Canada and the US | On by default, same protections as the rest of Copilot |
| A few advanced Claude models marked "with Data Retention" | Different terms apply and Anthropic keeps some data for a short period. These stay off until IT chooses to turn them on, and users see a banner when one is selected |
| Organizations in the EU or UK | Claude is off by default for data residency reasons. IT and legal decide whether to turn it on |
Source: Anthropic models in Microsoft Online Services, Microsoft Learn.
If you are in Canada or the US, Claude is probably already on in your tenant. My advice is to make that a decision you own, not a default you inherited.
Which model should employees pick in Copilot Cowork?
My rule is simple: use Auto, and switch to Opus for complex tasks. Microsoft's own guidance is also to leave the picker on Auto for most work.
I saw why this matters in a Prompt a Thon with a Canadian wealth management firm. Participants asked the question I now hear in almost every session: "What model should I use?" I gave them exactly that guideline, Auto for everyday work and Opus when the task is complex. They told us they felt more confident once they had a clear guideline to follow.
Here is the guideline I give employees:
| Option in the picker | When I tell people to use it |
|---|---|
| Auto (default) | Everyday work. Start here every time |
| Opus 5 | Complex, high stakes work: client deliverables, board papers, analysis across many documents |
| Claude Sonnet 5 | Quick drafts and lookups when speed matters more than depth |
| Any other model | Only when your team has tested it on a specific workflow, or IT has approved it |
The point is not to turn every employee into a model expert. Every extra decision at the moment of use lowers confidence. One default and one clear exception is something people remember after the workshop ends.
How model choice changes your Copilot Credits spend
Model choice is a cost decision. Cowork is billed on usage, and Microsoft says consumption depends on "model choice, context volume, orchestration, and tools used". Pay as you go is priced at $0.01 per Copilot Credit.
There is a second dial next to the model picker: reasoning effort, from Light to Max. Microsoft notes that higher effort "is slower and uses more credits," and the setting sticks until the user changes it.
I learned how quickly this adds up with a business unit that ran a pilot of 30 people using Cowork for a few months. They were champions who had come through our Prompt a Thon, so they were keen and capable. The guardrail we put in place was simple: do not use Cowork to clean up your inbox or for low value tasks, and check the cost after each task to decide whether the result was worth the spend. Microsoft lists inbox clean up as a Cowork scenario, and it works. My question is whether it is worth paying for when Copilot Chat or ten minutes of your own time would do.
To turn this into a policy, work through five steps:
- Confirm which model families are turned on in the Microsoft 365 admin center under Copilot settings.
- Set the default: Auto and Medium reasoning effort for everyone.
- Name the exceptions: which tasks justify Opus or higher effort.
- Set consumption limits per user or group before you widen access.
- Teach people to check the cost of each task, and review usage with your champions every month.
For the full credit math, forecasting and spend caps, see my guide to Copilot Cowork governance.
What IT should tell employees about which model runs their work
People will notice the model badge and the picker. If nobody explains them, they fill the gap with guesses: "Claude is smarter," "GPT is safer," or "Am I even allowed to use this?" A short, honest message prevents most of that.
I map the message to ADKAR so it does more than announce a feature:
- Awareness: Copilot Cowork can run on Claude or GPT models. Your data is protected the same way either way.
- Desire: Auto saves you a decision and keeps spend predictable for your team.
- Knowledge: Use Auto, switch to Opus for complex work, and know that every task has a cost.
- Ability: Practise on real tasks in a guided session, and check the cost of each one.
- Reinforcement: Managers and champions ask "was that task worth it?" in team check ins for the first month.
One admin detail matters before you send the message. Microsoft lets admins assign Claude access to specific users or groups, but the Cowork model page describes its Claude switch as tenant wide. Test what your tenant actually does before you promise a pilot group exclusive access.
Do you still need Claude Enterprise or ChatGPT if you have Copilot?
For most organizations already invested in Microsoft, no. Most organizations buy AI tools before they understand the investment they already made. I regularly meet clients with Copilot, Claude and ChatGPT all in the building, and the pattern is the same: spotty adoption across the board, with people simply using whichever tool they like most.
That overlap costs more than licences. It means more money, more confusion about who should use what, more security reviews for every tool, and no unified training plan, because nobody can train everyone on three assistants at once.
My recommendation is always the same. Get to know Copilot better and understand where the tools overlap, particularly if you are already invested in Microsoft. Copilot works across all the Microsoft 365 apps, and now that Claude runs inside it, much of the reason to buy Claude separately disappears. A standalone tool still makes sense when there is a clear use case Copilot does not cover, such as software development, specialist design work, or connections Copilot cannot reach. Make that a named exception with its own business case. For help deciding, see choosing the right AI for the work.
Unless there is a clear use case, stick to one tool. Claude inside Copilot means you can get the model without adding the vendor.
What IT leaders should do next
- Check whether Claude is on in your Microsoft 365 admin center, and decide whether that is your choice or just the default.
- Decide whether any Claude models marked "with Data Retention" are allowed, and write it down, even if the answer is no.
- If you operate in the EU or UK, confirm with legal and privacy that Claude is acceptable before turning it on.
- Publish a one line model rule: use Auto, switch to Opus for complex tasks.
- Set Copilot Credits limits per user or group before you widen Cowork access.
- Tell people which tasks are not worth delegating to Cowork, starting with inbox clean up.
- Brief managers and champions first, so they can answer "which model should I use?" without escalating to IT.
- Map your current AI tools and their overlap with Copilot before you renew or buy anything else.
How Change Champions helps
Since 2013, we have helped more than 500 organizations and 60,000 users adopt new technology, with a PROSCI certified approach focused on the people side. For multi model Copilot, that means turning admin settings into habits people actually follow, and spend you can predict.
- Microsoft Copilot adoption consulting: readiness, governance decisions and a rollout plan that includes a model guideline and Copilot Credits controls.
- Microsoft 365 Cowork Skill a Thon: a hands on workshop where teams learn what to delegate to Cowork, build reusable skills, and learn when Chat is the better (and cheaper) choice.
- Copilot Prompt a Thon: where most of the champions in the stories above started.
Ready to turn rollouts into real adoption?
Whether you're driving Microsoft 365 and AI adoption, reducing human risk, or leading a change program, a 30 minute discovery call is the fastest way to see if we're a fit.
