From tool to teammate: what Copilot Cowork changes about how we work

One of us was sitting in Amsterdam, the other in Miami, and we landed in the same place. People keep asking whether Copilot Cowork is better than Copilot. We think that is the wrong question.

Copilot helps you think, create, analyse and get things done. Cowork can take a goal and go do the work on your behalf, across multiple steps and applications. Microsoft describes it as an agentic system that can send emails, schedule meetings, create documents, post in Teams and manage files while keeping the user in control. (Microsoft Learn)

What changed is not simply the quality of the technology. It is the relationship we have with it. That sounds like a small difference until you notice how much of your own behaviour has to change before it pays off.

Between us, we use Cowork five or six times a day, plus a couple of tasks that run on a schedule without us touching them. Below is what a few months of real use taught us, including some of the lessons we paid to learn.

The way you learned to prompt is not the way you delegate

Prompting and delegating are not the same skill. Most of us learned generative AI in small steps. Make me a list. Now sort it. Now filter it. Now put it in Excel.

There is nothing inherently wrong with that. In a conversation with Copilot, you are thinking through the problem together and shaping the answer as you go.

Cowork changes the dynamic because you are no longer handing over only the next step. You can hand over an outcome. That means the better instruction is often not a sequence of tiny prompts. It is a clear description of what you need, the context it should use, the boundaries it should respect, and what “done” looks like.

Keep managing Cowork one tiny step at a time and you lose much of the advantage. You remain responsible for orchestrating the process while Cowork continues consuming credits as it works.

So the shift is not simply “learn a new tool.” It is define success before you delegate. Anyone who has onboarded a junior colleague already knows something about doing this. We just never had to think about it quite this way with software.

Cost is a management skill now, not a finance detail

Nobody was really trained for this part, so let’s be concrete.

Cowork uses Copilot Credits. Microsoft’s current pay-as-you-go price is $0.01 per Copilot Credit, approximately €0.01, although euro figures in this article are approximate conversions because Microsoft publishes the credit pricing in US dollars. (Microsoft CDN)

A very small task might cost only a few credits. A complex task that searches large amounts of information, reasons across sources, uses tools and produces multiple outputs can consume hundreds or considerably more.

That variability is intentional. Microsoft says Cowork usage depends on four things: models, context, tools and runtime. Its current planning guidance puts light tasks at roughly 100 to 300 credits, medium tasks at roughly 300 to 700, and heavy tasks above 700, while making clear that actual usage varies by workflow. (Microsoft CDN)

That means asking Cowork a simple question might cost around $0.01 / €0.01. A much heavier job can cost $25 / roughly €22, or more.

Now scale that to a role and it gets interesting. At Wortell, there is a project manager working thirty-six hours a week across twenty-four projects. Twenty-four.

With adoption consultants, architects, customers and internal stakeholders across those projects, Cowork takes on a meaningful share of the coordination. The usage costs roughly €2,000 / $2,315 a month.

The useful question is not simply whether €2,000 or $2,315 a month is expensive. The question is what that spend is buying.

If Cowork is saving someone a few minutes here and there, that is a personal productivity calculation. If it is absorbing a meaningful share of coordination work that would otherwise require additional capacity, that is a very different business case. Those are not the same value proposition, even if both have a cost.

A few habits help us understand the difference.

Run /cost after a task. Cowork reports consumption in credits. At Microsoft’s current pay-as-you-go rate, 300 credits is $3 / approximately €2.59. One of us ran a full inbox analysis on a quiet tenant and it came back at 66 credits, about $0.66 / €0.57. The same instruction against a large, complicated mailbox can be a very different workload.

But do not judge Cowork economics from one task. Look at your usage over time. Which kinds of tasks consume the most credits? Which ones recur? Which save meaningful time? Which ones are allowing you to do work you could not realistically have done before? Which ones are simply expensive ways of doing something Copilot could already do perfectly well?

One task gives you a price. Repeated use starts to show you where Cowork is actually worth the money.

For organisations, this is becoming a management discipline too. Microsoft now provides central cost management for Copilot Credits in the Microsoft 365 admin center, including spend visibility, budgets, policies and usage controls. (Microsoft Learn)

We have also noticed that producing finished artefacts can consume considerably more credits than producing the information behind them, particularly when the job involves documents, presentations or spreadsheets. That does not mean “files always cost more.” Microsoft makes clear that consumption depends on the complete mix of model, context, tools and runtime. (Microsoft CDN) But it does mean there are times when the smarter workflow is to use Cowork for the work that genuinely requires Cowork, then use Copilot to shape the final artefact.

The same thinking applies to email. Long email threads can become expensive because Cowork has more context to process. A fifty-reply conversation is a different workload from a five-message exchange. Worth remembering before you point it at your inbox.

And use PAUSE on anything big. Instead of:

“Answer all my emails from last week.”

Try:

Process the information. Tell me what you found. Propose what you intend to do. Then stop and wait for my approval.

You decide whether the plan makes sense before the expensive part begins. That pause can be the difference between valuable work and credits disappearing into a task you never defined properly.

Without a definition of done, “get my inbox to zero” is an instruction with a very large surface area.

Two adoption failures, both expensive

Give a team credits and you will usually see two reactions. Neither is the one you want.

The first grabs the popcorn. Great. I’ll use this until it runs out.

The second freezes. This costs money. I am not touching it. I am not going to be the person who burned the department’s budget.

We have seen far more of the second reaction than we expected. And it can be harder to spot because it looks like compliance.

What helps is not particularly exciting. Be explicit about how much budget exists and who it belongs to. Set different expectations for different roles because a project manager coordinating twenty-four projects should not necessarily have the same usage expectations as someone who wants a weekly summary.

And teach people to use Copilot alongside Cowork. This is the step adoption programmes cannot afford to skip. A lot of work does not need Cowork at all.

Knowing when Copilot is enough and when the work genuinely calls for Cowork is becoming an important AI literacy skill. The goal should not be to use Cowork as much as possible. The goal should be to use the right capability for the work. That is cheaper to teach than to discover through your monthly bill.

Skills are where the leverage sits

A skill is a set of instructions you write once. This is how my PowerPoints should look. This is my communication style in email. These are the steps I take every time a project changes phase.

The obvious benefit is consistency. The less obvious benefit can be efficiency because a well-designed skill gives Cowork instructions it would otherwise need to get from you repeatedly.

One of the most useful things you can do is build those skills with Cowork. Ask it to study examples of your email voice and help turn what it observes into reusable instructions. Review what it produces. Strengthen the weak parts. Remove anything it inferred incorrectly.

Mine landed on warm, direct and collaborative. I probably would not have written those three words on a blank page, but seeing them reflected back gave me something useful to work with.

Two practical notes. Skills can be stored as markdown files in your Cowork folder in OneDrive, which means the thinking behind them is portable. And skills are not becoming relevant only to Cowork. Instruction sets are increasingly part of how we work with AI across Microsoft 365.

Our advice is simple. Block half a day. Not “I’ll build skills while I work.” Actually sit down and ask yourself: what do I do the same way every single time?

Most people slowly create three useful skills by accident. You can create a meaningful set in one afternoon on purpose.

And where the risk sits

Skills are instructions guiding an agent that can take action. That distinction matters. If the goal is an email, Cowork is not only capable of telling you what the email might say. It can perform actions on your behalf, subject to the controls and approvals Microsoft provides. (Microsoft Learn)

So decide where you want the human checkpoint. Maybe you want the message drafted but not sent. Maybe you want to review a proposed plan before the actions start. Maybe you are happy for a low-risk recurring task to run automatically. Those are workflow design decisions now.

There is another issue. People are starting to share skills, instructions and agent components widely. That makes reviewing what you install increasingly important. Do not treat someone else’s instructions as harmless simply because they came in a convenient file.

Read them. Understand what they tell the agent to do. Check what information they can access and what actions they can initiate. Test the output. And stay deliberately in the loop when information is leaving your organisation or when the action has meaningful consequences.

Sharing is caring. But sharing is also sharing. What you make available to other people becomes available to other people. That deserves one deliberate thought before you click.

The upside is that learning to design and evaluate good instructions is not really a Cowork skill. It is becoming an AI skill. Getting good at it now will keep paying.

Scheduled tasks change the relationship again

Cowork no longer has to wait for you to come looking for it. Microsoft allows Cowork prompts to run automatically at a scheduled time or on a recurring basis. (Microsoft Learn)

One of us has a news briefing that runs at eight every morning. It uses a handful of trusted sources, filters what it finds through the context of the role, and delivers the result into Teams.

That delivery point matters. Cowork can bring the work to where it is useful to you, rather than forcing every result into the same destination.

Another pattern we like is simple: pull the transcripts from yesterday’s meetings and produce a list of everything I committed to doing. Deliver it at eight in the morning or at noon. Now the promise you made to a customer at four in the afternoon is much less likely to quietly evaporate.

There are all kinds of possibilities here: daily briefings, meeting reviews, recurring project updates, calendar checks, preparation work. Then you shape those patterns around how you actually work.

Two choices are worth making deliberately. First, where should the output land? Your Cowork conversation? Your inbox? A Teams message? Somewhere a colleague can use it? Not everything needs to come back to you.

Second, does the work need to stand alone or build over time? Separate runs are cleaner when you want to scan a specific day. An ongoing thread can make more sense when the work is cumulative and previous context matters.

The point is not scheduling for the sake of scheduling. It is designing the work around when and where it becomes useful.

It works across your applications, not only in its own window

Cowork can work across Microsoft 365 on your behalf. Set an out-of-office response. Find a time across calendars and create the meeting instead of simply telling you when everybody is free. Create documents. Send emails. Post in Teams. Manage files. (Microsoft Learn)

Plugins can extend that reach into additional systems. Connect the right business data and the question becomes much more interesting than “what is in my inbox?” Now the work can combine meetings, email, files and business systems into a picture that no single application holds on its own.

There is also browser use. Cowork can use Microsoft Edge on your device to work with websites you are already signed into. It operates through a hidden Edge tab while you continue interacting with Cowork, and it can hand control back when needed. (Microsoft Learn)

That opens up another category of work. One of us receives somewhere between seventy and ninety LinkedIn messages a day. That is not realistically manageable by hand. Cowork can help scan and organise them. What came in? What looks urgent? What can wait? Where might there be something worth acting on?

Replying automatically to every birthday wish is technically work you could delegate. It probably is not work worth paying Cowork to do. Finding a genuine opportunity, bringing it to your attention and helping move it toward a meeting? That might be. The ability to delegate something does not mean the whole thing should be delegated.

Barcelona: thirty minutes, three outputs

The clearest example we have is not a feature. It is half a day in Barcelona. Fly in, run a session, fly out that afternoon. Five customers in the building and almost no time.

So we spent thirty minutes doing something humans are particularly good at. Talking. A phone sat on the table recording while the facilitator agent ran. Three questions:

• What did you think of the event, and what were other people taking from it?
• What are you going to do with what you learned? What are your next steps?
• And what do you think about the organisation I work for?

At the airport, that conversation became three useful outputs. A summary for the board of what happened in Barcelona and what mattered. A story marketing could use about what customers learned and how they saw the organisation. And individual follow-ups for the account teams based on what their customers had said.

Thirty minutes of being properly present, followed by AI doing the work between conversation and action. That is the part that matters. The value was never the three hours we might previously have spent writing everything up. The value was in the conversation and in what happened because of it. Everything between those two points is exactly the kind of work we should be questioning.

The risk of not using it

People can list the risks of Cowork quickly. Security. Privacy. Compliance. Cost. They are all real. But there is another risk we talk about less. Not learning how to work this way at all.

What Cowork changed most for both of us is not simply output volume. It is capacity. Being able to look at twelve things and say: help me understand what should move first, organise the rest, take the work that does not require me.

Sometimes the answer is that Cowork can take it. Sometimes Copilot is perfectly capable of helping. Sometimes the task should disappear completely. And sometimes the honest answer is that the work belongs with another person because they need to own it. That is a much more interesting conversation than productivity.

There are things neither of us wants to do manually again. Hunting through calendars. Looking up names and addresses. Writing routine meeting invitations. Not because those things were hard. Because they were never really the work.

And perhaps the biggest shift is this: we can now do work we simply could not have done before. Not because we suddenly found more hours in the day, but because some of the work surrounding the work can finally be delegated. That is a very different value proposition from “AI makes me faster.”

Closing thoughts: keep the why

One caution to end on, because this one matters most. Do not hand over your judgement.

Those of us who already know how to learn, question, compare and challenge an answer have some protection against that. People just beginning their careers may have less. The habit of accepting a polished answer can form quickly.

So use AI to learn instead of using it to skip learning. Ask why the answer is good. Ask what another answer might look like. Ask what assumptions were made. Ask what evidence would change the conclusion. Ask whether the same decision still makes sense two years from now.

And ask your colleagues those questions too. We are suddenly surrounded by polished documents and confident presentations. That makes one question more valuable than ever: what was your intent here, and why did you make it this way?

You cannot automate away the need for judgement. Keep that part yours.