What does Copilot Cowork actually cost?

Since Microsoft made Cowork generally available in June, I keep getting the same question. What does that thing cost per month? And every time I have to give the same answer: it depends, and probably more than you would like.

That is down to the way Microsoft bills it. With Copilot you pay a fixed amount per user per month and that is that. With Cowork you do not. There you pay per job you have it carry out. The heavier the job, the higher the bill. You are no longer buying software, you are buying compute time.

Microsoft has a spreadsheet for this themselves. Useful, but it is full of terms you have to decode first and it leaves out a few things you actually want to know. I rebuilt it into a calculator you can simply fill in below.

First: what is a credit?

Cowork bills in Copilot Credits. One credit costs a cent. How many credits a job costs depends on four things: which model runs underneath, how much information it has to fetch, how many other systems it has to consult along the way, and how long it is busy.

Microsoft sorts jobs into three kinds:

Light. A simple question with little context. Around 125 credits, so a little over a euro.

Medium. Richer context, several sources, two or more results. Around 500 credits, so a bit over four euros.

Heavy. Pulling information together from many sources, the heavy model, a long run, several documents as the result. Around 1,200 credits, so a bit over ten euros.

Ten euros for one job sounds steep. Until you consider that a job like that replaces half an afternoon of manual work. The problem is not those ten euros. The problem is how often your colleagues are going to do it.

Why almost everyone underestimates this

Microsoft looked at what people in their early trial programme actually did. An average office worker issued 22 light, 11 medium and 5 heavy jobs per month. Work that through and you land on over €120 a month. For one person.

For technical staff it is higher, because they issue far more heavy jobs. They come out at almost €200 a month.

Put that next to the licence for Microsoft 365 Copilot, which costs around €26 per person per month. Consumption is then five to eight times as expensive as the licence itself. That is exactly what goes wrong in most quotes I see go past. The licence is calculated, and the consumption arrives later as a surprise.

And then the unpleasant part. The workflows that save the most time are also the workflows with the most steps. And more steps means more credits. So adoption pushes your bill up, not down. The better it works, the more expensive it gets.

How to fill in the calculator

It goes in four steps and you are through it in two minutes.

Number of people per role. An example company of 90 people is filled in. Replace those numbers with your own. The four roles come from Microsoft’s model, because their consumption differs measurably.

Adoption. Microsoft assumes that everyone with a licence will actually use Cowork. In the first year that is rarely the case. Set the slider to what you think is realistic.

Jobs per month. These are the figures from Microsoft’s trial programme. If you know your own people work differently, adjust them. This is the dial with the most effect on the outcome.

Rates. A cent per credit is list price. Buy a volume up front and your effective rate is lower, so enter that here. The licence cost and the dollar rate are there too, because you want those in the total picture.

How to read the outcome

At the top there are three numbers. The credits per month, what that costs in euros, and what it averages per user. Below them is the comparison with your licence cost. That is the sentence you want in your business case.

The chart underneath shows what one employee per role costs per month. The narrow dark block is the licence, the long coloured bar is the consumption. If you want one picture to convince somebody that this works differently from ordinary software, this is the one.

Right at the bottom there are three amounts side by side. Better than expected, your expectation, and worse than expected. I put those there deliberately, because one number gives false certainty here. Nobody knows in advance exactly how heavily their people will use Cowork, and you are allowed to admit that.

What I would do before you switch anything on

Cowork is off by default. An administrator has to switch it on, and that is just as well. Use that threshold.

Set a spending limit in the admin centre first, and set it at your upper bound rather than at your expectation. That way Cowork does not stall halfway through the month, and the bill cannot run on without limit either.

Then start with one group. Preferably the most expensive one, because that is where you learn the most. Let that group run for a month or two and see what actually happens. Treat those first months as paid research rather than as your structural cost.

Finally, look at the model running underneath. Microsoft’s figures assume the heaviest model they offer. Putting routine work on a lighter model saves credits immediately, and in most cases nobody notices the difference in the result.

Small print

The method and the default values come from the spreadsheet Microsoft distributes themselves. They base those figures on anonymised data from their trial programme, with a reference date of 27 May 2026, and state explicitly that it is an illustration rather than a prediction. I added the licence cost, the adoption rate, the conversion to euros and the range.

All amounts are indicative and exclude VAT. Prices in this market change every few weeks, so check the current rates before you put your signature under anything.