Today Microsoft made Copilot Cowork generally available. The headlines will be about what it does — and what it does is genuinely new: hand it a complex, multi-step job and it goes off, uses the tools, and comes back with the finished work instead of a draft.
That is the story everyone will cover. It is not the story that matters most to the person who has to pay for it.
The detail almost everyone will skim past is how Cowork is priced. And that detail quietly changes the shape of every AI budget in the country.
It is not in your seat
Until now, Microsoft 365 Copilot was a flat number. Thirty dollars per user per month, and you were done. You could put it in a spreadsheet, multiply by headcount, and know your AI bill for the year. A line item. Predictable. Boring in the way good budgets are boring.
Cowork breaks that.
The Copilot license is still there — it is the entry ticket. But Cowork itself is billed on top, by usage, in something Microsoft calls Copilot Credits. Pay-as-you-go runs a penny a credit, and what any given task costs depends on the model it picks, how much context it pulls in, how many tools it calls, and how long it runs. A quick task is cheap. A long, heavy, agentic one is not. It is off by default, and Microsoft has built an entire cost-management dashboard — spending limits at the tenant, group, and user level, budgets, alerts, hard caps — because you are going to need it.
Read that last part again. Microsoft shipped the governance tooling alongside the feature. That tells you exactly what kind of product this is.
We have seen this movie
If you ran IT through the move to the cloud, your stomach just did something familiar.
This is the same transition. Flat, predictable licensing gives way to metered consumption. The bill stops being a function of how many people you have and starts being a function of what they actually do. And the companies that got hurt in the cloud era were not the ones who adopted — they were the ones who adopted without visibility. They lifted, they shifted, they turned everything on, and three months later finance was staring at a number nobody could explain.
The meter does not punish usage. It punishes blind usage.
AI just crossed the same line. The license was a fixed cost you could ignore. Consumption is a variable cost you cannot.
You cannot govern what you cannot see
Here is where it stops being a Microsoft story and starts being a you story.
When the bill scales with usage, the question is no longer “how many seats do we buy?” It is “which of the things our people are doing with this are actually worth the credits — and how would we even know?”
That question has no answer in a vendor dashboard alone. Microsoft can tell you that a user burned a thousand credits last week. It cannot tell you whether the work those credits produced was worth more than it cost, because the value lives in your business, not in Microsoft’s meter. Connecting “we spent this” to “we got that” requires the boring layer underneath — the one that ties usage to outcomes, the one that lets you see which workflows pay for themselves and which are just expensive habits.
That layer is not a feature you switch on. It is something you have to build, and then watch.
The actual game
The companies that win this shift will not be the ones who spend the least on AI, and they will not be the ones who spend the most. They will be the ones who can see their own usage clearly enough to spend on purpose.
That is a familiar discipline. It is the same one that separated the companies who got the cloud right from the ones who got a surprise invoice. Treat AI like the metered utility it now is. Give it a budget, a meter you actually read, and an owner who connects the spend to the result.
The meter started running today. The only question that matters is whether you can see it.