Copilot Cowork is generally available. This is where it starts getting expensive.

Copilot Cowork is the most capable thing Microsoft has shipped under the Copilot name, and on 16 June 2026 it stopped being a preview. Hand it a real objective and it will go and complete the work, start to finish, while you get on with something else. For most businesses this quarter, the capability is not the story. The billing is.

Out of Frontier, onto the meter

Cowork spent its preview inside a programme Microsoft branded Frontier, which opened on 30 March. On 16 June it went generally available worldwide, and Microsoft changed the charging model with it. The licence has not gone away: you still pay for Microsoft 365 Copilot, around US$30 per user per month before any volume discount, just to have access. Cowork then bills again, on top, for the work it does.

That second bill is the new part. Microsoft meters it in a usage currency it calls Copilot Credits, priced at one US cent each. You can pay as you go, or commit to a block of usage up front for a discount. Billing started on 16 June, with one concession: tenants that had someone using Cowork during the Frontier preview get a short grace period and are not charged until 1 July 2026. Everyone else is on the meter now.

You pay for the licence, then you pay again for the work.

A prompt is one thing. A task is another.

A chat prompt is cheap because it does one small thing. You ask, the model answers, you read it, the transaction is over. A Cowork task is a different unit of work. You give it an objective and it runs the steps on its own: it searches, it reasons over your files, it calls tools and third-party plugins, it can drive a web browser, and it comes back with finished output instead of a suggestion. This is what the industry means by “agentic”: software that does not just answer, it goes and completes multi-step work for you. It is useful. It is also why it cannot be priced like a chat message.

Microsoft builds each task’s price from four inputs: the model it runs, the context it pulls in from your business, the tools it calls, and how long it runs. It sorts tasks into light, medium, and heavy. A light task reads a few sources, reasons lightly, and produces a single output. A heavy task aggregates broadly, reasons deeply, and produces several. The deeper and wider the task, the more credits it burns. Early analyses of those tiers put a light task at roughly one to three US dollars, a medium task at four to seven, and a heavy, multi-output task at seven and up. (Computerworld walked through the arithmetic the day after launch; Microsoft’s usage-based billing guide sets out the mechanism.) A cent a credit sounds harmless. A heavy task runs to hundreds of credits, and a workforce runs hundreds of heavy tasks a week.

Agentic AI doesn’t budget like a per-seat licence

We have spent fifteen years buying business software the same way: per user, per month, known in advance. You could put the number in a spreadsheet in January and still be right in December. Consumption breaks that. The bill is the sum of what your people actually did, and people are not predictable. Every heavy task, every vague request re-run three times because the first two missed, every workflow somebody rebuilds from scratch instead of reusing, lands on the invoice.

It is easy to make the figure large. One illustration doing the rounds, a public Cowork cost calculator built on Microsoft’s own Frontier-usage averages, models a mixed organisation of around 1,550 people and lands near US$418,000 a month of Cowork usage, on top of the licences. Treat that as one scenario and nothing more. Change the share of heavy users, or cap what the agent is allowed to do, and the figure moves a long way down. We built a Copilot Cowork cost calculator so you can model your own mix in Australian dollars instead of ours, and watch how far the number swings. The figure is not the lesson. The unpredictability is.

The quieter risk is the one that actually bites. Most organisations will not know what they are spending until the bill arrives. A per-seat licence cannot surprise you. A consumption bill can, and the first month is usually where it does.

If your staff use Cowork badly, you’ll pay for it

Consumption pricing turns everyday habits into a line item. Four of them cost real money:

• Vague tasking and reruns. A poorly framed task searches more, reasons more, and produces more than it needs to, and someone who is not sure what they asked for runs it again. Each run bills.
• Everyone rebuilding the same workflow. If forty people each invent their own way to produce the weekly report, you pay for forty heavy tasks instead of one shared skill the whole team calls.
• Ungoverned connectors and plugins. Every plugin a task is allowed to call is a tool call you can be billed for, and a path your data can travel. Left open, the surface only grows.
• Managers who can’t tell delegation from convenience. The judgement that now matters is knowing which work is worth handing to an agent at heavy-task prices, and which is a thirty-second job a person should just do.

The new AI literacy

For two years, “AI literacy” has meant one thing: can your staff write a good prompt. Cowork retires that definition. Prompt-writing is table stakes now. The literacy that earns its keep is judgement about cost and control. Which tasks are worth delegating to an agent, and which are the thirty-second job. Which justify a heavy task’s price, and which belong in ordinary Copilot Chat. Which teams get access at all, and which model each task should use. Which repeated workflows get promoted into a central skill so the team stops paying to reinvent them. Which plugins are approved to be called. And underneath all of it, how you measure whether the output came back worth more than it cost, before the usage runs away from you.

A work execution layer, not a chat box

It helps to be precise about what Microsoft shipped, because the capability is the reason any of this is worth the trouble. Cowork runs complex, long-running, multi-step tasks and, in Microsoft’s words, “returns a completed result, not just a draft or a recommendation.” The examples in the announcement are concrete: editing a batch of spreadsheets and drawing the dependency flow chart between them, comparing one file across several product versions, working a sales pipeline and handing back a ranked list of opportunities. It launches with nine third-party plugins, including Harvey for legal work, monday.com, and Miro, with more to follow. This is a layer that does work.

The capability that should get an administrator’s attention is the browser. Cowork can drive a local Microsoft Edge session, following the enterprise policies you already have in place. That matters more than it first sounds. A large share of real business work does not live in Microsoft 365 at all. It lives in portals, CRMs, finance systems, and supplier sites, behind a login, with no clean connector. If an agent can operate the browser, it can reach that work the way your staff do, by signing in and clicking. That closes a genuine gap, and it is the capability you most want governed, because an agent driving an authenticated browser session is an agent acting as a logged-in member of your staff.

Who controls the workflow, and who gets access

Two questions sit underneath all of this, and neither is technical.

The first is control. Once a single Cowork task chains research, then drafts the email, then writes the briefing, then builds the deck, then sets up the follow-up and posts the summary to Teams, a real slice of someone’s job now runs without them. Who owns that workflow? IT, the business unit that runs it, the person who kicked it off, the plugin vendor whose tool it called, or the agent itself? In most organisations on day one, nobody has decided, which in practice means the agent has.

The second question is sharper, and it is the one most leaders have not sat with yet. If Cowork lifts a person’s output, then access to it becomes an advantage at work. So who gets it? Everyone? Only senior staff? Only the revenue teams? Only your proven performers? Each answer has a consequence. Give it to a few and you create a tier of people who produce far more than the rest, while everyone else is told to do more with less. That shows up quickly in who hits their numbers, who gets the credit, who gets promoted, and who gets trusted with the bigger work. An access decision made casually, on cost grounds, can become a decision about your people’s careers without anyone meaning it to.

Controls exist. A dashboard isn’t a strategy.

Microsoft has done the controls properly, and it deserves the credit. Cowork is off by default. Admins decide when to switch it on and who gets access. You can set spending limits across the whole tenant (your company’s Microsoft environment), a group, or an individual, with customisable alerts when usage crosses thresholds you choose. When a user hits a wall mid-task, they can request more credits instead of running past the cap unseen. That is the right toolkit, and it is more than most consumption products ship with.

One piece is not there on day one. Per-task pricing shown to the user, the thing that would let a staff member see that the task in front of them costs eight dollars before they run it, is listed as coming after general availability. So at launch your people can spend without seeing the price of the thing they are about to start, while you watch the total climb on a dashboard. That gap is worth knowing about before you switch Cowork on.

A dashboard, however good, has a hard limit.

A dashboard doesn’t create a strategy. It only tells you what happened.

Deciding what should happen is your job, and the way to do it is not to turn Cowork on for the whole company and watch the graph. Run it like an experiment. Give it to a small group, on real workflows that matter, for a fixed period. Measure two things: what it cost, and what it produced. Compare that against the same work done in ordinary Copilot Chat, or by a person. Where a workflow is repeated, valuable, and cheaper or better as an agent task, promote it into a central skill the team reuses. Where the work is ad hoc, vague, or a person could do it in a minute, keep it out. Cowork earns its keep on reusable, high-value work. It bleeds money on the rest.

Genuine step forward, or expensive tool used badly?

Cowork is a real step. The open question is whether it is a step your organisation is set up to take well, or one it will take expensively and by accident. Two questions settle it.

Would you pay hundreds of dollars per user per month, on top of the licence you already buy, if the output justified it? For some teams and some workflows the answer will be a clear yes. For others it will be waste dressed up as progress, and the only way to tell them apart is to measure first, while the spend is still small.

And would you give it to everyone? Or are we about to watch a divide open up inside organisations, between the people with an agent doing their heavy lifting and the people told to keep pace without one? That is a leadership decision, not an IT one, and it is better made on purpose than by default.

This is the work we do with clients before the first big bill, not after it: switch Cowork on for a controlled group, put real spending limits and governance around it, measure cost against output on the workflows that actually matter, and decide who gets access, and why, on the evidence. If you are weighing up Cowork for your business, get in touch and we will help you run the experiment while it is still cheap to run.