Copilot Cowork went generally available worldwide today. After three months in the Frontier preview, more than half of the Fortune 500 is already using it. The headlines are about the price. I want to talk about everything the price tag hides, because that is where the actual leadership work lives.

Let me start with what changed, then walk through the four decisions that land on your desk the moment you read the announcement. Aanzetten, de modellen, de kosten, en welk werk waar hoort. None of them are settings. All of them are leadership.

What actually shipped

For two years we asked AI for drafts. We prompted, we generated, we admired the output, and we did the last mile by hand.

Cowork is a different kind of thing. You define the work, it runs end to end across your tools and files, and it hands back a finished result. Not a suggestion. The thing, done. Microsoft’s own examples make the shape clear: one team compared nearly four thousand files across two product versions, work that would have taken weeks. One sales lead pointed it at a stalled pipeline and got back a ranked list of at-risk deals with the exact follow-up that had gone cold on each, a week of review collapsed into a morning.

That is the line that matters. Cowork is built for long, multi-step, finished-while-you-are-in-a-meeting work. Hold that thought, because it decides everything that follows.

Decision one: it is off by default, and not only for budget reasons

Cowork does not switch itself on. An admin decides when it is enabled in your tenant, for whom, and with what spending limit. Nothing runs, and nothing gets billed, until someone makes that call.

For a Dutch company, that default carries a second meaning most of the noise is missing. Cowork runs on Anthropic’s models. They are excellent models. But Microsoft’s own documentation is explicit that Anthropic models sit outside the EU Data Boundary, and the processing happens on Amazon Web Services infrastructure in the United States. That is exactly why Anthropic models are switched off by default for every customer in the EU, the UK, and EFTA.

So before a single task runs on European work, someone has to decide to send that work across the Atlantic. Microsoft is refreshingly blunt about the conclusion: if strict adherence to the EU Data Boundary is non-negotiable for your organization, leave the access to Anthropic models off. Anthropic is a Microsoft subprocessor under the Data Protection Addendum, which removes the old friction of accepting a separate vendor’s terms. It does not remove the data-residency question. That one is yours to answer, against your sector, your policies, and the promises you have made to your own clients.

This is not “Claude is unsafe.” It is that the compliance posture changes the moment a non-EU model becomes the reasoning engine. Know that before you flip the switch, not after.

Decision two: the models, and the GPT question

“Then just pick GPT,” someone will say. Fair, and worth being precise about.

At general availability, Copilot Cowork runs on Anthropic models, Opus 4.8 and Sonnet 4.6. Model choice, including GPT 5.5, currently lives in the Frontier preview, not in the version that went broadly available today. Microsoft has also said a fine-tuned, lower-cost model called Cowork 1 is coming.

So for most organizations, right now, the real choice is not which model. It is on or off. And each model carries its own governance reality, so do not assume the EU posture is identical across every Copilot surface. Check each one.

Decision three: the money stops being a flat fee

This is the shift that changes behavior before it changes the budget. AI used to be a fixed line. A seat, a license, all-you-can-eat, predictable. Cowork bills per task, in Copilot Credits, on top of the Microsoft 365 Copilot license. Pay-as-you-go is priced at one cent per credit, with a commitment option for a discount.

Two things follow that almost nobody is connecting.

First, your AI cost will climb. Not because the price per task went up. Microsoft is right that unit costs will fall over time, as models get cheaper and the system gets better at matching the right model to each task. The total still rises, because cheaper work means more work. The line item grows even as the price drops. That is not a contradiction. It is the oldest pattern in technology.

Second, and this is the uncomfortable one: cost you can see is value you can finally measure, and most companies will not bother. Only around one in five CFOs can point to measurable value from AI today. The credit meter is about to make the cost visible to everyone. The value will stay visible to the few who do the harder work of measuring it.

The good news is that the control sits with you. Cowork ships with spending limits at the tenant, group, and user level, customizable usage alerts, and usage reporting that breaks spend down by person and feature. The budget discipline is available. Whether it gets used is a leadership choice, not a product feature.

Decision four: what do we use Cowork for, and what stays in M365 Copilot

Here is the trap. Because Cowork bills per task, asking it the weather in Amsterdam costs you a fraction of a cent. Almost nothing. But you just sent a Ferrari to pick up a carton of milk.

You now have two tools for two kinds of work, and the temptation is to use them interchangeably. Resist it.

M365 Copilot is for the work in the flow. Draft the email, summarize the thread, catch me up. Fast, light, already inside your per-user license.

Cowork is for the work you used to block half a day for. Five jobs it is genuinely built for:

  1. Respond to an RFP. Pull from every proposal you have ever won, your product docs, and your pricing, and draft a tailored first version. A week of work, back before lunch.
  2. Run due diligence on a company. Across filings, news, the open web, and your own CRM, and hand back a one-page memo with the red flags on top.
  3. Check a stack of contracts against your own policy. Every deviation flagged with the exact clause, not a summary that misses the one that matters.
  4. Build the board narrative. Take the actuals, last quarter’s deck, and the numbers that moved, and draft the story you will have to defend on Monday.
  5. Read the week’s meeting transcripts and hand back what you promised in each one, what is still open, and who owes whom by when. Your follow-up list, written while you slept.

See the pattern. Long, multi-step, finished while you are doing something else. Not one of them is a question you could have answered yourself in ten seconds. New partner plugins, browser use through Edge, and Work IQ context across your meetings, mail, and files are what make this range possible.

The skill that matters now is not prompting. It is judgment. Knowing which work belongs where. Spend Cowork on the heavy lifting, not the weather.

The bigger shift hiding under all of it

Step back from the features and you see the real change. We are moving from access-based pricing to outcome-based pricing, and that quietly forces a discipline we have avoided for years.

When every task has a price, you stop starting work on reflex. You think first. What is the outcome I actually want. Is it worth what it costs. What does good look like before I press go. The industry is heading the same way: consumption and outcome-based models, digital labor budgets, service levels for work measured in cost-per-outcome and success rate. The winners will not be the companies running the most tasks. They will be the ones running the right ones, on purpose.

That one in five number is not a verdict on the tools. It is a verdict on how we designed the work. Most of it was never built around an outcome in the first place, so when the meter switched on, there was no value to point to. The price tag just made the gap visible.

What I would actually do this week

Four moves, in order.

Decide your EU posture before you enable anything. Map which work would touch a non-EU model and hold that against your obligations and your client promises. For some of you, the honest answer is to keep Anthropic models off until the questions are answered. That is a legitimate, defensible call.

Set the guardrails before usage ramps, not after. Spending limits, alerts, and reporting exist on day one. Turn them on first.

Draw the line between Cowork and M365 Copilot for your teams, in plain language. Give people two or three concrete examples of each. The Ferrari does not go to the shop for milk.

Pick one heavy, recurring job and measure it properly. Time saved, cost per run, quality after review. One honest number beats ten enthusiastic pilots.

The tool arrived finished today. The thinking did not. Where your data lives, what the work is worth, which work belongs where, and how you redesign it so the answer is yes. None of that is in the settings menu. All of it is the job.