AI that does the work, not just answers.
Describe the outcome. Cowork breaks it into steps, carries the work across Outlook, Teams, and the M365 apps, and returns a finished result for your approval. Bring it into everyday work the controlled way: the right use cases, credit budgets, and a measurable pilot.
An agent that carries the work to the finish.
Cowork isn't the next version of Copilot chat. It's an agent that plans, does the work, and returns a finished result, not a draft.
- ▸ An agent in M365 Copilot, built on Anthropic's Claude
- ▸ Plans, does the work and returns a finished result, not a draft
- ▸ Cloud-based execution: the work continues even with the laptop closed
- ▸ You see the progress the whole time and can steer or stop it
- ▸ Generally available from June 2026, off by default
- ▸ Billed with Copilot Credits, one task at a time
The cheapest sufficient tool for every task.
Cowork is priced per task, so it shouldn't be used for everything. The right question isn't "can it do this" but "is this task worth the credits".
Copilot Chat
Questions, summaries, drafts. Included in the Copilot license, no metering.
Copilot in the apps
Word, Excel, PowerPoint, Outlook. The work happens where the content is. Included in the license.
Copilot Cowork
A multi-step task from start to a finished result. Billed in credits per task.
Copilot Studio
When the same work repeats for a team or department, it belongs in an agent, not a one-off task.
What the pilot group uses Cowork for first.
Preparing for client meetings
Background, email history, earlier conversations and open items collected into one finished brief.
Compiling monthly reports
Data from different M365 sources into one readable report without manual collation.
Calendar management and prioritization
The week's calendar organized by your goals, with rearrangements proposed in time.
Deep research work
Research combining web and M365 data, progressing in the background step by step.
Meeting follow-ups
A summary, CRM updates and follow-up messages drafted, always for a human to approve.
Recurring personal rhythms
A daily brief, weekly digest or end-of-day wrap-up without re-briefing.
Every task costs money. So we get consumption under control first.
Cowork is billed with Copilot Credits per task, and you see each task's credit cost as it runs. Without limits and monitoring, the monthly invoice can be a multiple of the forecast.
Frontier models cost more than efficient base models. Model choice is steered by the task.
How many documents, messages and how much data the task goes through.
Which applications and connectors the task touches.
How long the task runs. Long runs consume more.
Caps at tenant, team and user level before wide use. A named owner for the budget.
Consumption spikes are caught in days, not at the end of the month. Usage reporting per user and team.
A good task brief is a different skill from chat prompting, and it shows directly in credit consumption.
Time saved, result per credit, ROI per role. Budgeting from observed use, not wishes.
The same governance model scales as usage grows, and it is directly compatible with the Agent 365 governance model.
Honestly: what IT needs to know before activation.
Cowork is off by default: the admin decides who gets access and with what limits.
Sensitive actions behind approval gates: external emails, calendar changes, file sharing, system entries.
M365 controls apply: audit, eDiscovery, classifications and retention. DLP support is still coming; do not assume it.
EU Data Boundary: what this means in practice +
- ▸ Anthropic is not part of the EU Data Boundary commitment.
- ▸ Anthropic acts as Microsoft's subprocessor (DPA); the data is protected by contract.
- ▸ Flex routing (MC1269223) is in effect from 17 April 2026.
- ▸ In an EU/EFTA tenant, Anthropic is off by default.
- ▸ A GDPR assessment is recommended in regulated industries.
- ▸ A new EU data zone from Anthropic is coming; the schedule is not yet confirmed.
One team, four weeks, an expansion decision.
We don't recommend Cowork for the whole organization at once. A limited pilot group uses it on real work, and the expansion decision is based on experience and data.
Understanding
What Cowork is, Anthropic's role, credit pricing and the EU data boundary situation. Decisions are made on facts, not a sales deck.
Technical rollout
Activation, pilot group selection, access management (Entra ID), credit limits and alerts in place before the very first task.
Training for the pilot group
Cowork in practice, in people's own real work, not exercises. Briefing a task, acceptable use, security awareness.
Monitoring and the expansion decision
Usage data and feedback weekly for four weeks. At the end, an expansion decision based on experience and data.
Where you are after the pilot.
Validated use cases
You know where Cowork creates value in your everyday work and where it is not yet worth expanding.
Credit consumption under control
Budgets, limits and alerts running before wide use. No surprise invoices.
A trained pilot group
In-house champions who carry the rollout as Cowork expands to new teams.
Privacy questions handled
The EU Data Boundary and approval gates reviewed before the pilot, not after it.
A justified expansion decision
The decision forms in four weeks from real usage experience and data, not guesswork.
A path to agents open
Cowork makes the need for process automation visible. Recurring processes continue into Copilot Studio.
Does this sound familiar.
Your organization already uses Copilot and wants the next step.
Multi-step expert work takes hours that could be delegated to an agent.
You want credit costs and security under control before wide use.
Related reading.
Launch the pilot, in control.
In a first conversation, we'll pick the pilot group, the first use cases, and the credit limits. Privacy questions get answered before activation, not after.