Copilot Cowork: What It Is and How to Use It

At a glance

What it is
An AI agent that takes action inside M365 —
sends emails, posts to Teams, creates calendar blocks.

Key difference from Chat
Chat drafts — you implement.
Cowork executes — with your approval at each action.

Built on
Anthropic Claude agentic model,
wrapped in Microsoft enterprise governance.

Use case	What you describe	For
Meeting follow-up	Find my Teams transcript, draft and send the follow-up email to all attendees	Anyone who runs meetings
Weekly status post	Every Friday, pull my project emails, draft the Teams update, post after I approve	Project manager · Team lead
Focus time protection	Find gaps in my next 5 days, propose 90-min focus blocks, create them after I confirm	Anyone in back-to-back meetings
Commitment tracking	Find every "I'll send / I'll follow up" in my emails from the last 14 days, flag overdue	Account manager · Consultant
Conflict resolution	Find calendar conflicts next 7 days, propose resolutions, reschedule after I approve each	Manager · Executive
Renewal outreach	Search my history with [contact], draft a brief personalised email, send after I approve	Account manager · Customer success

Cowork vs Copilot Chat — the one distinction that matters

Aspect	Copilot Chat	Copilot Cowork
Output	A draft or answer — you decide what to do with it	An action taken — email sent, event created, post published
Execution	You read, copy, paste, act manually	Cowork executes; pauses for your approval at each step
Time span	Single response, immediate	Multi-step, runs over minutes with checkpoints
Steering	Static response; you decide next turn	You can interrupt, redirect, pause, or cancel mid-task
Scheduled tasks	Not available	Up to 5 recurring prompts that run on a schedule
Custom skills	Not available	Up to 50 saved instruction sets, auto-loaded per conversation

Billing — Cowork runs on usage, not a flat seat

Your M365 Copilot licence is the access gate. Cowork usage is billed separately on a consumption currency Microsoft calls Copilot Credits. Pay-as-you-go is $0.01 per credit; a commit-for-discount plan ("P3") is available for tenants that want a lower rate against a usage commitment. Usage billing is off by default — an admin must turn it on before any Cowork task will run. For a full breakdown of what Cowork consumption costs in Copilot Credits, see the cost model reference.

What you pay for

A task's cost = four inputs

Every Cowork task draws credits across four things combined: model use (the LLM tokens), context retrieval (searching your M365 data), tool calls (the actions it takes), and runtime (how long the task runs). A short "draft one email" task is cheap; a multi-step research-and-execute task that reads dozens of files costs more.

Model picker

Choose the model to manage cost

Cowork lets you pick the model per task. As of 14 August 2026 the Learn model page lists Auto (the default, Microsoft routes the task), GPT 5.5 (Frontier, hosted in Azure AI Foundry), GPT 5.6 Sol, GPT 5.6 Terra, Opus 5, Claude Sonnet 5 and Claude Fable 5. A separate effort control (Light, Medium, High, Extra High, Max) sets how hard the model works and how quickly it consumes a user's limits, which makes it a cost lever alongside the model itself. Use Sonnet for routine drafting and reserve the heavier models for complex multi-step reasoning to keep credit spend down. Governance flag: Fable 5 is in preview and stays off until an admin turns it on under Copilot settings in the admin center. It requires data retention, so prompts and responses for that model are kept by the model provider instead of following Cowork's default no-retention posture, and Cowork shows a persistent banner while it is selected. Treat it as a separate approval, not a dropdown choice. An admin can also switch the Anthropic model family off outright. Two entries recorded here on 5 August 2026, Claude Opus 4.8 and the paired Sonnet + Opus Advisor mode, are absent from the 14 August table; Microsoft publishes no diff and no retirement notice, so read that as the documentation changing rather than a confirmed removal. Model lists move monthly, so verify the live picker before writing it into a policy.

Off by default

No Cowork usage is billed until an admin explicitly enables usage-based billing for the tenant. Nothing runs up a bill silently.

Spend limits & alerts

Admins set spending limits and usage alerts at tenant, group, or individual user level — caps so a single user or team can't overspend.

Cost Management dashboard

A new Cost Management dashboard in the M365 admin center reports Copilot Credit consumption by user, group, and feature.

Frontier users — billing grace until July 1, 2026

Tenants that had a Cowork Frontier user during the preview window (Mar 30 – Jun 16, 2026) are not billed for Cowork usage until July 1, 2026. Use the grace window to set spending limits and alerts before the meter starts.

Setup — 4 steps

Verify your licence

Confirm you have the M365 Copilot add-on assigned (requires M365 E3 or E5 base). Check under portal.microsoft.com → Licences. Without this licence, Cowork is not accessible.

Confirm usage billing is set up

Cowork is metered in Copilot Credits and is off by default. An admin must turn on usage-based billing and set spending limits before tasks will run. Check with IT, or see the Billing section below for what each task costs and how limits work.

Open Cowork

Once your licence and usage billing are in place, Cowork appears in the left sidebar at m365.cloud.microsoft. Select it. The interface shows a chat input and a Tasks panel on the right.

Describe your first task

Type a task outcome in plain language — not a question, a delegation. "Find every unresponded email from this week and draft replies for each" is better than "Can you help me with emails?" See the Delegation Brief format below.

The Delegation Brief format

Structure every Cowork prompt with these 4 parts

How to write for Cowork — the 4-part structure

Part 1

Goal

The outcome you want — not the steps. State what done looks like, not how to get there. "Send a follow-up email to all attendees" is a goal. "Search Teams, then draft an email, then send it" is steps — Cowork works those out itself.

Part 2

Inputs

Which M365 sources to pull from. Name them explicitly: Outlook, Teams, Calendar, SharePoint folder, specific meeting transcript. Cowork can infer, but named sources produce fewer errors than "look at everything."

Part 3

Output

What to produce and in what format. "A Teams post in bullet format" is different from "a formal email under 150 words." The more specific the output description, the less iteration you need on the draft.

Part 4

Boundaries

What Cowork must NOT do without your explicit approval. "Show me the draft before sending." "Do not create any calendar event until I confirm the full list." "Do not contact anyone outside my organisation." These are the lines that prevent surprises.

Example — meeting follow-up

Goal: Send a follow-up email after my last meeting.

Inputs: Find the Teams transcript or my notes from the last 2 hours.

Step 1: Extract action items (what, owner, deadline), decisions made, and open
questions. Show me this list and wait for my confirmation before proceeding.

Step 2: Draft a follow-up email to all attendees with the above structure.

Step 3: Show me the full draft and the complete recipient list. Wait for my
explicit approval.

Step 4: Send only after I approve.

Actions Cowork can take — by category

Email & messaging

Draft and send emails
Reply to and forward emails
Post to Teams channels
Send Teams direct messages
Draft messages for your approval

Calendar & scheduling

Create calendar events
Add attendees to meetings
Block focus time
Reschedule events
Find free slots across attendees

Documents & files

Create Word, Excel, PowerPoint
Generate reports and summaries
Browse SharePoint and OneDrive
Read and analyse file content
Share documents with named users

Research & analysis

Search across all M365 data
Read meeting transcripts
Synthesise from email threads
Deep Research across enterprise
Daily and weekly briefings

What it cannot do

Delete files or folders
Create Loop pages
Upload files to Teams
Access local device files
Access systems outside M365
Read encrypted files
Process files over 200 MB

Plugins (May 2026)

HubSpot
Notion
Miro
monday.com
Dynamics 365
Power BI
Custom plugins (org-built)

Scheduled prompts & custom skills

The two features that make Cowork a system, not a tool

Scheduled prompts — 5 slots, use them carefully

Monday week-start briefing

Calendar, email priorities, and commitments due this week — before the day starts.

Highest priority

Weekly commitment tracker

Every "I'll send / I'll follow up" from the last 14 days, sorted by age, flagged if overdue.

High priority

Daily inbox priority scan

Top 5 emails needing action today, with a one-line "why this matters" per email.

Medium priority

Project status digest

Weekly summary from a specific project's emails, Teams, and calendar — posted to Teams channel.

Medium priority

Friday end-of-week digest

What was completed, what is pending, what needs attention Monday.

Optional

How to create a scheduled prompt: Describe the task to Cowork and include a time — "every Monday at 7 AM" or "every Friday at 4 PM." Cowork will propose the schedule and ask you to activate it. Manage existing schedules under Tasks → Scheduled tab. Each scheduled run still requires your approval before Cowork acts.

Custom skills — save a workflow, reuse it without retyping

How skills work

A skill is a saved instruction set — a SKILL.md file stored in your OneDrive. Cowork reads all your skills at the start of every conversation and knows to use them when a matching task comes up.

Limit: Up to 50 skills per user.

Storage path:
/Documents/Cowork/skills/{skill-name}/SKILL.md

Create a subfolder per skill, then create SKILL.md inside it. You can also add up to 20 companion files per skill (templates, reference docs) — stored in the same folder.

Updating a skill: Open the SKILL.md file in OneDrive and edit it directly. Changes take effect at the next conversation. Or ask Cowork to update the skill during a conversation — it will modify the file if you confirm.

SKILL.md format

---
name: Weekly Project Update
description: Posts a structured weekly
  update to a Teams channel from
  the last 7 days of project emails
  and Teams messages.
---

# Instructions below

Every Friday at 4 PM, search emails
and Teams from the last 7 days for
[project name].

Draft a Teams post:
- ✅ Done this week
- 🔄 In progress
- 🚫 Blocked (with owner)
- 📅 Next week

Show me the draft. Wait for approval.
Post only after I confirm.

What Cowork can read

Cowork reads from your M365 environment — email, Teams, calendar, files. It does not access systems outside M365, cannot read locally stored files, and cannot retrieve information from the internet.

Outlook email — inbox, sent, search

Outlook calendar — events, availability, invites

Teams channels & chats — read and post

Meeting transcripts — via Teams recording

SharePoint files — documents, folders

OneDrive files — personal and shared

Dynamics 365 — rolling out 2026

Power BI dashboards — coming soon

Local device files — not accessible

Internet / web search — not accessible

SAP · Primavera · external ERP — not accessible (custom plugins can bridge this)

Encrypted files — not readable even with access

Available now

In pipeline

Not accessible

The approval model — what Cowork shows you before acting

Low risk

Reading emails, searching Teams, browsing files, drafting content for review — no approval required.

Medium risk

Creating calendar events, creating documents, posting to Teams — shows approval dialog with preview.

High risk

Sending emails, modifying shared files, rescheduling others' meetings — shows approval dialog with prominent risk indicator.

01
Every sensitive action shows a full preview before executing. For emails: you see the full draft and recipient list. For calendar events: you see the proposed time and attendees. For Teams posts: you see the full message and the channel.
02
Cowork is not fully autonomous. There is no setting that removes approval requirements for sensitive actions. Human-in-the-loop is the design, not a limitation.
03
"Don't ask again" applies to the current conversation only. If you click the dropdown on an approval button and select "don't ask again," Cowork skips similar approvals for the rest of that conversation. It does not persist to future conversations.
04
You can pause or cancel mid-task at any point. If Cowork is mid-execution, you can interrupt, redirect, or stop it. Actions already approved and taken cannot be undone by Cowork — use normal M365 tools (e.g., Outlook recall) for those.

EU / EFTA / UK tenants — action required before Cowork works

Cowork is built on Anthropic's Claude model, which operates as a Microsoft subprocessor. In EU, EFTA, and UK tenants, Anthropic is disabled by default because data may be processed outside the EU Data Boundary.

To enable: admin.microsoft.com → Copilot → Settings → View All → AI providers operating as Microsoft subprocessors → Anthropic → Enable. This requires an admin to explicitly acknowledge that data will leave the EU Data Boundary. Users cannot enable this themselves — it requires tenant-level admin action.

Failure modes to expect

Cowork searches the wrong time window. Without a time boundary in your prompt, Cowork may pull from too broad a range (months of email instead of the last 7 days). Always state the window explicitly: "from the last 14 days," "since Monday," "from this week's meetings only."

The draft email includes the wrong recipients. Cowork infers recipients from context — which can produce unexpected results if a contact has multiple email addresses or if CC patterns from old threads are picked up. Always review the full recipient list before approving send.

A scheduled prompt fails silently. If a scheduled prompt encounters an error (no transcript found, a Teams channel permission changed), it may produce an empty or partial result without a visible error message. Check the Tasks panel for run history if a scheduled prompt seems to have not executed.

A custom skill is not applied to the task. Skills are matched by description at the start of the conversation. If the skill description is too vague, Cowork may not recognise that it applies. Write skill descriptions as specific outcome statements: "Generates the weekly HSE summary report for a named project from the last 7 days of incident data" rather than "weekly report."

Cowork cannot find the meeting transcript. Transcripts require Teams Meeting Recording to be enabled and the meeting to have ended. If the recording was off, or if the meeting was outside Teams, there is no transcript for Cowork to read. In that case, paste your own notes directly into the prompt.

Prepared by Mathieu Kessler · kesslernity.com
Sources: Microsoft 365 Blog — Cowork is now generally available · Microsoft Learn — Usage-based billing & Copilot Credits · Microsoft Learn — Copilot Cowork overview · Microsoft Learn — Cowork FAQ · Microsoft 365 Blog — Skills & Integrations
Updated August 14, 2026 — Learn documents the Cowork model picker and a separate effort control; Claude Fable 5 requires data retention. June 16 GA wave: Cowork GA worldwide + Copilot Credits usage billing + Work IQ APIs GA. Sources: MS365 blog, Microsoft Learn.