Every Copilot you’ve used so far is a drafting machine. It writes the email; you send it. It proposes the agenda; you schedule the meeting. Cowork breaks that contract: it’s the first Microsoft 365 Copilot that carries out the task — sends the email, creates the documents, books the meeting, searches the org — and shows you its homework in a folder when it’s done.
That’s a genuinely different tool, and it deserves a genuinely different mental model. Here it is.
Acts, not suggests
Standard Copilot Chat is a conversation that ends in text. Cowork is a conversation that ends in outcomes: files in an output folder, invites on calendars, mail in other people’s inboxes. The suggested starters tell you exactly what Microsoft thinks it’s for — Catch me up, Organize my inbox, Organize my week, Prep for a meeting, Plan an event, Prepare for my 1:1, Research a company. Notice the pattern: multi-step chores with a clear deliverable, not one-shot questions.
It also takes serious input — up to 250,000 characters per prompt, plus attached files from your device or cloud and your work context. You can hand it the entire vendor contract and the full email thread, not a summary of either.
Getting access (and the gotcha)
Cowork ships only through the Frontier preview program. Your admin enrolls the tenant — and here’s the trap that burns half the admins who try: if Cowork isn’t visible in the admin center’s Agent management, it’s usually because the admin account itself isn’t enrolled in Frontier (Copilot → Settings → Frontier). Enroll the account, and the management surface appears. If you’re a user waiting on access, send your admin exactly that sentence.
The approval model is the whole product
An agent that sends email on your behalf is either terrifying or liberating, and the approval flow decides which. Before any sensitive action, Cowork asks. Each request carries a risk-level indicator, and — this is the part most people skip — a Show parameters option that reveals the technical details of what’s about to happen: the actual recipients, the actual file, the actual meeting time.
Click it. Every time, at least for your first few weeks. The summary line says “send the update to the project team”; the parameters say which addresses. The gap between those two is where every agentic horror story lives. Reviewing parameters is to Cowork what checking citations is to Copilot Chat: the thirty seconds that separates users who get burned from users who don’t.
Two conveniences exist, and both are earned trust, not defaults: “don’t ask again” (scoped to the current conversation) and Approve All. Use them on task types you’ve watched succeed three times. Don’t use them on anything that leaves your tenant.
The approval gate inverts the usual risk math: because you review actions instead of drafts, you can delegate bigger, vaguer tasks than you ever would to standard Copilot — and still catch mistakes before they’re real.
The side panel: where trust gets built
While Cowork runs, the side panel is your window into its head. Learn its parts:
- Progress — the live step-by-step of what it’s doing right now. Watch this on your first few tasks; it teaches you how Cowork decomposes work, which makes your future prompts better.
- Input folder / Output folder — everything it consumed and everything it produced. You can download all outputs as a zip (up to 50 files / 500 MB). Outputs also land in your OneDrive Cowork folder, so nothing is trapped in the chat.
- Skills — chips showing which skills it pulled into this task. If a custom skill you wrote isn’t lighting up, your skill description isn’t triggering (see the SKILL.md guide).
- Schedule and Permissions — recurring runs and what it’s allowed to touch.
The panel is also your control surface mid-flight: pause soft (finishes the current step, then stops — use this when you want a clean checkpoint), pause hard (stops immediately — use this when you’ve spotted a wrong turn), resume, and cancel. Knowing pause-hard exists is what makes watching the Progress feed relaxing instead of anxious.
Finished and running work lives in the Tasks view — a Recent tab and a Scheduled tab, with each task marked In progress, Needs user input, Done, or Failed. “Needs user input” is the status to check before complaining a task stalled: half the time it’s waiting on an approval you didn’t see.
Scheduled prompts: the quiet superpower
You can tell Cowork, in plain language, to run a task on a schedule — “every weekday at 8, organize my inbox and give me a briefing.” Maximum five scheduled prompts, managed under Tasks → Scheduled. This is where Cowork stops being a tool you use and becomes staff you employ. Start with one: a daily catch-up. It’s the lowest-risk, highest-habit-forming task in the catalog.
The skills system
Cowork’s competence comes in three layers:
- Built-in skills — Word, Excel, PowerPoint, PDF, Email, Scheduling, Calendar Management, Meetings, Daily Briefing, Enterprise Search, Deep Research, Communications, Adaptive Cards. These cover the M365 surface and you get them free.
- Custom skills — markdown files you write. Each lives in your OneDrive at
/Documents/Cowork/skills/<name>/SKILL.md: YAML frontmatter (name,description) plus markdown instructions. They’re auto-discovered at the start of each conversation. Limits: 50 skills, 1 MB per file, 20 companion files / 10 MB per skill. You don’t even need to write the file by hand — ask Cowork to create a skill and it guides you through it in natural language. - Plugin skills — installable from the M365 App Store.
One detail worth knowing: SKILL.md follows the Agent Skills open standard (agentskills.io) — the same format used by Claude Code, GitHub Copilot, Cursor, and others. A procedure you write for Cowork is portable knowledge, not vendor lock-in.
Honest section: it’s a preview, act like it
Cowork is Frontier preview software, and it behaves like it. Rough edges you should expect:
- It’s confidently mid on judgment calls. It executes procedures well; it decides which procedure poorly when your prompt is vague. Vague delegation that works fine in chat (“handle this”) produces weird artifacts here, because the artifacts are real.
- Failed tasks fail quietly-ish. Check the Tasks view; don’t assume silence means success.
- Long multi-step tasks drift. The tenth step of a plan made at step one may no longer fit. Pause soft at natural checkpoints and review.
What not to trust it with yet: anything external-facing without review (customer email, exec comms), anything irreversible (declining meetings, deleting anything), and anything where a plausible-but-wrong output is expensive (numbers in a board deck). The approval gate catches the action; it doesn’t catch a wrong number inside an approved document. Review outputs in the Output folder like you’d review a new hire’s first deliverables — because that’s exactly what they are.
Your first five tasks
Run these in order. They escalate from read-only to real actions, which is how trust should escalate too.
- “Catch me up” — read-only, zero risk, teaches you the Progress panel.
- “Organize my inbox” — watch the approval flow on real actions in your own mailbox; click Show parameters on everything.
- “Prep for a meeting” (pick tomorrow’s hardest one) — output-folder deliverables you can judge against your own prep.
- “Plan an event” — a genuinely multi-step task: invites, documents, scheduling. This is where you learn pause/resume and per-conversation “don’t ask again.”
- Create your first custom skill — ask Cowork to build one for a chore you do weekly, then watch for its chip in the side panel on the next run.
By task five you’ll know, viscerally, what the marketing can’t tell you: where the line sits between what you delegate and what you still do yourself. That line is personal, it moves monthly, and finding it is the actual skill.