02 September 2026

Copilot Cowork: agentic tasks with human approval

Microsoft Copilot Cowork carries out multi-step Microsoft 365 work on your behalf. Approvals, usage-based billing, and Steady controls.

microsoftcopilotcoworkagents

DIGITAL SKILL WORKS LTD / 12 PANES

Cowork is documented on Microsoft Learn as an agentic experience inside Microsoft Copilot: you describe the outcome; it works across Outlook, Calendar, Teams, OneDrive, SharePoint, and related services — with approval before sensitive actions. It is not the same as Copilot Chat.

What it is

According to Learn, Microsoft Copilot Cowork carries out multi-step tasks on your behalf. Work or school Cowork is described as generally available; personal-account Cowork is described as preview. Access sits with Microsoft Copilot and usage-based billing / Copilot credits. Sessions show each step. Users can pause, resume, or cancel. Optional Anthropic models as subprocessors require admin consideration.

Cowork appears as a toggle beside Chat in Microsoft Copilot (browser, desktop, and mobile apps as documented).

What it does

Learn lists capabilities such as:

  • Sending and drafting email
  • Scheduling meetings and managing calendar
  • Creating Word, Excel, PowerPoint, and PDF outputs
  • Posting in Teams
  • Searching the organisation and running deep research
  • Managing files in OneDrive and SharePoint
  • Preparing briefings and stakeholder communications
  • Scheduling prompts and event-driven tasks

Skills (Word, Excel, email, deep research, and more) load as needed. Custom skills and plugins extend behaviour. Memory is conversation-scoped; approval preferences can persist for a session. Cowork is not intended for high-stakes decisions that bypass human review (for example legal filings or medical decisions).

Why you should care

Board. Cowork moves from “assist” to “act”. Governance must match.

CISO. Action approval is a control, not a courtesy. Review plugins, connectors, and model subprocessors.

CFO. Usage-based billing can spike with long autonomous runs. Set budgets and effort levels deliberately.

Adoption. Train on when to use Chat versus Cowork. Chat for think; Cowork for multi-app do.

How it maps to 12 Panes

  • Pane 04 Security: approval, plugins, sensitive directories of work.
  • Pane 05 GRC: register Cowork as an agentic capability; literacy on approval.
  • Pane 06 Cost & Value: credit metering and FinOps owners.
  • Pane 08–09: behaviour change and task redesign.
  • Pane 12 Frontier: fence advanced models and experimental skills.

Suggested actions

  1. Confirm usage-based billing and credit owners before enabling Cowork broadly.
  2. Review Anthropic / AI provider settings with Security and Legal if required in your tenant.
  3. Pilot with a small cohort; require approval for send/post/create; collect failure examples.
  4. Add Cowork to the AI register and incident playbook.
  5. Keep Wave 0 Steady controls ahead of organisation-wide promotion.

Communication needed

Security, GRC, and FinOps first. Adoption and Comms: explain Chat versus Cowork, approval prompts, and how to cancel a runaway session.

Benefits and risks

Benefits. End-to-end task completion across apps with visible steps and approvals. Scheduled recurring work. Extensibility via skills and plugins.

Risks. Credit burn. Over-approval fatigue (“Always allow”). Plugin supply chain. Treating Cowork output as final without review.

Sources

Back to blog