Copilot Cowork: Delegating Multi-Step Work in Microsoft 365
The first phase of enterprise AI was session-based: the user asks, the system answers, and the user carries the work forward. Copilot Cowork takes over multi-step work. It builds a plan from a desired outcome, uses email, meetings, files and data from Microsoft 365, works in the background and asks for approval at checkpoints. Admins control billing, plugins, browser use and audit logs. In this article, I explain why this execution layer changes how organizations should measure AI value.
With Copilot Cowork, Microsoft moves its AI offering from answering questions to delegating work that runs over longer periods across Microsoft 365.
The first phase of enterprise AI was session-based. The user asked for a summary or a draft, the system responded, and the user carried the work forward. Cowork handles long-running, multi-step work: it creates a plan, grounds itself in work context, progresses in the background and pauses at checkpoints so the user can review, redirect or approve.
From assistance to delegation
A conventional assistant speeds up individual tasks. Cowork carries a task through to completion. The user describes the desired outcome, and Cowork turns it into a plan. It draws on email, meetings, messages, files and data, asks when something is unclear and requests approval before it applies changes.
This matches how knowledge work runs. Hardly any task is done in one step. Most involve several documents, applications and people, and priorities change along the way. The value lies in keeping the work moving.
Grounding in Microsoft 365
Cowork runs on Work IQ, Microsoft's intelligence layer for work context. It uses Outlook schedules and email threads, Teams conversations, files, spreadsheets and connected plugins. Microsoft's examples show why this matters: calendar cleanup, meeting packet preparation, company research and launch planning. Preparing a meeting, for instance, involves previous emails, calendar availability, an analysis in Excel and a follow-up message afterwards.
The plan as the core of the product
Cowork makes its execution path visible. Users see what the system intends to do, which information it uses, where it stands and when it needs approval. They can intervene without starting over. Organizations need autonomy they can steer, inspect and interrupt. Many AI tools generate quickly but leave the path to the result opaque, and Cowork makes planning and approval part of the product.
Governance built in
The Microsoft Learn guidance lists what admins control: usage-based billing, discoverability for users, plugins, model availability, browser policies, audit logs and approval patterns. Browser tasks inherit the organization's Microsoft Edge policies. Automated tasks run with the user's permissions. Shared actions such as sending messages or changing shared systems require approval by default.
Enterprises therefore do not have to trust a more capable AI blindly. They define how far the capability reaches, monitor it and control the cost.
Microsoft's advantage
Many AI products generate text or summarize documents. An execution environment inside the daily working tools is harder to replicate. Microsoft brings the applications, the identity and permission layer, the compliance framework, the business context and the extensibility model. Cowork combines these assets into a layer for delegated work inside the enterprise boundary.
Consequences for measuring value
If this model matures, organizations will measure AI value differently. The number of prompts or drafts matters less than delegated tasks, shorter cycles across teams, less preparation time for recurring tasks and fewer manual handoffs between applications. This view of ROI is more operational and more realistic. Enterprises need AI that works reliably in the flow of work without creating new governance problems.
Chat remains, but it becomes the starting point rather than the endpoint. Enterprise AI will increasingly be judged by whether a system takes an intent, builds a plan, uses the right context, keeps the user informed and completes the work responsibly.