Copilot Cowork: Microsoft Shifts from AI Assistant to Agent That Executes Work

Copilot Cowork: Microsoft Shifts from AI Assistant to Agent That Executes Work

Since the launch of Microsoft 365 Copilot, enterprise AI usage has largely followed a simple logic: the user asks a question, and Copilot answers.

Summarize a Teams meeting. Retrieve information from a document. Draft an email. Analyze data. Create the first draft of a presentation.

With Copilot Cowork, Microsoft aims to go further.

The difference can be summed up in one sentence: users no longer just ask Copilot to help with a task, they can delegate an entire mission to it.

Copilot Cowork can turn a request into a plan, traverse multiple sources, use different tools, execute several steps, and deliver outputs. Users can monitor its progress, provide new instructions, or halt execution.

For CIOs, this shift is significant. Microsoft 365 is no longer just a conversational assistant, it is becoming an environment where AI agents can truly take action.

What is Copilot Cowork?

Copilot Cowork is an agentic system integrated into Microsoft 365 Copilot, designed to execute long, complex tasks requiring multiple steps.

Microsoft describes its operation as follows: the user defines the expected outcome, Cowork leverages available work context, builds a plan, executes the necessary actions, and delivers the result.

It can work with emails, meetings, files, and data accessible to the user via Work IQ. It can also create documents, manage files in OneDrive or SharePoint, draft Teams messages, schedule meetings, or take action in Outlook.

This is what sets Copilot Cowork apart from a simple chatbot.

In a classic chatbot, the loop looks like this:

User → question → answer.

With Cowork, it becomes more like:

User → goal → plan → search → actions → checks → deliverables.

The change may seem subtle. In practice, it fundamentally transforms how a company can use anMicrosoft AI agent.

Copilot Chat answers, Copilot Cowork executes

Let’s take a simple example.

A CIO might ask Copilot Chat:

"Summarize the main topics discussed in this week’s ERP project meetings."

Copilot searches for relevant information and provides a summary.

Now, imagine a different request:

"Analyze the ERP project’s progress based on Teams meetings, reports, and available files. Identify blocking issues, prepare a report for the executive committee, create a summary presentation, and draft an email to the relevant managers asking for updates on delayed items."

This request involves multiple operations.

It requires searching for information, understanding context, cross-referencing multiple sources, identifying issues, creating deliverables, and preparing actions.

This is precisely the type of work Copilot Cowork is designed for. Microsoft states that Cowork can handle complex, time-consuming tasks across multiple tools, then deliver a final result rather than a simple recommendation.

Microsoft clearly distinguishes between use cases. Copilot Chat is suited for quick questions and exchanges. A specialized agent is for repetitive, well-defined tasks. Cowork targets missions that span multiple steps and applications.

For businesses, this distinction is critical.

The focus is no longer just on content generation. It’s now on work execution.

Copilot Cowork starts by building a plan

One of the key features of Microsoft Cowork is its ability to break down a task.

When a user delegates work to it, the agent doesn’t just try to produce an immediate answer. It turns the request into distinct steps.

This plan remains visible.

Users can track progress, provide additional guidance, adjust the direction, or halt execution. Cowork can also request clarification if information is missing.

This approach brings the experience closer to that of a colleague entrusted with an objective rather than a basic instruction.

For a CIO, this opens up far more compelling scenarios than simple assisted writing.

A Microsoft 365 autonomous agent could, for example, be tasked with preparing a monthly review, gathering the necessary elements from multiple spaces, analyzing available data, and generating the corresponding materials.

The user then primarily intervenes when a decision or validation is required.

Work IQ provides the agent with the company’s context

An agent capable of reasoning is only truly useful if it has the right context.

This is where Work IQ.

Microsoft presents Work IQ as the engine enabling Cowork to leverage the user’s professional context: emails, meetings, messages, files, and data within Microsoft 365.

This changes a great deal.

A general-purpose AI can understand a request, but it may not know the company’s clients, project files, decisions made in recent meetings, or exchanges with a supplier.

Copilot Cowork can use the information the user already has access to in their Microsoft 365 environment.

For a CIO, this integration is likely one of the main advantages of Microsoft’s approach. The agent is not added on top of the information system as a standalone tool, it works directly with the applications and data already used by teams.

A Microsoft AI agent capable of using multiple tools

Cowork is not confined to a single application.

It can operate across multiple Microsoft 365 environments and leverage additional integrations.

Microsoft highlights Outlook, Teams, OneDrive, and SharePoint, as well as Dynamics 365, Fabric, and plugins that extend the agent’s capabilities.

Organizations can also develop their own plugins to connect Copilot Cowork to their internal systems and processes.

This extensibility is strategic.

In most companies, a process never lives within a single application.

A sales review may require CRM data, Excel files, emails, and a PowerPoint presentation.

Project tracking can involve Teams, SharePoint, internal documents, and data from a business tool.

An incident may require collecting information, analyzing it, producing a report, and then contacting multiple people.

It’s in this type of workflow that the Microsoft AI agent becomes valuable.

Copilot Cowork can also use the browser

Microsoft is also adding a particularly significant capability: browser usage.

As part of the Microsoft Copilot Frontier program, Cowork can use a local Edge browser to access the web while respecting the company policies already applied to the user.

Browser access significantly expands an agent’s scope of action.

A business process isn’t always limited to Microsoft 365 data. Some information may be accessible via a web app, a supplier portal, or an external source.

The agent can thus gradually shift from a Microsoft 365-centric environment to a broader digital execution logic.

For CIOs, however, this capability warrants special attention. The more tools an agent has, the more critical questions of permissions, traceability, and control become.

The admin documentation gives concrete answers here. Because the browser tab runs on the user's own machine, it inherits the policies already in place: conditional access, DLP and the tenant's Edge browsing rules. A "Cowork Browsing" tenant setting turns the capability on or off for the whole organisation, Edge allowlists and blocklists still apply, and every browser task Cowork starts on a user's behalf is recorded in the unified audit log. Reference: managing Copilot Cowork for your organisation.

Microsoft is also betting on a multi-model architecture

Another key development: Microsoft doesn’t want to lock Copilot into a single AI model.

Microsoft 365 Copilot now adopts a multi-model approach. Cowork can use the model deemed most suitable for the task, while some Frontier experiences also allow model selection.

That multi-model approach is more concrete than it sounds: Cowork ships with several Anthropic Claude models, Opus and Sonnet variants, a Sonnet+Opus pairing, alongside GPT 5.5 and an image generation model. Two points matter directly to an IT department. First, the Anthropic model family can be turned off from Copilot settings in the admin centre. Second, some preview models require data retention on the provider's side: the user's prompts and responses are retained there. Choosing a model is therefore not only a quality or cost trade-off, it is a confidentiality one. Documentation: managing Cowork's models.

This approach addresses a practical issue.

Not all tasks require the same level of reasoning.

Summarizing a document, analyzing thousands of files, or managing a complex workflow don’t demand the same resources.

Model selection can thus become both a functional and economic lever.

Microsoft also bills Cowork based on usage via Copilot Credits. Costs depend notably on the models used, context retrieval, tool calls, and execution time.

For an IT department, AI governance will therefore no longer focus solely on the number of licenses. It will also need to track the cost of tasks assigned to agents.

That point actually gates access itself: to open Cowork to users, an admin must first enable usage-based billing. Without it the experience may stay discoverable, but users can only request access, leaving the admin to arbitrate on policy, cost and compliance. Copilot Credits are paid as consumed, $0.01 per credit on pay-as-you-go, with a discounted volume-commitment option, and are consumed by model responses, tool calls, image generation and browser tasks. Limits can be set per user or per group, and a cost estimator is provided. Documentation: usage-based billing and Copilot Credits.

The CIO’s role becomes even more central

With an AI assistant, the main risk often concerned accessible data and generated responses.

With a Microsoft 365 autonomous agent, a new dimension emerges: action.

An agent can prepare a document, modify a file, send a message, or intervene in a process.

Microsoft has therefore implemented control points. Cowork can request user approval before sensitive actions, such as sending an email or making significant changes. Its actions are also integrated with Microsoft 365 governance mechanisms.

Automated tasks deserve attention, because that is where the risk shifts: a user can create scheduled tasks, or tasks triggered by an incoming email or Teams message, that run with nobody at the screen. Microsoft applies three safeguards. Each task runs with its creator's permissions and therefore sees only that person's data. Any shared action, sending an email, posting a message, changing a shared system, is prepared and then submitted for approval, unless pre-authorised, and that pre-authorisation only holds for the current session. Finally, rate limits and loop protection stop a task from re-triggering itself. Detail: the safeguards on automated tasks.

At general availability, Microsoft announced Cowork’s integration with audit functions, eDiscovery, Insider Risk Management, Data Security Posture Management, and various compliance policies. Sensitivity labels can also be preserved on generated content.

For the CIO, deployment cannot therefore be limited to enabling a new Copilot feature.

They must determine which users can access Cowork, which data is properly protected, which use cases are acceptable, which actions require validation, and how to monitor consumption.

Microsoft also allows administrators to define access to Cowork and set spending limits at the tenant, group, or user level.

From Microsoft Copilot Frontier to production use

It is also important to distinguish Cowork from the Microsoft Copilot Frontier.

Frontier provides early access to certain experimental Microsoft 365 Copilot features. Administrators can control which users have access via the Microsoft 365 admin center.

Copilot Cowork itself began in Frontier in March 2026. But Microsoft announced its global general availability on 16 June 2026 for Microsoft 365 Copilot customers.

Some associated features, however, continue to be tested via Frontier, including Cowork’s use of Edge and certain model selection mechanisms.

This distinction is important for an IT department preparing a roadmap. Not all capabilities presented around Cowork necessarily have the same level of maturity or the same access conditions.

What Copilot Cowork truly changes

The challenge of Copilot Cowork is not ultimately about getting better answers in a chat window.

It’s about changing the unit of work entrusted to AI.

Yesterday, we asked:

"Write this email."

"Summarize this file."

"Analyze this data."

Tomorrow, the request could become:

"Prepare the monthly review for this activity and give me the deliverables needed for Friday’s meeting."

The user describes the expected outcome in greater detail. The agent then chooses part of the path to achieve it.

It’s this shift from command to delegation that makes Microsoft Cowork particularly compelling.

And it explains why CIOs must look beyond the novelty effect surrounding Copilot.

Preparing your IT systems for agents that truly work

The arrival of Copilot Cowork marks a new stage in Microsoft’s AI strategy.

Microsoft 365 is gradually becoming an environment where multiple types of agents can search, reason, produce, and act.

For CIOs, the real question is no longer just: "Should we deploy Copilot?"

It becomes: what tasks do we truly want to entrust to AI agents, with what data, what permissions, and what controls?

The companies that derive the most value from this shift will likely not be those that enable the most features. They will be the ones that identify the right processes, clearly define responsibilities between users and agents, and then build appropriate governance.

It’s precisely on this front that the agency Scroll helps businesses: define AI use cases, identify processes worth automating, and design agents tailored to existing tools. With the arrival of Copilot Cowork and autonomous agents in Microsoft 365, this framing work is now as critical as the technology itself.