Microsoft unveils the new Copilot: building an “AI operating system” for work

Microsoft has redesigned and introduced an entirely new Copilot, bringing conversations, Office collaboration, app building, and continuously running Agents into a single entry point for work. The direction of this integration suggests that Microsoft wants to build an “AI operating system” for work: you set a goal, and AI uses enterprise data and tools to move the work forward through files, applications, and concrete processes.

The new Copilot’s main capabilities and characteristics fall into six areas:

• Home: ask questions, delegate tasks, and edit files in one place. Chat handles immediate conversations, while Cowork takes on complete tasks. Built-in Word, Excel, and PowerPoint capabilities turn the results directly into real files that people can edit and collaborate on.
• Code: build working tools using natural language. Describe what you need to create a dashboard, tracker, desktop widget, or internal app, then connect it to data and share it with your team through a managed environment.
• Autopilot: let an Agent keep following through. Give it a role, a goal, and boundaries. It monitors progress in the cloud, runs recurring tasks, follows up with people, and can resume moving a project forward days later.
• Enterprise context and plugins: help AI understand the business and use its tools. Microsoft IQ connects enterprise knowledge and data; plugins extend the actions Copilot can take; IT centrally manages permissions and available capabilities.
• Subscription plus usage-based billing: separate everyday use from long-running execution. Subscriptions cover everyday conversations and Office use. Capabilities such as Cowork, Code, and Autopilot consume credits based on usage, with budgets and spending policies available to organizations.
• More proactive personal and team collaboration: prepare the next steps in advance. The upcoming Today experience will organize tasks and prepare drafts and follow-ups. The new @Copilot in Teams will use shared context to help teams understand decisions and coordinate work across departments.

These capabilities are at different stages of rollout, including general availability, early access, and private preview. The sections below explain the problems they address, how they connect, and the conditions to understand before using them.

Home: connect conversations directly to the team’s ongoing work

Home is the starting point for the new Copilot. It lets you review recent activity, pick up where you left off, and see suggestions for work that might be suitable to delegate to Copilot. It contains two different ways of working: Chat and Cowork.

Chat is suited to immediate back-and-forth exchanges. You might look up an answer, organize information, or write a first draft. You ask, it responds, and you remain in the conversation deciding what comes next.

Cowork is suited to delegating a task with a clear deliverable. Microsoft’s examples include responses to requests for proposals (RFPs), product launch materials, customer briefings, and financial close packages. These usually involve gathering information, organizing material, producing an output, and iterating. The product’s aim is for AI to execute the full task and return something you can continue refining, reducing the need to direct every action individually.

A more immediate change is Office in Copilot: Word, Excel, and PowerPoint capabilities come into Copilot, creating or updating editable documents, workbooks, and presentations.

For example, you can ask it to draft a launch briefing, build a budget model, or create a presentation, then keep editing the file alongside the conversation. Existing team files can also be used. When you @mention a colleague, their edits appear in real time, and the file stays synchronized with Office apps. The output of an AI conversation can therefore become part of a file maintained jointly by the team, reducing copying and pasting, passing versions around, and explaining the background again.

Microsoft also described more specific capabilities inside Office apps. PowerPoint can maintain brand consistency, adjust slides and images, and apply custom templates. Excel will explain what Copilot or a collaborator changed and why, and recommend appropriate charts. Skills bring specialist working methods into apps, such as financial skills for Excel and legal skills for Word. The aim is to turn methods used repeatedly in a particular role into reusable workflows that a team can follow.

For now, it remains necessary to understand the distinction between Chat and Cowork. Microsoft says that, in the future, users will simply describe what they want to accomplish, and Copilot will choose the appropriate capability across Chat, Cowork, and Code. Automatic delegation is a future direction; it should not be treated as an existing feature that already covers every request.

Code: deliver a usable tool, beyond a file

Documents are useful for explaining things, spreadsheets for organizing and calculating data, and slides for presenting to others. Some work, however, calls for a tool that can be used repeatedly.

Microsoft positions Code as a way for ordinary business users to describe an app, tracker, dashboard, automation, or workflow in natural language, with Copilot choosing how to implement and build it. The result could be a widget that stays on the desktop, an interactive dashboard for exploring data and comparing outcomes, or an internal cloud app shared with colleagues.

The official interface shows a concrete request: “Build a pipeline dashboard by region and stage.” A sales pipeline tracks potential deals as they move from initial contact to a completed sale. Compared with a static chart, the value of such a tool is that the team can keep examining the business by region and stage. When a business question needs to be answered repeatedly, the tool can become a lasting interface for that work.

Microsoft calls this a “fourth” type of output for knowledge work, alongside documents, spreadsheets, and presentations: small solutions built for specific problems. That is also what makes the capability worth watching. Business users may not have to fit every requirement into the fixed features of existing software; they could turn some repeated work into tools of their own. This is a product direction. Whether a particular tool can operate reliably still depends on data connections, implementation quality, and its running environment.

Code uses the same underlying technology as GitHub Copilot. Microsoft continues to position GitHub Copilot as the everyday tool for professional developers, while strengthening its connections to the Copilot platform. The two serve different users and entry points for work.

Once it is built, where does it run?

This step determines the distance between “generated code” and “something the team can actually use.” A program produced in a chat window still needs a running environment, data connections, access controls, and a way to share it.

The newly introduced Copilot Managed Runtime is intended to handle these needs. A runtime is the environment in which a program actually executes; “managed” means the platform operates that underlying infrastructure. Microsoft says it allows apps to run within the company’s Microsoft 365 environment under IT management, while users can share them with colleagues, connect live data, and access them from different locations.

Code runs in a sandbox and can also be hosted within the enterprise tenant. A sandbox is an isolated environment that limits what a program can do. A tenant is an organization’s separately administered space in Microsoft’s cloud services. Microsoft IQ provides work context for building, and plugins are synchronized as well. The idea is that business users describe their needs, AI builds the tools, and the company can still manage where those tools run and which data they can access.

Managed Runtime also supports apps built through Cowork, Code, and Copilot Studio, and is open to third parties and professional developers. It is currently in preview. This launch is therefore also adding a shared foundation for running apps, beyond a new input box for writing code.

Autopilot: remember a task and keep moving it forward

While Cowork focuses on delivering a delegated task, Autopilot emphasizes continually fulfilling a role. Previously called Scout, it works, in Microsoft’s description, by receiving a name, a role, and a goal. It can then keep monitoring channels, following discussions, running recurring tasks, and even resuming work on a project days later, without waiting for a new prompt each time.

A vendor review is the announcement’s end-to-end example. A review requires scheduling, working backward to plan preparation, organizing meetings, following up on action items, and asking relevant people for updates. Microsoft wants Autopilot to take on this whole process, extending beyond generating a vendor review template.

This also explains why it needs cloud hosting, memory, and its own workspace: the work spans multiple conversations, applications, and points in time. The process must continue after the person leaves their computer. When they return days later, it needs to know how far the work has progressed.

According to Microsoft, Autopilot has its own identity, memory, computer, and workspace within the enterprise tenant. Microsoft IQ grounds it in business context, and it appears in the places teams already use: Teams, Outlook, chats, channels, and documents. You can @mention it much as you would a colleague.

Having “its own identity” is particularly important. An Agent that keeps taking action, accessing information, and following up with people needs to receive permissions, and the organization needs to track what it has done. Microsoft includes permissions, auditing, and governance in the capability, with users setting goals and boundaries.

Users also need to be clear about both “what to accomplish” and “what actions are allowed” when delegating. For a vendor review, for instance, whether the task is limited to organizing material or also permits contacting people and advancing arrangements directly changes the Agent’s responsibilities. Such working agreements are better suited to long-running tasks than a vague “handle this for me.”

Autopilot remains in an expanded private preview. The capabilities described above are those Microsoft has introduced and planned; they do not establish that all subscribers can already hand an entire vendor review over to it.

The differences become clearer when the four ways of working are considered together:

• Chat: returns an answer, explanation, or first draft for the current question and continues exchanging messages with you.
• Cowork: returns the completed result of a defined task, such as a set of customer briefing materials.
• Code: returns a tool that runs and can be reused, such as a business dashboard.
• Autopilot: keeps advancing work toward a goal, such as review follow-ups spanning several days.

Home is the starting point for this work, rather than another model on the same level as Chat. Auto in model selection is also different from Autopilot: Auto chooses a model for a request; Autopilot is an Agent that keeps executing work.

How does it understand the company’s business?

Connecting these capabilities in practice requires another problem to be addressed: AI’s ability to write something does not mean it understands your company’s customers, metrics, and business processes.

Microsoft calls this layer of enterprise background Microsoft IQ. It can be understood as a contextual foundation connecting enterprise knowledge and business operations, allowing Copilot and other Agents to use organization-specific information while remaining subject to enterprise permissions and governance.

Within this layer, Fabric IQ brings the context of enterprise data into Copilot. The announcement mentions more than 20 million semantic models in Power BI. These are not 20 million large language models. They are data models that organize business data, metrics, and relationships—for example, how tables connect and how a metric is defined. The number describes the scale of those data models; it does not mean that every company automatically gains access to all their data.

According to the announcement, Fabric IQ support for Chat and Cowork is generally available, with Code integration to follow through the Frontier program.

Data and workflows from Dynamics 365 and Power Platform will also be connected, with public previews rolling out over the month following the announcement. Microsoft gives the example of a salesperson writing a proposal: Copilot could bring deal history and support tickets directly into the document. The proposal can then draw on the customer’s previous business interactions, reducing the need to search across applications and move information between them.

Beyond data, there is the question of which actions it can take. Plugins combine Skills, connectors, and other capabilities to extend the workflows Copilot can complete. A new plugin registry will put plugins from Microsoft, partners, and the organization itself into a unified catalog. IT can approve and manage them centrally, while developers publish once to extend across multiple supported experiences.

Along the work process, these layers have distinct roles: Microsoft IQ supplies business context, plugins supply connections and actions, Managed Runtime runs applications, and enterprise permissions and governance constrain the whole process. A stronger model alone cannot provide all of these working conditions.

Why does AI that keeps working need a different pricing model?

Microsoft separates Copilot costs into two parts.

The first is a user subscription license (USL), billed per user per month, covering Chat and everyday Copilot use in Word, Excel, PowerPoint, Outlook, and Teams. The second is usage-based billing (UBB), which consumes Copilot Credits for long-running capabilities such as Cowork, Code, and Autopilot, as well as frontier models such as Astra and Fable. The official pricing explanation makes clear that these UBB capabilities still require a USL: users need a subscription license first.

This means that “I bought Copilot” and “I can run background Agents without limits” are separate questions to check.

Why is Auto suited to everyday tasks?

Microsoft offers one explanation in its pricing discussion: some tasks have a performance “saturation point.” Turning meeting notes into key points, for example, may not require the strongest model. Once a more powerful model no longer changes the result materially, additional computing has limited value.

Auto selects the model and reasoning effort for each request according to accuracy, speed, and cost requirements. Reasoning effort can be understood as how much computation the model devotes to processing a problem before answering. Microsoft uses this routing to support a fixed subscription price for everyday use and reduce the burden of choosing a model each time.

Users can also select a model manually. The pricing article lists GPT-5.6 and Sonnet 5, and says newer models such as Opus 5 will be included with limits. Subscriptions remain subject to fair-use policies and applicable allowances. Heavy users receive a warning as they approach a limit and can switch back to Auto at no additional charge or choose to use Credits. A fixed subscription does not mean unlimited access to every model at every level of reasoning effort.

Long-running Agents are different. A task lasting several hours may repeatedly process context, call tools, and run programs. Its cost is affected by the model, the volume of material, the tools, and the running time. Microsoft therefore uses UBB for this work, while retaining room for users to choose models and spending levels according to the task’s value.

What does the official cost chart tell us?

Microsoft also provides a cost comparison chart to explain why it favors a combination of “subscription for everyday work plus usage-based billing for advanced tasks.”

For the everyday user’s mix of 15 tasks, the chart puts Copilot Auto at 30 dollars, Claude Opus 5 at 84 dollars, and ChatGPT GPT-5.6 Sol at 38 dollars.

The advanced user’s mix consists of 20 everyday tasks plus 5 complex, long-running tasks. Copilot with Cowork using Opus 5 is shown at 73 dollars, compared with 149 dollars for Claude with Cowork. Copilot with Cowork using GPT-5.6 Sol is shown at 69 dollars, compared with 80 dollars for ChatGPT with Work.

The methodology is central to interpreting the chart. Microsoft’s internal analysis used two sets totaling 40 prompts, with each tested 10 times. Copilot’s everyday tasks used Auto, while competitors used the models listed in the chart; reasoning settings also differed for some app tasks. Copilot costs were calculated using the subscription list price and Cowork variable rates. Competitor costs were modeled using public list prices and API prices, excluding discounts, promotions, and negotiated pricing.

The chart therefore supports Microsoft’s cost claim under a particular task mix and set of pricing assumptions. It does not establish that every user switching to Copilot will obtain the same savings, nor does it provide results sufficient to show that output quality was identical across all three providers. Differences between the bars also reflect model routing and pricing methods, so totals alone cannot establish which model is cheaper.

A more useful comparison for an organization is to examine total spending and usable results for its actual tasks. If everyday summaries account for a large share, subscription coverage matters. If long-running tasks account for more, both the credits consumed per execution and whether the returned work is worth using again need to be considered.

How is billing enabled, and who controls it?

The pricing explanation states that enterprise UBB services are off by default. Billing does not begin until an administrator creates a spending policy in the Microsoft 365 admin center. Seeing a capability in the interface does not mean the organization has authorized employees to use and pay for it.

Administrators can set tenant and user-group budgets, individual limits within group policies, threshold notifications, and requests for additional credits. Spending can also be assigned to the relevant Azure subscription or resource group. Reports break it down by user, group, workload, policy, and payment method, so costs can be linked to the responsible team.

Microsoft calls this ongoing management of AI spending and business value FinOps for AI. This expansion includes cost management for Code and Managed Runtime, APIs for managing spending policies in bulk, integration of credit requests with existing approval processes, and visibility into which Cowork scenarios consumed credits and what they produced. Administrators can also restrict model families by user group, which constrains Auto’s choices as well. The detailed pricing article says this model management will launch in Cowork first and then expand to other workloads.

Users can see their own credits used, remaining allowance, and history in Copilot. As AI does more work, the way that work consumes budgets needs to become visible too.

What is available, and what is still in preview?

These updates are being introduced in stages. According to the launch announcement:

• Home and Code: rolling out through the Frontier early-access program. Code is scheduled to enter Frontier at the end of the announcement month, with broader availability in the following weeks. A preview for Microsoft 365 Premium and Pro subscribers is planned for later that year.
• Autopilot: expanding its private preview at the end of the announcement month, rather than opening to everyone.
• Copilot Managed Runtime: already in preview.
• Fabric IQ: generally available in Chat and Cowork, with Code integration to follow through Frontier.
• Dynamics 365 and Power Platform context: public previews rolling out over the month following the announcement. The plugin registry is beginning its rollout and is planned to become generally available across supported entry points in the following weeks.

Microsoft also previewed two more proactive work experiences.

The first is called Today. It will appear in Home, bringing together email, calendars, Teams discussions, meetings, and tasks to show what you missed, what needs attention now, and what can wait. It will also prepare work in advance: drafting text for you to edit, suggesting schedule changes, and preparing context-rich follow-ups for you to send. This approach of preparing the work before handing it back to a person can also reduce the time spent finding what to do next.

Today is planned to enter private preview in Copilot in the month after the announcement, with Outlook and Teams to follow.

The second is @Copilot in Teams, which helps a whole team using the shared context and permissions of a channel, group chat, or meeting. If someone asks why a product decision was made, it can bring together the discussion behind it. When a decision affects manufacturing, packaging, and launch timing, it can also propose preparing a cross-functional review that brings dependencies, owners, open questions, and decisions together. This capability is likewise planned to enter private preview at the end of the announcement month.

Considered alongside Autopilot, these two experiences reveal Microsoft’s direction: AI is participating earlier in organizing work, helping people decide what to handle first while beginning to take on subsequent execution and collaboration.

For everyday users, the question to watch is whether results can flow directly into real files and team processes. For business teams, it is whether repeated operations can become reusable tools. For enterprise IT, it is whether these continuously running apps and Agents can operate within clear permissions and budgets. The new Copilot brings those questions into one product system. The real test will come through sustained use in specific business settings.