On September 25, 2026, Microsoft announced a redesign of its enterprise AI application Copilot around three pillars: Home, Code, and Autopilot. Home creates actual Word, Excel, and PowerPoint files alongside conversations; Code assembles small business applications from natural language; and Autopilot continues working even after the user has stepped away. What's grown here isn't simply the number of features—it's the range of deliverables Copilot now handles. The product is moving from an AI that returns answers and documents toward a business platform that runs code and takes on extended work.
However, these three features have not all reached general availability for the same audience simultaneously. Code is rolling out in stages starting from an early access program, Microsoft Copilot Managed Runtime (hereafter, Managed Runtime)—which runs the code—is in public preview, and Autopilot is in an expanding private preview. This is more accurately understood as an announcement of the direction Microsoft intends to take Copilot, rather than a finished, monolithic product.
From "responses" to "working deliverables"
Home places Chat, which answers questions, and Cowork, which delegates substantial tasks, at the same entry point. According to Microsoft's explanation, Chat handles short tasks like searches and drafts, while Cowork carries multi-step work through to completion—responding to RFPs, assembling a full set of product-launch materials, or preparing financial closing documents. There's also a future vision in which users won't need to select a mode at all; the system would route work to Chat, Cowork, or Code based on content. But this is not a current feature.
What stands out in Home is "Office in Copilot." What Copilot creates here isn't a fragment of text displayed in a chat window—it's an editable Word document, Excel workbook, or PowerPoint presentation. Users refine content alongside the conversation, bring in colleagues, and changes made within the Office apps themselves stay in sync. In other words, this reduces the friction of copying an AI's response into a separate file, redesigning the workflow so the output is treated as a shared artifact from the start.
This direction extends Microsoft 365's longstanding treatment of the file as the fundamental unit of knowledge work. Code, however, expands that unit. When a user describes what they want in text, Copilot builds desktop widgets, dashboards for exploring data, internal apps, automations, and workflows. Microsoft distinguishes this from GitHub Copilot, which developers use daily, positioning Code instead as an entry point where non-developers, too, can build small tools for their own work.
Three entry points, three different deliverables and operating models
Laying out Microsoft's official materials as of September 25, 2026 across deliverables, persistence, governance, billing, and rollout stage reveals the distinctions. This isn't a comparison table of hardware performance or pricing—it's an organization of the roles and conditions Microsoft has announced.
| Entry Point / Platform | Primary Deliverable | Execution & Persistence | Governance & Billing | Status as of Sept. 25, 2026 |
|---|---|---|---|---|
| Home / Office in Copilot | Synced Word documents, Excel workbooks, PowerPoint presentations | Created and updated alongside conversation; syncs with Office apps and co-editors | Day-to-day Chat and Office work is covered by the base user license | Staged rollout from the Frontier program |
| Code | Widgets, dashboards, internal apps, automations, workflows | Generated in a sandbox; can be deployed onto Managed Runtime | Code is metered/usage-based; runs under admin-defined policy | Rolling out to Frontier late September; broader availability within weeks. Individual Premium/Pro preview planned within the year |
| Managed Runtime | Runtime environment for apps built via Code and other tools | Hosted within the Microsoft 365 tenant boundary; ongoing deployment, updates, and monitoring | Centralized management via Entra ID, connection policies, audit, and inventory | Host and SDK in public preview |
| Autopilot | Agents with their own identity, memory, compute, and workspace | Runs persistently in the cloud; supports recurring tasks and resuming work days later | Extended work is metered/usage-based; subject to permissions, audit, and governance | Expanding private preview as of late September |
Home's deliverable is a collaboratively editable Office file; Code's is a small app running inside the tenant; Autopilot's is a continuously running agent with its own identity and memory. Everyday work falls under the standard user license, while Code and Autopilot are subject to usage-based billing. Home and Code are rolling out in stages starting from Frontier, with Code's broader availability following within a few weeks; Managed Runtime is in public preview; and Autopilot is in an expanding private preview.
Ignoring these distinctions risks the mistaken impression that Copilot's new interface has made every feature immediately usable. What Microsoft has actually presented is a shared entry point and an operating philosophy—eligibility, billing, and product maturity all need to be verified separately.
Code's real substance is the operating infrastructure, not the generative model
Even if an app can be generated from natural language, enterprise use requires further work downstream: provisioning cloud resources, authenticating users, and restricting which data and external services the app can connect to. It's also essential to be able to run a new version alongside the previous one before switching over, and to trace problems if they arise. What determines Code's success isn't whether it can generate code—it's who takes on the burden of everything that comes after generation.
Microsoft's answer is Managed Runtime. The company explains that it hosts generated code within the Microsoft 365 tenant boundary, unifying identity management through Microsoft Entra, organization-defined policies for connections and data, Git-based source control, and deployment and version management. Apps built through Cowork, Code, and Copilot Studio all run on this same platform, which is also open to compatible third-party tools and professional developers.
The SDK and command-line tools aren't merely for packaging a finished app. They scaffold new projects, define necessary data connections, and generate typed services in TypeScript for handling connectors. Development-time execution and preview, through deployment and updates, all flow through the same path—and the code itself remains editable and stored in Git. The design intent is that when a developer picks up a prototype built by a non-developer, they don't have to rebuild the runtime infrastructure from scratch.
For administrators, an Apps view is provided within the Microsoft 365 admin center. It reportedly lets admins see at a glance who has access, how much an app is being used, whether it's running properly, and which policies apply—along with the ability to enable or disable apps. That said, both the host and SDK remain in public preview. Since this material doesn't include independent verification of safety, availability, service guarantees, or the quality of generated apps, Microsoft's design claims and real-world production performance should be evaluated separately.
Persistent agents come bundled with permission and cost controls
Autopilot elevates what was previously called "Scout"—a persistent agent—into a core Copilot feature. Once a user gives it a name, role, and goal, it watches channels, follows threads, carries out recurring tasks, and returns to the same matter days later. Microsoft describes it as having its own identity, memory, compute, and workspace, callable like a colleague within Teams, Outlook, and documents.
The June announcement of its predecessor, Scout, revealed how Microsoft intends to govern persistent agents. It's built on the open-source OpenClaw, and rather than using a shared service account, each agent is assigned a managed Entra ID. Access and credentials are scoped narrowly to the task at hand, and sensitive operations can require human approval. Microsoft Purview's sensitivity labels and data loss prevention policies are also reportedly applied before any send or write action.
However, as of June, Scout was limited to select customers and Frontier organizations, requiring Intune configuration and explicit enrollment procedures. These conditions shouldn't be treated as the final specification for the September version of Autopilot. Still, the underlying design philosophy is consistent: if an agent is going to act on a person's behalf, there needs to be a mechanism that can record who authorized what.
The cost structure is also changing. Microsoft includes questions in Chat and drafting, summarizing, and analysis in Office within the standard user license, while splitting off Cowork, Code, and Autopilot's extended tasks and advanced models into usage-based billing. Once work extends beyond conversation time, execution volume—not seat count—determines cost. Administrators can set budgets, usage caps, and which model families are available, while users can check their remaining balance and usage history. FinOps isn't an afterthought billing screen here—it's the control plane needed to bring persistent agents into an enterprise.
Adoption decisions should start with "what stage is this at," not "does it work"
The announcement mixes general availability, staged rollout, public preview, private preview, and future plans. Data context for Chat and Cowork via Fabric IQ has reached general availability, but integration with Code is planned to roll out first through Frontier. Features that draw on Dynamics 365 and Power Platform data and business processes are entering public preview over the month following the announcement. A unified plugin registry is also rolling out, with general availability across supporting surfaces expected within a few weeks.
Furthermore, "Today"—a feature that surfaces priorities from email, calendar, Teams, meetings, and tasks—is planned to enter private preview in October 2026, and the feature that automatically selects among Chat, Cowork, and Code remains a future plan. Announced concepts shouldn't be mistaken for features immediately usable under current contracts. Regional and language-specific availability, exact usage-based pricing, and Copilot Credits consumption rates also can't be confirmed from the materials obtained for this piece.
What enterprises should verify first isn't whether a demo could build an app—it's which contracts and preview stages apply, what data connections are permitted, and what the procedures are for stopping, updating, and auditing. Beyond that, once broader availability arrives, organizations need to measure the quality of generated apps and the visibility of the admin console, and verify which operations require human approval under Autopilot and what extended tasks actually cost. Once that combination of numbers and operational track record is in place, Copilot can move from being a document-drafting assistant to a platform capable of managing small internal software and ongoing business operations.