Microsoft’s new Copilot has three names for three kinds of work: Home for the starting point, Code for building small solutions, and Autopilot for persistent delegation. The announcement is significant, but it is not a single general-availability launch. Rollout stage and usage-based costs matter before a team puts these features into a working process.
What Microsoft announced—and when you can use it
In its September 25 announcement, Microsoft described Home, Code, and Autopilot as connected parts of a new Copilot experience. Home brings Chat, Cowork, and editable Word, Excel, and PowerPoint work into one starting point. Code builds purpose-made apps, dashboards, and workflows from natural-language instructions. Autopilot is a cloud-hosted agent that can keep working between prompts.
Availability differs. Microsoft says Home and Code will begin rolling out through its Frontier program in the coming weeks. The announcement gives Code an end-of-September Frontier target and says broader availability follows later. Autopilot expands to private preview at the end of September. Copilot Managed Runtime, the hosting foundation for apps made through these workflows, is already in preview. A feature described on a launch page may therefore be absent from your tenant.
Choose the mode by the outcome, not the label.
For a quick answer, summary, or first draft, Chat remains the natural starting point. Cowork is for an end-to-end deliverable that needs iteration, such as a briefing or launch package. Home makes those paths easier to find and adds Office files that remain editable rather than returning only a block of generated text. Microsoft says automatic routing among Chat, Cowork, and Code is coming soon; do not assume the app already chooses every mode for you.
Code targets a different output: a working tool. A team might request an internal tracker for approvals, a dashboard for a dated dataset, or a small app colleagues use. That raises a review question a document draft does not: who owns the app after it is generated? Before sharing it, verify its data connections, access rights, error handling, and behavior when the source changes. Managed Runtime may simplify hosting inside Microsoft 365, but a preview runtime is not a substitute for acceptance tests.
Autopilot shifts the boundary again. Microsoft says it can watch channels, follow up on threads, and resume work over days. It has its own identity, memory, computer, and workspace inside the tenant. That makes the permission design central: specify which channels it may read, which people it may contact, what it can change, and which actions require a person’s approval. A scheduled agent should have an owner and an audit trail, not merely a broad instruction to “keep going.”
The subscription and agent-usage bills are different.
Microsoft separates everyday Copilot subscription use from usage-based billing for long-running agentic work. Its announcement places Cowork, Code, and Autopilot on the usage-based side, alongside certain frontier model choices. It also describes administrator spending policies, credit approvals, model-family controls, and end-user usage history. A business case based only on seat cost will miss the variable cost of delegated tasks.
For a pilot, budget each task, define a completion condition, and track whether the output was actually used. Compare the time saved against credits consumed and human review time. A dashboard that looks finished but takes hours to fix is not a successful low-cost run. Our look at Microsoft Foundry agent traces covers a related need: seeing what an agent did, not just what it produced.
A sensible first pilot
Pick one low-risk workflow with a known answer and a named owner. In Home, test whether an Office deliverable stays editable and correctly reflects its source files. In Code, request a small internal app with sample data and test permissions and refresh behavior. If your organization receives Autopilot access, use a non-sensitive channel and require approval before messages or changes reach other people—record which feature was actually available in your tenant on the test date.
The wider trend is a move from answering questions to carrying out work across files, apps, and time. That is our editorial reading of the release, not a promise of measured productivity. For now, the useful distinction is concrete: Home helps you start and edit, Code creates software-shaped outputs, and Autopilot persists between instructions—each with its own rollout and control boundary.
