Microsoft announced Copilot Home, Code, and Autopilot on September 25, 2026, bringing conversational work, app creation, and persistent agents into a more unified experience. These capabilities have different rollout schedules: the announcement describes Frontier access for Home and Code and an expanding private preview for Autopilot.
Three work modes, with different operating commitments
According to Microsoft’s announcement, Home combines Chat, delegated Cowork tasks, and editable Office files. Code builds small applications in a sandboxed environment. Autopilot is a cloud agent with its own identity, memory, computer, and workspace. The distinction matters because answering a question, generating an application, and operating a recurring process create different responsibilities.
A team evaluating Home can begin with document workflows and inspect the output directly. A team evaluating Code needs a software owner who can review the generated application’s inputs, storage, sharing, and maintenance. An Autopilot trial needs a process owner who can specify when work should start, what completion means, and which actions require approval. The same successful demo cannot establish all three.
Tenant context is valuable only when its ownership is clear
Microsoft’s separate Fabric and Azure update identifies governed business context as part of this direction. For organizations, the useful opportunity is to connect work with agreed business definitions rather than another isolated copy of company information.
Before granting access, identify who owns each connected dataset and how permissions change when an employee changes roles. A shared agent should not inherit a powerful employee’s unrestricted access simply because that employee created it. Test a user who lacks access to a document, a revoked connection, and a task that spans two departments before expanding the trial.
Budget delegated work separately from everyday chat
Microsoft distinguishes subscription-backed everyday use from usage-based agentic work. Administrators should therefore evaluate a recurring workflow as a recurring operating expense. An agent that checks a channel all day has a different cost profile from a person requesting a single summary.
Our editorial recommendation is to set a budget and approval boundary for one measurable process first. Record completed tasks, correction time, and charges together. Keep an existing manual handoff available until failures are understood. The launch offers a broader set of tools; deciding which work to delegate still requires evidence from the organization’s own workflow.