On September 25, Microsoft announced a major restructuring of Copilot at its annual fall event. The assistant is now organized around three surfaces: Home (chat and collaboration), Code (natural-language tool building), and Autopilot (long-running agents that operate in the cloud). Word, Excel, and PowerPoint are embedded inside Copilot rather than accessed through it.
Microsoft CMO Jared Spataro framed the change directly: “Just as Office defined how work got done in the PC era, the new Copilot is designed to define how work gets done in the AI era.”
The comparison is not rhetorical. It describes a change in what work is made of.
The Fourth Unit of Knowledge Work
Microsoft's internal framing describes knowledge work as having three units: documents, spreadsheets, and slides. Code introduces a fourth: small, purpose-built tools that any employee can create by describing what they want.
The tools run in a sandboxed environment, built on the same underlying technology as GitHub Copilot. Microsoft says “learning to build these small tools is becoming a skill as fundamental as writing a memo or making a budget.”
That sentence matters more than the product description. If a memo and a tool are equivalent skills, the document is no longer the container that work lives in. It becomes one possible output. The primary act shifts from editing something that already exists to describing something that should.
Home reflects the same logic. Chat and Cowork are merged into one surface. Microsoft says users “won't need to choose a mode” — they state a goal, and Copilot routes the request to the appropriate capability automatically.
The Meter Is the Product
Microsoft split pricing into two layers. The base subscription (USL) covers Chat, Office integration, and model selection. A new usage-based billing layer (UBB) covers Cowork, Code, Autopilot, and access to frontier models including Astra and Fable.
Usage billing is off by default. Administrators set spending caps and manage it through FinOps for AI. But the structure itself is the signal: AI work now has a variable cost, not a fixed one.
ZDNET's assessment was blunt: “Nobody knows how much this will cost. Not even Copilot itself.”
Autopilot runs continuously in the cloud, in Microsoft's words, “even while you sleep.” A long-running agent with persistent memory and compute has no natural spending ceiling. That is a different procurement problem than a per-seat license. CFOs who approved Copilot as a fixed line item now face a consumption model with soft limits and uncertain unit economics.
Autopilot Has an Identity
Microsoft describes Autopilot as having “its own identity, memory, computer, and workspace in your tenant.” It can be @-mentioned. It works in Teams, Outlook, channels, and documents. It can set up a vendor review process, manage meetings, and follow up with stakeholders.
That is not a tool description. It is a description of a digital colleague.
Microsoft says the necessary “permissions, audit, and governance mechanisms” are in place. It has not said who defines Autopilot's boundaries, who audits its decisions, or how an organization revokes its access when a project ends. A human employee has a manager, a role description, and an offboarding process. Autopilot has a spending cap.
The governance gap is not unique to Microsoft. It is the same question Anthropic faces with embedded evaluators and OpenAI faces with agent logs. When an AI entity operates inside an organization with its own identity and persistent memory, the question of who it reports to does not have an obvious answer.
What the Restructuring Establishes
Three things are now clear. Microsoft is positioning Copilot as the platform layer for AI-era work, not as a feature inside Office. The economics have shifted from subscription to consumption, with frontier model access billed separately. And Autopilot is being introduced as an entity with identity and autonomy, not as an interface.
The open question is the same one that follows every agent deployment: when the agent has its own workspace and the meter runs continuously, who decides when it stops?
P.S. Microsoft says Autopilot will be available to Frontier program customers first, with broader rollout later. The company has not published a rate card for usage-based pricing, and the FinOps for AI controls are designed for administrators, not end users. An employee who asks Autopilot to manage a vendor review will not see the meter running.
Frequently Asked Questions
Q: What did Microsoft announce?
A: On September 25, Microsoft restructured Copilot around three surfaces: Home (chat and collaboration), Code (natural-language tool building), and Autopilot (long-running cloud agents). Word, Excel, and PowerPoint are now embedded inside Copilot.
Q: What is Code?
A: A surface that lets any employee create small, purpose-built tools by describing what they want. Microsoft says building these tools is becoming “a skill as fundamental as writing a memo.”
Q: How does pricing change?
A: The base subscription covers Chat and Office integration. Cowork, Code, Autopilot, and frontier model access move to usage-based billing, off by default with administrator spending caps.
Q: What is Autopilot?
A: A long-running agent with its own identity, memory, computer, and workspace inside the customer's tenant. It can be @-mentioned, works in Teams and Outlook, and runs continuously in the cloud.
Q: What governance questions remain?
A: Microsoft has not said who defines Autopilot's boundaries, who audits its decisions, or how an organization revokes access when a project ends. It has a spending cap, but not a manager.
