Home, Code and Autopilot, the three pieces of Microsoft’s latest Copilot announcement, move Copilot from answering questions to taking on work. Most enterprise AI so far has aimed at individual productivity and left the operating model alone. This release points at the operating model. Until now Copilot has been a feature IT licenses per user and switches on. In our view it is becoming a platform IT operates, with a consumption budget, new identities in the tenant, apps built outside IT, and a dependence on how ready the organization’s data and permissions are. We’d settle each of those before delegation opens to the wider organization.
Home and The Shift to Delegation
The biggest shift, in my view, is from assistance to delegation. Home puts Chat, for questions and drafts, beside Cowork, which takes complex tasks such as RFPs and briefings end to end. Microsoft says users soon won’t need to choose a mode, because Copilot will route each request to Chat, Cowork or Code. I agree with the principle underneath: a person describes an outcome, and the system works out the reasoning, tools and data it takes. That is closer to delegating to a capable colleague than to operating software.
Copilot Code and The Economics of Small Software
Code builds an app, tracker, dashboard or workflow from a natural-language description. It runs sandboxed on the same technology as GitHub Copilot. Copilot Managed Runtime hosts what it builds inside the organization’s Microsoft 365 environment, with governance for IT over what runs.
I don’t expect your finance manager to become a programmer. The change is economic. Plenty of small business problems would benefit from purpose-built software but could never justify a development project, and a problem too small, specialized or temporary to build for may now be solvable.
That software needs an owner. Copilot Managed Runtime and Microsoft’s cost controls cover where these apps run and what they spend. In our view they leave open who maintains an app once its author changes roles or leaves, a question IT leaders have met before with citizen-developed apps. We’d decide ownership, lifecycle and retirement before opening Code broadly, because an app nobody owns still has users depending on it.
Autopilot Keeps Working Between Prompts
Autopilot, previously called Scout, is given a named role and goals, then watches channels, follows threads, runs recurring work and resumes projects days later without a prompt. Microsoft’s example is a full supplier review. It lives in the tenant with its own identity, memory, computer and workspace, and people @mention it in Teams, Outlook and documents as they would a colleague.
An agent that carries on after the prompt holds a responsibility, which makes Autopilot a new identity to administer. I’d manage it like any identity with a lifecycle: an owner, access scoped to its role, monitoring and evaluation of what it does, human review, and a way to switch it off. Agents should not get unrestricted authority on day one. Our own agent work bears this out: once an agent moves from an individual experiment to shared production, each of those becomes a design decision settled before launch. Copilot is grounded in organizational context through Microsoft IQ and Fabric IQ, and in our view an agent working under its own identity can reach whatever its permissions allow. Oversharing becomes a question of agent authority as well as search quality.
Status and billing, from Microsoft’s announcement:
| Capability | Status | Billing |
| Home (Chat plus Cowork) | Frontier, coming weeks | Chat: user subscription license. Cowork: usage-based |
| Code | Frontier, end of month; wider in coming weeks | Usage-based |
| Copilot Managed Runtime | Preview | Not stated |
| Autopilot | Private preview, expanding at end of month | Usage-based |
Delegation Runs on Usage-based Billing
Microsoft has split Copilot licensing in two. A user subscription license covers Chat and Copilot in Word, Excel, PowerPoint, Outlook and Teams at a fixed cost. Cowork, Code, Autopilot, long-running agentic work and frontier models run on usage-based billing.
The delegation capabilities at the center of this release are metered. The budget gets a consumption line beside the license, and that line moves with how much work gets handed to Copilot and which models run it. We’d configure FinOps for AI before any of those capabilities reach users, using the controls Microsoft lists: spending policies managed through an API, approval workflows for credit requests, model families set per user group, and Agent 365 cost management expanding to Code and Copilot Managed Runtime.
The Application Stops Defining The Work
For most of the history of enterprise computing, the application has defined the boundary of work: the CRM for customers, the ERP system for finance, Word for a document. AI begins to invert that. The first question moves from “Which application do I need to open?” to “What am I trying to accomplish?”
Microsoft is building in this direction, with Office working inside Copilot and grounding in Dynamics 365 and Power Platform data rolling out in public preview. The systems of record stay. They become context and tools that an AI system uses to reach an outcome.
Where IT Leaders Should Start
Start with the work, and choose the Copilot experience once you understand it. Take an important business outcome and trace how it gets produced today: where the information originates, where it gets copied, where somebody waits, where judgment matters, and where people perform mechanical work because the systems around them were never designed to do it. Then ask one question of that process:
If we designed this process today, knowing what AI can reliably do, would we design it the same way?
The first moves we’d make:
- Join the Frontier program with a scoped pilot group.
- Model the budget under both the user subscription license and usage-based billing.
- Set spending policies and approval workflows before enabling Cowork, Code or Autopilot.
- Name an owner for every agent and every app built with Code.
- Review permissions and oversharing before widening Copilot’s grounding.
The organizations that get the most value from this generation of AI will redesign work around what people do best and what machines do best, and decide deliberately where that boundary sits. More licenses or more agents will not get them there on their own.
Microsoft’s announcement gives a much clearer picture of the technology. Building the operating model around it is the harder work, and it is still ahead.