Summary
- Microsoft has reorganised Copilot around Home, Code, and Autopilot.
- Autopilot is intended to persist across longer-running work rather than ending with a single chat session.
- As agents gain more autonomy, identity, permissions, hosting, governance, and variable AI consumption become part of Microsoft 365 administration.
Microsoft is rebuilding Copilot around software that can complete work, create internal applications, and continue operating when a user is not sitting in front of the screen, pushing its enterprise AI product further from the familiar assistant window.
The new Copilot centres on three capabilities: Home, combining conversational Chat with Cowork; Code, for creating applications and workflows from natural-language instructions; and Autopilot, a persistent agent intended to continue work over longer periods.
The product direction changes what a workplace AI licence is expected to do. Microsoft began by placing generative functions inside Word, Excel, Outlook, Teams, and other productivity applications. It is now building an environment where AI can assemble software, invoke tools, retain context, and pursue work across several steps.
Persistent agents need persistent identity
Autopilot is the clearest expression of the shift. Microsoft describes it as a persistent and proactive agent that can continue working while the user is elsewhere rather than ending when an individual conversation closes.
That persistence creates a different enterprise security problem. An agent working independently needs an identity, defined permissions, and an audit trail because it may continue accessing tools and information after the employee who initiated the work has moved on to something else.
Organisations will need policies around who owns such agents, whose authority they inherit, which resources they can access, when long-running tasks expire, and what happens when the sponsoring employee changes role or leaves.
The problem is more difficult than governing a conventional chatbot because the system may act between human review points rather than merely returning an answer to a person.
Copilot also moves into software creation
Code extends the product towards internal application development. Employees can describe a tracker, dashboard, workflow, or other tool, drawing on technology related to GitHub Copilot to generate software from inside the Copilot environment.
That can broaden the population capable of creating internal applications, although producing code is only one part of deploying software safely. Hosting, identity, data access, permissions, policy, updates, and ongoing operation remain necessary after a generated application first works.
Microsoft is addressing some of that gap with managed runtime infrastructure and deeper integration into the Microsoft 365 estate, where administrators can apply organisational controls around applications and agents.
The approach resembles the next stage of low-code development: employees describe the intended outcome rather than manually assembling every component, while the platform underneath handles more of the deployment environment.
That convenience can also accelerate application sprawl. If every department can create dozens of small AI-generated tools, organisations will need stronger inventories, ownership rules, access policies, and processes for retiring software that no longer has a responsible maintainer.
Autonomy changes the software bill
Longer-running agents also change how AI consumption is financed. Simple prompts and document assistance are relatively bounded workloads, while an agent operating over several hours may query data repeatedly, invoke tools, use different models, and consume cloud infrastructure throughout the task.
Microsoft is consequently bringing usage controls and cost management further into Copilot administration. The shift makes AI spending resemble cloud FinOps more closely than conventional per-seat software licensing.
That could improve the way companies assess AI projects. A fixed licence can disappear into departmental overhead, while an autonomous workflow with measurable consumption invites a direct comparison between the cost of the agent and the time or output it replaces.
Microsoft’s broader strategy is therefore moving Copilot from a feature beside work towards infrastructure underneath it. Greater autonomy may create larger productivity gains, but it also moves model failures closer to operational systems.
The next phase will be measured less by how many employees open a Copilot chat window and more by which pieces of work organisations are prepared to delegate. Microsoft is assembling the identity, application, runtime, and cost-control layers needed for that delegation; enterprises now have to decide where autonomy is reliable enough to earn it.












