Microsoft Copilot isn't just answering questions anymore. It's being given jobs

Keep up with IT Club
Add IT Club as a preferred source in Google Search.
Microsoft's September 2026 Copilot announcement introduced Home, Code and Autopilot: a new starting point for work, natural-language app creation, and an agent intended to keep working on assigned objectives. This article explains what Microsoft says each feature does, what remains in preview, how usage-based billing differs from a per-user licence, and why any agent needs a clear job, limited permissions and an accountable owner.
A chatbot waits for a question. An agent can be asked to do a job. That difference is the important part of Microsoft’s September 2026 Copilot announcement—not the new names or the reshuffled app.
Microsoft introduced three capabilities called Home, Code and Autopilot. The first is a new starting point for work, the second is meant to help people build small software tools, and the third is intended to keep working towards objectives after the user has moved on. It is an ambitious direction. It is also a set of preview features, not a sudden upgrade that every Microsoft 365 customer already has.
Home: a place to pick work back up
Home brings Copilot Chat and Cowork together with recent activity and suggestions. Microsoft also describes Office in Copilot, where people can create and edit Word, Excel and PowerPoint files as collaborative documents that stay in sync with the Office apps.
The business question is less “Does Copilot have a new home screen?” and more “Which information does it use when it helps with this task?” A document, chat and spreadsheet can each contain different versions of a decision. Bringing them into one experience may save switching between apps, but it does not remove the need to check which file is authoritative and who is allowed to see it.
Code: describing a tool instead of starting from code
Code is Microsoft’s natural-language route to creating small apps, trackers, dashboards, automations and workflows. A user describes what they want to do; Copilot helps build a tool intended to do it. Microsoft says the experience runs in a sandbox and can be hosted within a company tenant. Those are product claims, not a substitute for checking how a specific app handles data, access and updates.
A useful internal tracker can still make a poor business system if nobody owns it, if its data goes stale or if an employee builds in assumptions nobody else understands. A prototype that works on sample data is not automatically ready to handle customer records or financial decisions. Decide who will test and maintain the result before treating it as production software.
Autopilot: the more important change
Autopilot is the feature that makes the shift from chatbot to agent clearest. Microsoft describes it as a persistent, proactive agent that can continue working on assigned objectives, including following up or carrying out recurring tasks. It is presented as having an identity, memory and workspace in the organisation’s environment, and as appearing across places such as Teams, Outlook, chats and documents.
That description is a preview of what Microsoft intends the agent to do; it is not evidence that every task will run correctly without oversight. If an agent keeps working while nobody is looking at the screen, the business needs to know what it can read, what it can change, what it can send, and who will notice when it misunderstands the goal or gets stuck.
The shift
Ask AI something → delegate a task → give AI an ongoing job
A job description includes permissions
A human employee gets a role, a manager, policies and limits on what they can approve. An agent needs equivalent boundaries, even if its job description is written in a prompt. Before assigning work, a business should be able to answer:
- What is the specific outcome, and what is outside the job?
- Which documents, mailboxes, customer records or systems can it read?
- Can it change records, send external messages, create users or approve spending?
- Which actions require a person to review or approve them?
- Who checks its activity and notices if it has stopped or gone off track?
- Who can pause the agent, remove its access and preserve its audit history?
- What happens to its identity and permissions when its owner changes role or leaves?
Start with the minimum access needed for one clearly defined task. Read-only access is a safer first test than permission to edit live records. Keep a named human owner, log the actions that matter, and decide in advance which mistakes would require an immediate stop. These are sensible controls for an automated process, not special legal duties created by a product announcement.
Your AI Did Exactly What You Asked. Just Not How You Expected. →
Preview means preview
Microsoft’s 25 September announcement said Home and Code would begin rolling out through Frontier, while Autopilot would expand through private preview at the end of September. Those words describe staged early access. They do not establish that the rollout is complete, that every eligible organisation has access, or that the features are generally available.
Microsoft describes Frontier as a pre-release programme managed at tenant level. Administrators control participation and which users can access preview features, and the features may change. The licence, tenant settings and exact feature entitlement all matter. A Microsoft 365 account on its own is not proof that Home, Code or Autopilot is available.
That is worth checking before promising a team that an agent is ready to take over a workflow. If the feature is not enabled for your tenant, the useful next step may simply be to map a suitable process and decide what controls it would need—not to buy something in anticipation of a rollout.
Microsoft Learn: get started with the Copilot Frontier programme →
A job that runs in the background needs a cost boundary too
Microsoft says Cowork, Code and Autopilot are among the Copilot experiences that use usage-based billing. This is different from the ordinary user subscription licence: some advanced work is metered using Copilot Credits and requires an organisation to configure an appropriate billing policy. The exact feature eligibility and administration can depend on the tenant and service.
No per-credit rate or typical bill should be guessed from the announcement. The practical point is simpler: when work can continue without someone continually prompting it, a business should decide what usage it permits and who can see it. Set an agreed budget, turn on the relevant usage reporting or alerts, and start with a bounded task before expanding it.
This is not a prediction that an agent will automatically run up a large bill. It is ordinary cost governance applied to a tool whose activity may be less visible than a person clicking through a workflow. Scope, duration, access and retry behaviour all affect what an agent is doing—and may affect what it consumes.
If AI use grows, who is paying for the computing? →
A sensible first pilot is small and reversible
For a business that does gain access, a useful pilot might be an internal task using a limited set of non-sensitive documents, with a person reviewing the result. Keep the first version away from payment approval, changes to important customer records or unreviewed messages to customers. Define what “done” looks like, how long it can run, who checks the output and how to stop it.
Then compare the complete workflow with what staff do today: time saved, corrections required, quality, reliability and any extra administration. If the agent needs broad access or frequent rescue to finish a simple task, that is useful information too. A pilot should be allowed to conclude that a chatbot, a rule or a human process is the better tool.
The IT Club view
A chatbot needs a prompt. An agent needs a job description, permissions and boundaries. The more persistent the work, the more important it is to know who owns the agent, what it may do, how its actions are checked and how its authority can be removed.
There is no need for an SME to adopt every new agent the week it is announced. If a real process is worth automating, start with the smallest useful task, current availability and licensing, limited access, a named owner and a tested stop button. The goal is not to make AI sound like an employee. It is to make sure the work still belongs to the business.
Can AI Really Act as Your PA? →
Sources and further reading
The feature descriptions and rollout language below are Microsoft’s own announcements and documentation. Availability and billing can change; check current tenant and product documentation before planning a deployment.
Microsoft: Introducing the new Copilot with Home, Code and Autopilot →
Microsoft Copilot Blog: evolution of the Copilot pricing model →
Microsoft Learn: user subscription licence and usage-based billing →
Microsoft Learn: get started with the Copilot Frontier programme →
Found this useful? Forward it to someone who might too.
Plain-English Takeaway
A chatbot needs a prompt; an agent needs a job description, limited permissions, an audit trail, a spending boundary and a reliable way to stop it.
Related Articles
How to Find Every Photo on Your Windows PC
Photographs often become scattered across a Windows PC — saved in Pictures, Downloads, Desktop, OneDrive, project folders and places you have long forgotten. File Explorer can search across the entire computer and display image files from many locations in one results view. Here is how to do it safely.
Read articleGemini can phone businesses—but its first trial is US-only
Google's Gemini can make some routine calls to businesses on a user's behalf, but its first Call for me rollout is limited to eligible US customers. UK firms are not included.
Read articleAI is getting a face. Will that make us trust it too much?
Google’s Live Avatar adds a generated face to spoken AI conversations. It may make an interface feel more natural, but a friendly expression is no evidence that an answer is right.
Read article