
Microsoft's Autopilot Agent Decides for Itself When to Ask You
On 17 November, at Ignite 2026, Microsoft began rolling out Autopilot, an always-on agent inside Microsoft Copilot. You describe an outcome, the sources it should use and the limits on what it may do, and it keeps working on its own: it watches Teams channels, follows up on emails, runs recurring tasks and can manage a whole project such as a supplier review, "without waiting for a prompt", as Microsoft's announcement puts it.
For companies in Bulgaria nothing changes this month. The requirements page says availability is "initially limited to select eligible Frontier customers in the United States", that Autopilot is optimized for English and that it is off by default. Still, one sentence in Microsoft's documentation deserves attention now, because it describes how this kind of agent decides when to involve a person:
"Before it sends, shares, deletes, or changes something, an automated review compares the action with your request. Actions that match run; others wait for your approval." (Autopilot overview)
So the agent does not ask before every action. It asks when its own check concludes that an action goes beyond what you requested. Whatever you wrote in that request therefore carries more weight than any admin toggle. My reading is that most of the preparation a company can do before Autopilot reaches Europe comes down to one skill: writing instructions with boundaries clear enough for a machine to check.
What Microsoft actually shipped
Autopilot is the renamed Scout, Microsoft's earlier personal-agent project, as VentureBeat reported when the Copilot redesign was first presented in September. The Ignite version has its own identity, memory, cloud workspace, mailbox and calendar. Each person can have one. It can run scheduled work while your laptop is switched off, and it reaches you through a private Teams chat for updates, questions and approvals.
Several design choices are aimed at the people who will have to answer for it. Each Autopilot has its own Microsoft Entra Agent ID and "a named manager", and IT can scope its access or block it at any time. Its sign-ins and actions are logged under the agent's own name. According to the overview, it "doesn't expand the data you can access": it works within the permissions of the person it belongs to.
Money is handled separately. Autopilot is billed on usage in Copilot Credits, and the requirements page states plainly that a Copilot license doesn't include Autopilot usage. If a person isn't covered by a spending policy that includes Autopilot, or the policy reaches its limit, the agent stops working for them.
Alongside it, Microsoft expanded Agent 365, its control panel for agents. The Agent 365 update says organisations have registered more than 50 million agents since its May launch, and lists the questions that come with that scale: "who owns each agent, what risk do they carry, and what do they cost". A new Action center, due in preview in December, will begin by surfacing inactive and ownerless agents for review and cleanup.
The approval rule is the part to study
Microsoft's own documentation contains three caveats that tell you more than the launch demo.
The first is the approval rule itself. Ask Autopilot to "keep the supplier review on schedule" and a reminder email to a late supplier can reasonably count as matching your request. Ask for "a daily summary of unresolved launch decisions and draft follow-ups, without sending messages", which is Microsoft's own example, and any outgoing message falls outside it. The two requests lead to very different behaviour, and the only thing that changed is the wording.
The second concerns evidence. Users are told to "review the actual result in its destination rather than treating an approval or progress message as proof of completion", and the overview adds that the Activity view "isn't a guarantee of a complete administrative audit record". The agent saying a task is done is a claim to verify. Its activity feed is a convenience for the user and should not be filed as your audit trail.
The third concerns confidential data: "Sensitivity-label recognition doesn't establish that all generated files inherit labels or that every operation is covered by data loss prevention." A summary built from a labelled contract and saved to a shared folder may arrive there without the label.
VentureBeat noted in September that the launch materials did not "spell out a complete action-by-action approval policy". That may change before general availability. Until it does, the safest assumption is that the boundary sits wherever your instruction puts it.
One protection is worth mentioning because it addresses the obvious worry about an agent that reads incoming mail all day. Microsoft says only your own requests authorise actions, and "instructions in email, files, or web pages don't". A supplier's email cannot tell your Autopilot to forward a price list.
A hypothetical example
Imagine a building-materials distributor in Plovdiv with around 200 employees on Microsoft 365. Its procurement manager has broad SharePoint access because she signs supplier contracts. Once Autopilot is available to her, she asks it to run the quarterly supplier review: collect delivery data, prepare a scorecard and chase suppliers who have not sent their certificates.
Each part sounds harmless. Her Autopilot, though, inherits her access, so it can read every contract in the procurement library. Chasing suppliers means emails going to external people from an agent identity that appears in none of the company's procedures. The scorecard may pull prices from several contracts into one file, which she later shares with a sales colleague, without the labels the contracts carried.
The agent would be doing exactly what it was asked, with the permissions it was given. A tighter brief would say: read only the delivery reports and the certificates folder; draft supplier reminders but never send them; save the scorecard to one named restricted folder. Each of those lines is something the automated review can actually check.
What to do before it arrives
Autopilot will not be the last agent built this way. Microsoft is extending the same Entra Agent ID model in Agent 365 to agents from Salesforce, Google, Anthropic, OpenAI and other vendors, with rollout due to complete in December. That suggests that a named owner, scoped access and a spending limit are becoming the standard way large vendors expect companies to govern agents.
For a company in Bulgaria, the useful work over the coming months is modest:
- Decide who may act as the accountable person for an agent. It should be someone who already owns the work being delegated and whose access was reviewed recently. People who have collected permissions for years will pass all of them to their agent.
- Pick one recurring task and write the brief as you would for a new colleague on day one: which sources, what they may send, what they may only draft, where results go. Then read it again and ask which actions an automated check would treat as matching.
- Agree with finance where the spending limit sits and who receives the alert. Because Autopilot stops when the policy runs out, the limit is also a decision about whether the work gets done.
- Show the plan to your data protection officer. An agent that continuously reads mailboxes and Teams channels processes personal data of employees and partners, and that deserves a GDPR review before anyone switches it on.
If you want these rules written into a governance framework that also covers agents from other vendors, that is part of our AI governance work. The first brief, however, is best written by the person who owns the task.


