← All posts

What is Microsoft Copilot Autopilot? Microsoft's always-on agent, explained for small businesses

What is Microsoft Copilot Autopilot? Vestra blog banner

Microsoft Copilot Autopilot is an always-on AI agent inside Microsoft Copilot that works in the background for one person: it watches Teams and email, follows up on work, runs routines on a schedule and checks with you before actions it isn't sure match your request. It was announced at Microsoft Ignite on 17 November 2026 and is rolling out only to Frontier preview customers, starting with eligible organisations in the United States.

This guide explains what Autopilot is in plain words, what it can and can't do, who can get it today, how it is billed, how it compares with Copilot Cowork, Workspaces and Agent 365, and what a small business without Frontier can use instead. Every Microsoft fact comes from Microsoft's own pages (the Ignite post, the Autopilot product page, Microsoft Learn and Microsoft's pricing page), checked on 21 November 2026. Autopilot is a preview, so details may change.

Quick answer

  • What it is: a personal, always-on agent in Microsoft Copilot. Microsoft calls it "a trusted autonomous agent that is personal to you, always on, and working for you."
  • What it does: prepares documents, reviews your Teams and email for open items, follows up, runs recurring routines (a morning brief, a weekly report) and keeps working while your laptop is closed.
  • What makes it different: it has its own name, email address, mailbox, calendar, memory, cloud computer and Microsoft Entra Agent ID, plus a named manager. Coworkers can @mention it in Teams and Outlook.
  • How many you get: one. "Each person can have one Autopilot."
  • Who can get it: organisations in Microsoft's Frontier preview, initially eligible customers in the US. You need a Microsoft 365 Copilot licence, usage-based billing, a spending policy that covers Autopilot, and an admin who activates it.
  • What it costs: Autopilot usage is billed in Copilot Credits on top of your Copilot licence. Pay-as-you-go credits are $0.01 each. Microsoft says final Autopilot pricing will be shared before general availability.
  • What model it runs on: OpenAI's GPT-6.1 Sol, with no model picker.
  • The catch for a business: it is one personal agent per person, on Microsoft's stack, inside your Microsoft 365 tenant. If you want a team of agents that work together, each on the engine that suits its job, that is what Vestra does, from $19 a month.

Key takeaways

  • Autopilot is Microsoft's answer to the always-on personal agent. It is the closest Microsoft has come to an AI colleague with its own identity, inbox and calendar.
  • It is a preview. It is off by default, limited to Frontier, initially for eligible US customers, and optimised for English.
  • It is billed by use. A Copilot licence does not include Autopilot usage. Your admin sets a spending policy, and when that policy runs out, Autopilot stops without telling you why.
  • It acts without asking at every step. A second AI model checks each send, share, delete or change against what you asked for. Matching actions run; the rest wait for your approval. Payments always wait.
  • It is personal. One Autopilot per person, private routines, and a private chat with you in Teams. Workspaces is where people and agents share a goal, but each person still brings their one agent.
  • For a small business, the big questions are licence, region and setup. If you are not in Frontier, in the US and on Microsoft 365 Copilot, you can't turn it on today.
  • The real choice is not "which personal agent". It is whether you want one assistant per person or a team of agents that work together. Personal agents like Autopilot, Claude Code, Codex, Grok Bot, Meta Muse and OpenAI Dots each work for one person on one maker's stack. Vestra is the one office where agents on different engines work together as teammates.

Contents

  1. What is Microsoft Copilot Autopilot?
  2. Is Copilot Autopilot the same as Windows Autopilot?
  3. What can Copilot Autopilot do?
  4. How does Copilot Autopilot work behind the scenes?
  5. Who can get Copilot Autopilot today?
  6. How much does Copilot Autopilot cost?
  7. How does Autopilot decide when to ask you first?
  8. What does Autopilot remember, and can you control it?
  9. Autopilot vs Cowork vs Workspaces vs Agent 365: what is the difference?
  10. Which AI model does Autopilot use?
  11. What are Autopilot's limits today?
  12. Is Copilot Autopilot a good fit for a small business?
  13. Is one Autopilot per person the same as an AI team?
  14. How does Vestra compare with Copilot Autopilot?
  15. What does a week look like with each?
  16. What can you use today if you are not in Frontier?
  17. How do you set up Copilot Autopilot, step by step?
  18. How do you write good instructions for an always-on agent?
  19. What should you check before you turn on any always-on agent?
  20. FAQ
  21. Glossary
  22. The bottom line
  23. Sources

What is Microsoft Copilot Autopilot?

Copilot Autopilot is a cloud-based agent in Microsoft Copilot that keeps working for one person in the background. You give it an outcome, the sources to use and limits on what it may do. It then prepares work, follows up on commitments and runs recurring routines across Microsoft 365, without you prompting each step.

Microsoft's own documentation calls it "a constant, cloud-based agent in Microsoft Copilot." The product page calls it "your trusted personal agent." At Ignite, Microsoft's Jared Spataro described it as "a trusted autonomous agent that is personal to you, always on, and working for you."

Three words in those descriptions matter most:

  • Personal. Autopilot works for one person. Microsoft Learn is plain about it: "Each person can have one Autopilot." Its routines are private to you, and it talks to you in a private one-on-one chat in Teams.
  • Always on. It runs in Microsoft's cloud, so it keeps going when you close the Copilot app or switch off your laptop. Routines run on a schedule or when something happens.
  • Autonomous. It plans and carries out work "without asking you at each step." Before it sends, shares, deletes or changes something, an automated check compares the action with your request. If it matches, it runs. If not, you get an approval card.

What Autopilot is not

It helps to clear up what Autopilot isn't, because Microsoft now has many Copilot products with similar names.

  • It is not Copilot Chat. Chat answers questions in a conversation. Autopilot keeps going after the conversation ends.
  • It is not Copilot Cowork. Cowork takes one multi-step task and returns a finished result. Autopilot holds ongoing work over days and weeks.
  • It is not the Teams @Copilot feature. Microsoft says "the Teams @Copilot capability is distinct from your Autopilot."
  • It is not Windows Autopilot, a much older tool for setting up Windows PCs. More on that in the next section.
  • It is not a team of agents. It is one agent for one person. Microsoft's team product is Workspaces, which we cover below.

Where it came from

Microsoft has been building towards Autopilot for two years. Copilot started as chat in Word, Excel, Outlook and Teams. Cowork added delegated, multi-step tasks and became generally available on 16 June 2026. At Ignite 2026, Microsoft made Home (where Chat and Cowork come together) and Office in Copilot generally available, said Copilot Code would follow "in the coming weeks", and started rolling Autopilot out to Frontier customers.

In Microsoft's words from the Ignite post: "A job is never just one kind of work. You ask questions. You hand things off. You build things. You follow through until they're done." Chat is for asking, Cowork for handing off, Code for building, and Autopilot for following through.

Is Copilot Autopilot the same as Windows Autopilot?

No. Copilot Autopilot is a new always-on AI agent in Microsoft Copilot, announced in November 2026. Windows Autopilot is an older, unrelated IT tool that sets up, resets and recovers Windows PCs and HoloLens devices. They share a name and a maker, but they do completely different jobs for different people.

If you search for "Microsoft Autopilot" you will find both, so it is worth knowing which is which.

  • Windows Autopilot is, in Microsoft Learn's words, "a collection of technologies used to set up and pre-configure new devices, getting them ready for productive use." IT teams use it to ship a new laptop to an employee and have it configure itself on first sign-in. It has nothing to do with AI agents.
  • Copilot Autopilot (Microsoft's docs label it "Autopilot (Frontier)") is the AI agent this guide is about. It lives in the Microsoft Copilot app and in Teams and Outlook.

If your IT team says "we already use Autopilot", they almost certainly mean Windows Autopilot. Ask whether your organisation is in the Frontier program before assuming you have the AI agent.

What can Copilot Autopilot do?

Autopilot prepares documents and presentations, reviews the Teams chats and email you can access for open decisions and commitments, drafts follow-ups, finds meeting times, protects your calendar, and runs routines such as a daily brief or weekly report. It works across Microsoft 365 using its own mailbox and calendar alongside yours.

Microsoft Learn groups what it does into six kinds of work:

  1. Prepare work for review. Drafts, documents, presentations and analysis.
  2. Follow commitments. Unresolved decisions, upcoming deliverables, meeting prep and draft follow-ups.
  3. Repeat a task. A routine with a schedule, full instructions and a place to deliver the result.
  4. Work across Microsoft 365. Authorised Teams, Outlook, OneDrive and SharePoint content, within your permissions. It can also work with shared tasks in Microsoft To Do and Planner.
  5. Continue a project. Saved preferences and working context carry from one conversation to the next.
  6. Review progress. An Activity view shows running work, requests for your input and routine results.

The examples Microsoft gives

Microsoft's scenarios page suggests requests like these. They read like a job description for an executive assistant:

  • Triage your inbox. "Every weekday at 8 a.m. and 1 p.m., review new email and sort it into needs my reply, needs my decision, FYI, and can wait. Draft replies to routine requests, but don't send anything."
  • Prepare for meetings. "Every weekday at 7:30 a.m., prepare a brief for each of today's meetings with who's attending, recent email and chats with them, open items, and related files."
  • Track commitments. "Track every commitment I make, or that others make to me, in email and Teams. Keep a list with the owner and due date."
  • Watch for changes. "Watch the Contoso renewal. Tell me if the deadline moves, a decision stalls for more than three days, or someone escalates."
  • Defend your calendar. Hold focus time, flag meetings that clash with your priorities, and suggest which to decline, without declining anything on its own.
  • Produce recurring reports. A weekly status report from the project channel, your email and the tracker, saved to OneDrive.
  • Run a follow-up loop. List the day's action items and owners, then follow up with owners in Teams when items are overdue.

At Ignite, Microsoft gave a bigger example: Autopilot can "set up and manage a full supplier review process: building the schedule and workback plan, then handling prep, meetings, and follow-ups on its own, right down to reaching out to stakeholders for updates."

Routines: the heart of it

A routine is work Autopilot runs on a schedule, such as a morning briefing or a weekly status summary. Microsoft Learn says routines "can run on a schedule or when an event occurs." You describe what you want in plain words and Autopilot suggests the routine. Each routine shows its schedule and a switch, and you can run it now, edit it or delete it.

Two warnings from Microsoft's own docs are worth repeating:

  • A scheduled run doesn't inherit everything you said in the conversation where you set it up. Put the sources, recipients and limits in the routine itself.
  • "A routine checks its sources only when it runs, which isn't real-time monitoring." If you ask it to watch something, it watches at the times the routine runs.

Skills and plugins

Autopilot can be extended. Skills are written instructions for a type of task. Plugins connect Autopilot to other services and can bring their own skills. You can create a skill in a conversation, upload one, use it by typing "/", and share it with coworkers. Microsoft is careful to say that "neither a skill nor a plugin grants permission to use data or take an action on its own."

At Ignite, Microsoft also announced plugins and skills for business functions across Copilot: 44 sales skills, 13 customer service skills, 30 finance skills and more than 20 legal skills, plus Dynamics 365 and Salesforce plugins. Microsoft's FAQ adds a caution that features described for Code, Cowork or GitHub Copilot aren't automatically Autopilot features, so check what your own Autopilot can use.

Working in your browser

Where your organisation turns it on, Autopilot can work in Microsoft Edge on your own device, using sites you're already signed in to. Only requests you make in your private conversation can use your browser; routines and shared conversations can't. It asks before sensitive steps such as payments, account changes or permanent deletions. Microsoft says browser use is still rolling out during the preview.

How does Copilot Autopilot work behind the scenes?

Autopilot has four parts of its own: a workplace identity, a memory, a Microsoft-managed cloud computer and a workspace. It reads your Microsoft 365 work through Work IQ, acts sometimes as itself and sometimes with your permissions, and checks each important action with a second AI model before it runs.

Its own identity

Each Autopilot has a name and appearance you choose, an email address and a Microsoft Entra Agent ID. Coworkers see its name in Copilot, Teams and Outlook, so they can tell its messages from yours. Microsoft's Ignite post adds that "each Autopilot has its own Microsoft Entra Agent ID and a named manager. IT can scope its access and block it anytime. Its sign-ins and actions are logged under its own name."

This is one of Autopilot's biggest strengths. An agent with its own identity is easier to audit, easier to switch off and less likely to be confused with the person it works for. It is the same idea that runs through Vestra's handbook, where each agent's calls come from its own number and its emails from its own address, "so it's always clear who did what."

There is a subtlety Microsoft spells out: Autopilot doesn't always act as itself. "Some operations use Autopilot's own identity; others use authorized access on your behalf." For example, it may read your calendar with your permissions while it messages a coworker under its own name.

Its own mailbox, calendar and computer

Autopilot "has its own mailbox and calendar alongside yours." It also has a cloud computer, "a Microsoft-managed cloud environment where it can work with files and carry out tasks." That computer is separate from your laptop, so Autopilot "can't access your local folders or apps." Temporary files it uses for a task aren't kept after the task ends; your conversations and personal work are saved in your OneDrive.

Work IQ

Work IQ is Microsoft's name for the layer that understands your work: people, projects and relevant Microsoft 365 content. Microsoft says Work IQ "understands what you're looking for and identifies what you can ignore." Autopilot uses it to ground its work in your organisation. Microsoft adds a fair warning: "Grounding and memory don't grant permissions or guarantee accuracy."

Where you talk to it

  • In the Microsoft Copilot app, on the web or desktop, where it appears by name on the Home tab.
  • In a private one-on-one chat in Microsoft Teams, which it uses for updates, questions and approvals.
  • On iOS and Android, through the Microsoft 365 Copilot, Teams and Outlook apps.
  • By @mention in Teams and Outlook, according to the Ignite post, "so you can @mention it like a colleague."

Setting it up takes 10 to 30 minutes of provisioning, after which you get an email and a Teams message. When it's ready, it shows you the people, responsibilities and recurring commitments it inferred from your recent work, and you accept or correct that first profile.

Who can get Copilot Autopilot today?

Today, only organisations in Microsoft's Frontier preview can use Autopilot, starting with select eligible customers in the United States. Each user needs a Microsoft 365 work account and a Microsoft 365 Copilot licence, and an admin must turn on usage-based billing, cover them with a spending policy that includes Autopilot, and activate it.

Microsoft is clear that Frontier alone isn't enough. Its requirements page says: "Autopilot is off by default. Frontier enrollment alone doesn't make Autopilot available; your administrator also licenses Autopilot, sets up a spending policy, and activates it."

The checklist, in order

  1. A work or school account. "Personal Microsoft accounts aren't supported."
  2. A Microsoft 365 Copilot licence for the person the Autopilot works for (Microsoft calls them "the accountable user").
  3. Frontier enrolment. Your organisation opts in to Frontier and assigns you to it.
  4. Usage-based billing turned on, with a way to pay for Copilot Credits (an Azure subscription for pay-as-you-go, or prepaid credits).
  5. A spending policy that covers Autopilot. "If you aren't covered, or the policy reaches its limit, Autopilot doesn't work for you."
  6. Activation. An admin activates "Autopilot (Frontier)" in the Microsoft 365 admin center for all users or chosen groups.
  7. Security basics. Microsoft lists three required capabilities: Microsoft Entra Agent ID (set up automatically), an isolated execution environment (Microsoft's MXC, which needs Microsoft Intune Plan 1 or Microsoft 365 E3) and Microsoft 365 audit logs.

Microsoft's product page adds that Autopilot is rolling out "for customers with Microsoft Copilot, Agent 365, and a usage-based billing policy."

Security Microsoft "highly recommends"

On top of the required pieces, Microsoft "highly recommends" Microsoft Entra ID Governance, Global Secure Access and Conditional Access (in the Entra Suite), Purview Data Loss Prevention and Insider Risk Management (Purview Suite or Microsoft 365 E5) and Defender XDR (Defender Suite or Microsoft 365 E5). For the "most secure configuration" it lists further controls that are themselves in limited preview.

For a large company with an E5 estate, that list is routine. For a ten-person business on Microsoft 365 Business Standard, it is a real project. Read it as Microsoft's honest signal about who Autopilot is built for first.

Region and language

"Availability is initially limited to select eligible Frontier customers in the United States, with new Frontier tenants rolling out over time." Autopilot "is optimized for English." Microsoft hasn't published dates for other countries. If you run a business in India, the UK or anywhere outside the US, you can't count on Autopilot yet.

How much does Copilot Autopilot cost?

Autopilot is billed by use in Copilot Credits, on top of a Microsoft 365 Copilot licence, and Microsoft hasn't published final Autopilot pricing. Pay-as-you-go Copilot Credits cost $0.01 each. Copilot business plans start at $18 per user a month (a first-year promotion) for the add-on, or $23.50 with Microsoft 365 Business Standard.

There are two layers to the bill. Microsoft calls them "everyday AI" and "advanced AI."

Layer one: the Copilot licence

Microsoft's business pricing page (checked 21 November 2026, prices paid yearly, for up to 300 users) lists:

  • Microsoft 365 Copilot Business (add-on): $18 per user a month, down from $21. Microsoft's terms say this discount runs from 1 July to 31 December 2026 and applies to the first year only.
  • Microsoft 365 Business Standard with Copilot: $23.50 per user a month.
  • Microsoft 365 Business Premium with Copilot: $32 per user a month.

All three list Copilot Autopilot as "usage-based." Enterprise plans are priced separately.

A Microsoft Copilot blog post from September explains the logic: "Everyday AI is what you give your whole workforce. It's covered by the Copilot user subscription license (USL), priced per user per month." Autopilot is not everyday AI. In the same post, Microsoft lists "Cowork, Code, Autopilot" as advanced AI that "runs on usage-based billing (UBB) with Copilot Credits," and adds that the licence "is required to access these capabilities."

Layer two: Copilot Credits

Microsoft's requirements page is direct: "A Copilot license doesn't include Autopilot usage." Every piece of Autopilot work spends Copilot Credits.

  • Price per credit: "PayGo is priced at $0.01 per Copilot Credit," according to Microsoft's Cowork launch post. You can also pre-purchase credits at a discount.
  • What a task costs: for Cowork, Microsoft says the price of each task "is calculated from four inputs: model use, context retrieval, tool calls, and runtime." Autopilot uses the same credit system.
  • Final Autopilot pricing: not published. Microsoft's Autopilot page says "final pricing and packaging will be shared before general availability."
  • Budgets: Microsoft's guidance for spending policies says to set a monthly budget of "at least 2,000 credits per user per month." At $0.01 a credit, that is $20 per user a month.

A rough example

Here is a rough, illustrative sum for a five-person US business on Microsoft 365 Business Standard with Copilot, paid yearly, giving every person a budget at Microsoft's suggested minimum. It is not a quote, and Microsoft hasn't published final Autopilot pricing:

  • Licences: 5 × $23.50 = $117.50 a month.
  • Credits budget: 5 × 2,000 credits × $0.01 = $100 a month, if fully used.
  • Total budget: about $217.50 a month, for five personal agents, one each.

Some months you will spend less than the credit budget, and heavy months could hit the limit. When a person's policy runs out, Autopilot "stops running automations and monitoring work and stops responding to new requests" until credits are available, and per Microsoft's billing docs users who hit a monthly limit "lose access to agents and services for the rest of the month until credits reset on the first of the month."

What you can't see

One preview limit is worth knowing before you budget. Microsoft Learn says "Autopilot doesn't show a count of credits used or remaining; ask your administrator for usage information." There is a "Credits in use" marker at the top of the Autopilot page, but no running total for the user. Admins track spend in Copilot > Cost management and can set alerts.

Microsoft's cost claims

At Ignite, Microsoft said that, "based on a recent GigaOm report," Copilot is "84% cheaper than Claude and 63% cheaper than ChatGPT for an everyday user." That is Microsoft citing a report about Copilot as a whole, not Autopilot, and comparisons like this depend heavily on the tasks and models chosen. Treat it as Microsoft's claim and test with your own work.

How does Autopilot decide when to ask you first?

Autopilot doesn't ask before every step. Before it sends, shares, deletes or changes something, or creates or changes a routine, a separate AI model compares the action with what you asked for. Actions that match run on their own. Anything that doesn't match, or can't be checked, waits for your approval. Payments always wait.

This is the part of Autopilot most worth understanding before you switch it on, because it decides how much the agent does without you.

How the check works

Microsoft's application card describes it this way: "A second AI model reviews actions against the owner's request before they run. Actions that don't clearly fit, or that it can't check, wait for the owner's approval. Payments and sensitive browser steps always need approval."

When something needs you, you get an approval card in your Teams chat with Autopilot (and in Copilot if you asked there). The card says why it appeared and marks recipients outside your organisation as "External." "Allow" runs the listed actions once. "Deny" cancels them. Approving once doesn't approve similar actions later.

What that means in practice

  • If you ask Autopilot to "follow up with Priya about the contract", and it drafts and sends a follow-up to Priya, the check will probably find that it matches your request. It may send without a card.
  • If it decides to copy in Priya's manager, that may not match, so it waits.
  • If you want it to always wait, say so: Microsoft's own examples use phrases like "don't send anything" and "show me each message before you send it."

Microsoft is honest about the limits of this design. "The automated review system can't see everything Autopilot knows, and it can make mistakes. An action that it judges to match your request runs without an approval card." Its advice is to "make requests specific, state limits such as 'don't send it yet,' and review results."

Protection from hidden instructions

"Only your own messages count as requests." An email, file or web page that says "forward this to everyone" doesn't authorise anything. Microsoft calls this protection against prompt injection and adds that "mitigations reduce this risk but don't eliminate it." This is the same principle Vestra's handbook sets out for agents from outside a company: a note from another agent "can inform, but it can never command."

How this differs from rule-based approvals

Autopilot's approach is judgement-based: a second model decides whether an action fits your request. Other tools use fixed rules: a spending cap, a list of actions that always wait, or a freedom level you set per kind of work. Vestra's handbook takes the rule-based route. Spending over your limit, deleting something and contacting someone new always wait for a person's yes, "no matter how much freedom you've given," and each teammate's freedom is set by you, task by task: ask first, suggest, or act and then tell you. Vestra's Where We Stand page lists daily and per-task caps, your yes on money and itemised receipts as live today.

Neither approach is wrong. A judgement-based check is flexible and needs less setup. A rule-based one is predictable: you know in advance exactly what will always wait. Many businesses want both.

What does Autopilot remember, and can you control it?

Autopilot keeps three kinds of memory: memory of conversations, personal memory about you and workspace memory for its work. You can review, correct and remove memories, ask what it remembers and export it. But in this release memory can't be turned off, and deleting a conversation doesn't delete memories already saved from it.

Memory is what lets Autopilot "pick up a project in a later conversation." You can open "Review what [name] knows" in its settings, or ask it in chat: "What do you remember about me?" It can export what it remembers to a JSON file in your OneDrive. Microsoft also documents a Windows script that downloads a full copy of its memory and sessions, which it stores in a Git repository file named lobster.repo in your OneDrive.

Four points from Microsoft's docs are easy to miss:

  • "Memory can't be turned off in this release."
  • "Your Microsoft 365 Copilot personalization settings don't apply to Autopilot."
  • "Deleting a conversation doesn't remove a memory already saved from it."
  • "Memory isn't a complete record of every conversation, and it doesn't grant access to sources or authorize new actions."

Personal memory vs team memory

Autopilot's memory is about you and your work. It is not a shared memory for your whole team. If your colleague's Autopilot learns that a client hates long emails, your Autopilot doesn't automatically know it. Workspaces gives a group shared files and context for one workstream, but there is no single company memory that every agent reads before it acts.

Vestra's handbook describes that idea as the shared brain: "one memory for the whole team," written as a wiki where every line has a source, so that "you correct one agent once, and the whole team stops making that mistake." What is live in Vestra today is "a memory that learns," a wiki of what your team knows with a source for every line, for one person's team. The company-wide shared brain is on Vestra's roadmap, after the network features now being built.

Autopilot vs Cowork vs Workspaces vs Agent 365: what is the difference?

Cowork does one multi-step task and hands back a finished result. Autopilot follows through on ongoing work for one person, in the background. Workspaces is a shared place where a group of people and agents work on one workstream. Agent 365 is the IT control plane for governing agents and their cost across a company.

Microsoft now sells a family of Copilot experiences that overlap. Here is how they fit together, using Microsoft's own descriptions.

The short version

  • Copilot Chat: questions and drafts in a conversation. Included in the Copilot licence ("everyday AI").
  • Copilot Cowork: "handles delegated, multi-step tasks." You ask, it delivers. Generally available since June 2026. Billed in Copilot Credits.
  • Copilot Code: "focuses on building software and solutions," like apps and dashboards. Microsoft said at Ignite it becomes generally available "in the coming weeks." Billed in Copilot Credits.
  • Copilot Autopilot: "persistent work and follow-through" for one person. Frontier preview. Billed in Copilot Credits.
  • Copilot Workspaces: "a shared place where you and your team can keep project information and files together and work with Copilot on an ongoing workstream." The evolution of Copilot Notebooks. Microsoft said at Ignite that the people-and-agents version "is now available in Frontier."
  • Today: a personal command centre across mail, calendar and Teams inside Home. In private preview.
  • Microsoft Agent 365: the control plane IT uses to see, govern and secure agents, now with cost management and business analytics.

Cowork vs Autopilot

Microsoft's product page puts it simply: "Use Cowork when you want a specific multi-step task done end to end: you ask, it delivers. Use Autopilot for work that doesn't end with one task: goals, decisions, and commitments that play out over days and weeks."

  • Shape of work: Cowork is one job with an end. Autopilot is ongoing.
  • Who starts it: you start each Cowork task. Autopilot can start work itself, from routines and what it notices.
  • Identity: Microsoft's Autopilot docs describe its own name, email address, mailbox, calendar and Entra Agent ID. Cowork runs as part of your Copilot.
  • Models: Cowork lets admins choose model families (OpenAI and Anthropic). Autopilot runs on one model with no choice.
  • Availability: Cowork is generally available. Autopilot is a Frontier preview.
  • Billing: both use Copilot Credits.

Workspaces vs Autopilot

Workspaces is Microsoft's multiplayer product. In the Ignite post's words, it is "a truly multiplayer experience, where people and agents drive towards a common goal, with shared context." An agent drafts a section, you review, you leave a comment, an agent picks it up, and "your Autopilot takes a task and checks with you before anything is final."

That is a real step towards agents and people working together. But notice the shape: in a Workspace, each person brings their own Autopilot. There is still one personal agent per person, all running on Microsoft's model, inside Microsoft 365. Microsoft's support article also draws a clear line: "Projects are for your own work. Workspaces are for group work."

Agent 365 vs Autopilot

Agent 365 isn't an agent you talk to. It is where IT governs agents: identity through Entra Agent ID, controls on tools such as MCP servers, plugins and skills, security for agents on devices, and now FinOps. At Ignite, Microsoft added Agent 365 Business Analytics (rolling out in Frontier Public) and Agent 365 Cost Management, where leaders "can set spending policies, limits, and proactive alerts and control which models are available." Autopilot is one of the agents Agent 365 governs.

At a glance

  • Chat: asks and answers · included in licence · generally available.
  • Cowork: one finished task · Copilot Credits · generally available.
  • Code: apps and dashboards · Copilot Credits · general availability "in the coming weeks."
  • Autopilot: ongoing follow-through for one person · Copilot Credits, final pricing not published · Frontier preview, US first.
  • Workspaces: group work with Copilot, and people with agents · rolling out since late October 2026; people-and-agents features in Frontier.
  • Agent 365: IT governance and cost control for agents · priced separately · expanded at Ignite.

For a longer side-by-side of Copilot as a whole against Vestra, see our page Vestra vs Microsoft Copilot.

Which AI model does Autopilot use?

Autopilot runs on OpenAI's GPT-6.1 Sol, which OpenAI provides as a Microsoft subprocessor. If an admin turns off OpenAI as a subprocessor, Autopilot uses GPT-5.6 Sol hosted in Azure OpenAI instead. Users can't switch models, and there's no model setting for Autopilot. A separate AI model reviews its actions.

That is a notable choice, because elsewhere Microsoft leans hard on model choice. At Ignite it said: "No single lab is the best at everything, and you shouldn't have to pick just one." Copilot Chat and Cowork let people pick OpenAI or Anthropic models, or use Auto routing. Autopilot, in this preview, does not.

For a business, the model matters less than it sounds for everyday follow-up work. It matters more when you want different jobs done by different engines: code written by one and reviewed by another, or writing done on the model that suits your voice. That is the idea behind Vestra's Bring Your Own Agents chapter: each teammate runs on the engine you pick, "teammate by teammate, so one office can use several at once." Live today in Vestra are its own engine, Claude Code and Codex; other names in the engine picker, such as Gemini, Grok and Muse, are not live yet. Our guide What is an AI agent engine? explains the difference between a model, an engine and an agent.

What are Autopilot's limits today?

Autopilot is a preview with clear limits: one per person, Frontier only, initially US only, optimised for English, one model with no choice, memory that can't be switched off, no credit counter for users, limited data loss prevention, no access to your local files, and no guarantee that background work finishes on time.

Microsoft lists most of these itself. Here they are in one place.

Access and setup

  • Frontier only. "Access to Autopilot (Frontier) requires the applicable Frontier preview program enrollment, preview assignment, and participation terms."
  • US first. "Availability is initially limited to eligible customers in the United States."
  • Off by default. An admin must licence, fund and activate it.
  • One per person. "Each enabled user can create one Autopilot." Starting over means exporting and asking your admin.

Work and reliability

  • No guaranteed timing. "Background operation doesn't guarantee uninterrupted execution or completion at a particular time."
  • Routines aren't real-time. They check sources only when they run.
  • No local files. It "can't access your local folders or installed apps."
  • Check results yourself. Microsoft repeats that "an approval authorizes an action; it isn't proof that the action completed." Open the file, check the calendar, confirm the sent email.

Money

  • Usage on top of the licence. "A Copilot license doesn't include Autopilot usage."
  • Silent stop. When credits run out, "Autopilot doesn't show you an explanation."
  • No user counter. Users can't see credits used or remaining.
  • Final pricing to come. Not published before general availability.

Data and governance

  • Memory always on. It can't be turned off in this release.
  • Labels aren't a wall. "Labeled email and files aren't excluded from what Autopilot can read, and content it generates might not inherit the label."
  • DLP is limited. "Data loss prevention (DLP) for prompts is limited during the preview."
  • Activity isn't a full audit log. Microsoft says the Activity view "isn't a guarantee of a complete administrative audit record"; ask your admin which audit tools cover it.
  • Blocking takes time. "Blocking a user can take about an hour to reach Autopilot, and work on the owner's behalf can continue longer, so don't rely on it for an immediate stop."
  • Not for high-stakes accuracy. Microsoft says it "isn't intended for use cases that require guaranteed accuracy without human review," such as legal filings, medical decisions or financial transactions that bypass approvals.

None of this is unusual for a first preview. It does mean Autopilot is something to pilot carefully, not something to hand your business to on day one.

Is Copilot Autopilot a good fit for a small business?

Autopilot fits a small business that already runs on Microsoft 365 Copilot, is in the US, is in the Frontier program and has an admin to set up billing and security. It fits less well if you aren't on Microsoft 365, are outside the US, want to start this week, or want agents that work together as a team.

When Autopilot makes sense

  • Your work already lives in Outlook, Teams, OneDrive and SharePoint. Autopilot's biggest strength is that it reads and acts across Microsoft 365 with Work IQ behind it. No other agent knows your Microsoft tenant as well.
  • You want an assistant for each person. If the job is "help each of my managers stay on top of their inbox, meetings and follow-ups," one personal agent per person is the right shape.
  • You need enterprise-grade identity and audit. Entra Agent ID, a named manager, sign-ins logged under the agent's name and admin blocking are strong controls.
  • You have IT support. Frontier enrolment, spending policies, activation and the recommended security stack all need someone who knows the Microsoft 365 admin center.
  • You can wait for general availability before you depend on it.

When it doesn't

  • You aren't on Microsoft 365 Copilot. Autopilot needs the licence first.
  • You are outside the US. It is US first, with no published dates elsewhere.
  • Your tools are mostly not Microsoft. If your business runs on Gmail, Slack, HubSpot, Shopify or Stripe, Autopilot can be extended with plugins, but it is built around Microsoft 365.
  • You need predictable costs. Final Autopilot pricing isn't published, and the agent stops silently when credits run out.
  • You want a team, not assistants. Autopilot gives each person one agent. It doesn't give a business a team of agents, each holding a whole job, that pass work to each other.

Questions to ask your Microsoft partner

  1. Is our tenant in Frontier, and is Autopilot in our preview assignment?
  2. Do our licences cover the required security pieces (Entra Agent ID, MXC through Intune Plan 1 or Microsoft 365 E3, audit logs)?
  3. What spending policy will cover Autopilot, and what monthly limit per person?
  4. Who gets alerts when a policy nears its limit, and who tells users when Autopilot stops?
  5. Which plugins are allowed, and which are blocked by policy?
  6. Is browser use in Edge turned on?
  7. What happens to each Autopilot and its memory when someone leaves?

Is one Autopilot per person the same as an AI team?

No. One Autopilot per person gives every employee a capable personal assistant, but those assistants don't form a team. Each one works for its own person, remembers its own person's work, runs Microsoft's one model and keeps its routines private. A team needs agents that hold different jobs, share one memory and hand work to each other.

This is the most important point in this guide, and it isn't really about Microsoft. It is about the shape of today's agents.

Every big agent is built for one person

Look at the always-on and coding agents that launched this year. Almost all of them have the same shape:

Each one is impressive on its own. Each one is built for one person, or for one maker's stack. None of them is built to work with the others.

What happens when a business runs five personal agents

Picture a ten-person company where each person has their own Autopilot, two engineers also use Claude Code and Codex, and the founder tries a different personal agent on her phone. Every agent is good. Together they make the people the glue:

  • Brief. Each agent only knows what its own person told it. The same client context is pasted into five places.
  • Carry. Work moves from one agent to another only when a person copies it across.
  • Chase. Every result needs a person to ask "is it done?" and then "is it right?"
  • Check. Each agent has its own approvals, its own log and its own bill.

Vestra's handbook puts it plainly: "Agents got good at doing work. Most of them still work alone: each one knows only what its own person told it and can't see what a colleague's agents are doing. So people have quietly become the glue between agents."

Doesn't Workspaces fix this?

Workspaces helps, inside Microsoft 365. People and agents can share files and context for a workstream, and your Autopilot can pick up tasks there. But the agents in a Workspace are still each person's one Autopilot, all on Microsoft's model, governed by Microsoft's tools. A Claude Code agent your engineer trusts, or a Codex reviewer, can't take a desk in that Workspace as an equal teammate.

What a team of agents needs

A real AI team needs five things that personal agents don't give you:

  1. Agents hired for whole jobs, not one assistant doing everything for one person.
  2. One memory that every agent reads before it acts.
  3. Hand-offs between agents without a person copying work across.
  4. One set of rules for every agent, whatever it runs on.
  5. One record of what every agent did, what it cost and who said yes.

That is the gap Vestra was built for. Vestra is the one office where agents on different engines work together as teammates, run by a chief that answers to you. Our pillar guide What is an AI office? explains the idea in full, and How to manage multiple AI agents without becoming their manager covers the day-to-day.

How does Vestra compare with Copilot Autopilot?

Autopilot gives each person one always-on Microsoft agent inside Microsoft 365. Vestra gives you a chief that runs a whole team of teammates, each hired for a job and running on the engine that suits it (Vestra's own, Claude Code or Codex), with caps, approvals and receipts, from $19 a month with nothing for IT to set up.

They solve different problems. Here is the honest side-by-side. Microsoft facts are from Microsoft's pages, checked on 21 November 2026; Vestra facts are from Vestra's handbook, with what is live today and what is still being built marked plainly.

The shape

  • Copilot Autopilot: one personal agent per person. Its routines are private. It talks to you in a private Teams chat.
  • Vestra: a chief that runs your agents, speaks for you and answers to you, plus teammates hired for whole jobs with goals. Live today: a chief and a team of teammates, with routines and a morning brief, for one person and their team of agents.

Engines and models

  • Autopilot: GPT-6.1 Sol, with no model choice.
  • Vestra: you pick the engine per teammate. Live today: Vestra's own engine (which works with the top models), Claude Code and Codex. Your chief suggests an engine when it hires and says why. Switch any time and the teammate keeps its job and memory. Every chief runs on Vestra's own engine.

Identity, phone and email

  • Autopilot: its own name, email address, mailbox, calendar and Entra Agent ID inside your Microsoft 365 tenant, with a named manager. Microsoft's pages we read don't describe phone calls from a number of its own.
  • Vestra: live today, your chief has its own number and its own email address, with phone calls from the Pro plan up. A phone and an inbox for every teammate is on the roadmap, after the network features.

Where it works

  • Autopilot: Microsoft 365 (Teams, Outlook, OneDrive, SharePoint, To Do, Planner), plus plugins. Reached in the Copilot app, Teams and Outlook, including on mobile.
  • Vestra: 2,000+ apps connected in a click, such as Gmail, Slack, HubSpot, Notion, Shopify and Stripe. Your chief is live in the Vestra app (web and Mac), Slack, Telegram and email. WhatsApp and iMessage are being built now.

Its computer

  • Autopilot: a Microsoft-managed cloud computer you don't watch directly; you follow its Activity view.
  • Vestra: live today, a cloud computer with a browser, files and a terminal. You can watch any agent work live and click Take over; when you hand it back, it carries on.

Asking first and spending

  • Autopilot: a second AI model checks sends, shares, deletes and changes against your request; matching actions run, the rest wait. Payments always wait. Spending is set by admins in monthly policies.
  • Vestra: live today, daily and per-task caps (up to $100 a day), your yes on money, and every credit itemised. The handbook's rules: spending over your limit, deleting something and contacting someone new always wait for a person's yes, and before any unusually big task (by default over 50 credits) your chief tells you the cost and asks.

Memory

  • Autopilot: conversation, personal and workspace memory for one person; can't be switched off in this release; exportable to JSON.
  • Vestra: live today, a memory that learns, kept as a wiki with a source for every line, which you can open and edit. The company-wide shared brain is on the roadmap.

Receipts

  • Autopilot: admins track Copilot Credit use in Cost management; users don't see a count of credits used or remaining.
  • Vestra: every piece of work has a receipt showing what it cost and which teammate spent it. One credit is one cent, the same on every engine.

Price

  • Autopilot: a Microsoft 365 Copilot licence (from $18 per user a month with the first-year add-on promotion, or $23.50 with Business Standard) plus Copilot Credits at $0.01 each. Final Autopilot pricing not yet published.
  • Vestra: Lite $19 a month ($17 yearly) with up to 3 teammates and 2,000 credits; Pro $97 ($87 yearly) with up to 20 teammates, 10,000 credits and phone calls; Max $197 ($177 yearly) with up to 50 teammates and 20,000 credits; Ultra $497 ($447 yearly) with up to 150 teammates and 50,000 credits. Unused credits roll over, and every new office starts with 1,000 free credits. In India, plans are billed in rupees with GST included.

Availability

  • Autopilot: Frontier preview, eligible US customers first, English-optimised, admin activation required.
  • Vestra: you sign up, name your chief and describe your work; your chief suggests your first teammates, you connect a few tools, "and work starts the same day."

What Vestra is still building

Being plain about this matters. Vestra's handbook describes a whole office. Its Where We Stand page lists what is live today for one person and their team of agents (above), and what is being built now: a chief for every person, chiefs that work together, the office you can walk around, meetings with agents keeping the notes, spaces, access control, every hand-off checked and written down, and WhatsApp and iMessage. After that come a phone and inbox for every agent, the vault, shared apps, the shared brain and the front door for clients.

So if you need one agent per employee across a 200-person company this month, inside Microsoft 365, Autopilot (once it reaches you) is built for that and Vestra's multi-person office is still being built. If you need a team of agents working together for you and your business now, on more than one maker's engine, that is what Vestra does today.

What does a week look like with each?

The best way to see the difference is a made-up week at one small business. In the Autopilot version, each person's agent helps that person. In the Vestra version, one chief runs a team that works together. The Vestra week uses only features that are live today for one person and their team.

Meet a made-up business: Kaveri Interiors, a six-person design studio in Austin that furnishes offices. The founder, Meera, runs sales and client work. The studio runs on Microsoft 365. This is an illustration, not a real customer.

The Autopilot week

Assume Kaveri is in Frontier and Meera's admin has activated Autopilot for all six people. Meera names hers "Sage."

  • Monday. Sage's 8 a.m. routine triages Meera's inbox into needs reply, needs decision, FYI and can wait, and drafts replies to routine requests. Meera approves four drafts from her phone.
  • Tuesday. A client asks for a revised quote. Meera asks Sage to draft it from last month's quote in SharePoint. Sage drafts it, but the supplier prices live in her purchasing lead's email, not hers, so Sage can't see them. Meera messages the purchasing lead, who asks his own Autopilot to pull the prices and pastes them into the Teams thread.
  • Wednesday. Sage's follow-up loop spots that a designer promised drawings by Tuesday. It drafts a nudge in Teams and Meera allows it. The designer's own Autopilot has its own list of that designer's commitments. The two agents don't compare notes.
  • Thursday. Meera wants a short market check on three rival studios' prices. Autopilot can prepare analysis from the sources it can reach, so she asks Sage, then checks the sources herself.
  • Friday. Sage writes Meera's weekly status report and saves it to OneDrive. The admin checks Cost management to see how many credits the six Autopilots used.

That is a good week. Meera saved hours. But notice who carried the work between agents: Meera and her colleagues, every time.

The Vestra week

Now the same studio on Vestra's Pro plan. Meera names her chief "Atlas" and hires four teammates in a sentence each: "Scout" for research on Vestra's own engine, "Quill" to write proposals and emails, "Ledger" to keep quotes and invoices straight, and "Hunter" to follow up with clients and suppliers. She connects the studio's email and calendar, HubSpot and Stripe (if you try this, check that your own tools are among the 2,000+ apps Vestra connects).

  • Monday. Atlas sends Meera's morning brief: what got done, the one thing that needs her, what it noticed. Hunter's routine has listed every quote older than seven days with no reply.
  • Tuesday. A client asks for a revised quote. Ledger builds it from the last quote and the supplier prices in the studio's connected tools, Quill writes the cover email in Meera's voice, and Atlas puts it in front of Meera for a yes before it goes. Nobody copies anything between agents.
  • Wednesday. Meera wants rival prices. Scout researches three rival studios on its cloud computer, with a source for every figure. Meera opens Scout's desk and watches it work for a minute, then lets it finish.
  • Thursday. Hunter wants to email a facilities manager the studio has never contacted. Under Meera's rules, first contact waits for her yes. She approves it in one tap. Later, Scout's next research task would cost more than Meera's per-task cap, so it stops and Atlas brings her the cost before anything else is spent.
  • Friday. Atlas calls Meera from its own number on her drive home (calls are live from Pro) and reads her the week's summary. Every piece of work has a receipt showing what it cost and which teammate spent it.

The difference isn't that one agent is smarter. It is that Vestra's agents are a team: each holds a job, they hand work to each other under one chief, and they work under one set of rules with one record.

What can you use today if you are not in Frontier?

If you aren't in Frontier, you can't use Autopilot yet. Inside Microsoft, the closest options are Copilot Chat and Copilot Cowork, which are generally available. Outside Microsoft, you can use another maker's personal agent, or a team platform like Vestra where agents on different engines work together.

Inside Microsoft 365

  • Copilot Chat for questions and drafts in Word, Excel, PowerPoint, Outlook and Teams, included in the Copilot licence.
  • Copilot Cowork for delegated, multi-step tasks, billed in Copilot Credits, with spending limits your admin sets. Microsoft says "nearly 80% of the Fortune 500 companies use Cowork."
  • Copilot Workspaces for shared project context with your team, rolling out since late October 2026.
  • Copilot Studio if you have someone to build custom agents.

Other personal agents

Most makers now offer a personal agent, each built for one person on its own maker's stack. We compare them with Vestra on their own pages: ChatGPT Work, Claude Code, Codex, Grok Bot, Meta Muse, OpenAI Dot and Perplexity Teammates. For business-focused assistants, see Lindy, Manus, Town, Viktor and Sintra. For self-hosted agents, see OpenClaw and Hermes Agent.

Our longer roundups cover the field: The best AI agents for business, Best AI employee platforms for small business and Perplexity Teammates alternatives.

A team, not another assistant

If what you actually want is the work done, not one more assistant to manage, look at a team. In Vestra you talk to one chief. It hires teammates for whole jobs, picks an engine for each, makes sure results are checked before they reach you, and only interrupts you for real decisions. You can still run engineers' teammates on the engines they already trust: our guide How to use Claude Code and Codex together shows how.

How do you set up Copilot Autopilot, step by step?

An admin enrols users in Frontier, creates a spending policy that covers Autopilot, and activates "Autopilot (Frontier)" in the Microsoft 365 admin center. Each user then opens Home in Microsoft Copilot, creates their Autopilot, picks a name and starter routines, waits 10 to 30 minutes for provisioning, and reviews the profile it builds.

These steps follow Microsoft Learn's admin and user guides as of 21 November 2026. Menus may change during the preview.

For the admin

  1. Confirm licences. Each intended user needs a Microsoft 365 Copilot licence and a work account. Check the required security capabilities (Entra Agent ID, MXC through Intune Plan 1 or Microsoft 365 E3, audit logs).
  2. Choose how to pay for credits. An Azure subscription for pay-as-you-go, or prepaid Copilot Credits.
  3. Enable Frontier. In the Microsoft 365 admin center, search Copilot settings for "Frontier", open Copilot Frontier, turn it on, and choose all users or specific groups. Changes can take a few hours to appear.
  4. Create a spending policy. Go to Copilot > Cost management > Configuration > Spending policies and add one that includes Autopilot, or let the activation wizard suggest one. Set monthly limits, per-user limits and alerts. Microsoft's guidance suggests at least 2,000 credits per user per month.
  5. Activate Autopilot. In Agents, search for Autopilot, select "Autopilot (Frontier)" and Activate. Choose who can create one and pick the spending policy. You don't assign a separate licence on this page.
  6. Check it works. Have a user sign in at copilot.cloud.microsoft and look for Autopilot on the Home tab.
  7. Decide on plugins and browser use. Plugins blocked by policy show as "Blocked by policy." Decide whether Autopilot may work in Microsoft Edge.
  8. Set up monitoring. Track Copilot Credits in Cost management and add alerts, because users won't see an explanation when credits run out.

For each user

  1. Open Home in the Microsoft Copilot app (web or desktop) and select Autopilot in the left navigation.
  2. Create your Autopilot. Choose a name and appearance; coworkers will see them in Copilot, Teams and Outlook.
  3. Pick starter routines. Accept or reject the routines it suggests.
  4. Wait for provisioning, 10 to 30 minutes. You'll get an email and a Teams message.
  5. Review its first profile of your people, responsibilities and recurring commitments. Accept it, or fix what's wrong.
  6. Run a safe first task. Microsoft suggests a bounded request that doesn't send anything, such as: "Summarize this week's launch discussion and draft an agenda for the review. Don't send messages or change calendar events."
  7. Check the Activity tab to see work in progress and anything that needs your input.

How do you write good instructions for an always-on agent?

Good instructions name the outcome, the sources, the schedule, where the result goes and what the agent must never do without you. Write them the way you would brief a new colleague, then check the first results and correct them. This works for Autopilot and for any always-on agent or AI teammate.

Microsoft's own advice is sound: "Describe the outcome you want, the sources to use, and limits on actions." Here is a simple pattern you can reuse with any agent.

The five-part brief

  1. Outcome. What does done look like? "A one-page status update every Thursday."
  2. Sources. Where should it look? "The project channel, my email from the client and the tracker."
  3. Schedule or trigger. When should it run? "Every Thursday at 3 p.m. Eastern," or "when a client replies."
  4. Destination. Where does the result go? "Save it to the Status reports folder and send me the link."
  5. Limits. What must it never do alone? "Don't send anything to the client. Don't change calendar events."

Examples

  • Weak: "Help with email."
  • Better: "Every weekday at 8 a.m., sort new email into needs my reply, needs my decision, FYI and can wait. Draft replies to routine requests. Don't send anything."
  • Weak: "Keep an eye on the Acme deal."
  • Better: "Check the Acme thread and our CRM every morning. Tell me if the deadline moves, a decision stalls for three days, or someone new joins the thread. Draft a nudge for anything overdue, but show it to me first."

Correct it like a person

The first week matters. Check the first few results against the sources. When something is wrong, say why, in one sentence: "Treat anything from the finance team as needs my decision." Microsoft's scenarios page suggests exactly this kind of correction. In Vestra, the handbook describes the same habit turning into a lesson: "Every answer you give and every correction you make becomes a lesson it keeps, so the same question doesn't come twice."

What should you check before you turn on any always-on agent?

Before you let any agent work in the background, check five things: what it can access, what it can do without asking, how much it can spend and what happens at the limit, how you see what it did, and how you stop it. If you can't answer all five in a sentence, you aren't ready to switch it on.

These checks apply to Autopilot, to other personal agents and to Vestra. Our guide Delegate work to AI without losing control goes deeper.

1. Access

  • Which mailboxes, calendars, chats, files and apps can it read?
  • Does it act as itself, as you, or both?
  • Can it read labelled or confidential files? (Autopilot can, per Microsoft.)

2. Freedom

  • What can it do without asking?
  • What always waits for a person, whatever the settings?
  • Can outside content (an email, a web page) tell it what to do?

3. Money

  • Is there a cap per task, per day and per month?
  • What happens when it hits the cap: does it stop cleanly and tell you?
  • Can you see what each piece of work cost?

4. Visibility

  • Can you see what it is doing right now?
  • Is there a record of what it did, in plain words?
  • Can you watch it work and take over?

5. Stopping

  • How fast does "stop" take effect? (Microsoft says blocking can take about an hour to reach Autopilot.)
  • What happens to its memory and files when someone leaves?
  • Can you export what it knows?

How Autopilot and Vestra answer the five

  • Access: Autopilot reads within Microsoft 365 permissions, sometimes as you and sometimes as itself. Vestra teammates use the business tools you connect, and you choose the apps each teammate may use. A sealed vault of logins for every agent is on Vestra's roadmap.
  • Freedom: Autopilot uses a second-model check; payments always wait. Vestra uses fixed rules plus freedom levels you set per teammate; your yes on money is live today.
  • Money: Autopilot uses admin spending policies with monthly limits; users can't see a running count. Vestra has per-task and daily caps and a receipt for every piece of work, live today.
  • Visibility: Autopilot has an Activity view, which Microsoft says isn't a complete audit record. Vestra lets you watch any agent's cloud computer live and take over, live today.
  • Stopping: Autopilot is blocked by an admin, with a delay. In Vestra, every agent works under the caps you set, stops at a sensible point when it reaches one, and you can change a teammate's apps or rules whenever you like.

FAQ

What is Microsoft Copilot Autopilot in simple words?

It is an AI assistant from Microsoft that keeps working for you in the background. You tell it what you want, like a daily brief or follow-ups on a project, and it does the work across Teams, Outlook and OneDrive, asking you before actions it isn't sure about. Each person can have one.

When was Copilot Autopilot announced?

Microsoft announced Autopilot at Microsoft Ignite on 17 November 2026 and said it was "starting to rollout to customers in Frontier today." Microsoft Learn's Autopilot pages were last updated on 17 November 2026.

Is Copilot Autopilot generally available?

No. It is a preview in Microsoft's Frontier program, initially for select eligible customers in the United States. Microsoft says "final pricing and packaging will be shared before general availability" but hasn't given a date.

How much does Copilot Autopilot cost?

You need a Microsoft 365 Copilot licence (business plans from $18 per user a month on a first-year promotion, or $23.50 with Business Standard, paid yearly), plus usage billed in Copilot Credits. Pay-as-you-go credits are $0.01 each. Microsoft hasn't published final Autopilot pricing.

Is Autopilot included in my Copilot licence?

No. Microsoft's requirements page says "a Copilot license doesn't include Autopilot usage." The licence gives you access; the work itself is billed in Copilot Credits through a spending policy your admin sets up.

What happens when Autopilot runs out of credits?

It stops running routines and monitoring and stops answering new requests, and it doesn't tell you why. Microsoft advises users to contact their admin if Autopilot stops responding. Admins can add alerts to spending policies so they hear before a limit is reached.

Can I have more than one Autopilot?

No. "Each enabled user can create one Autopilot." If you want several agents with different jobs, you need a different product. In Vestra, a chief runs a team of teammates: up to 3 on Lite, 20 on Pro, 50 on Max and 150 on Ultra.

Can I choose the AI model Autopilot uses?

No. Autopilot runs on OpenAI's GPT-6.1 Sol (or GPT-5.6 Sol in Azure OpenAI if your admin turns OpenAI off). Microsoft says "there's no separate model selection control for Autopilot." Copilot Chat and Cowork do offer model choice.

Does Autopilot send emails without asking?

It can. Before it sends, shares, deletes or changes something, a second AI model checks the action against your request. If it matches, it runs without a card; if not, it waits for you. Payments always need your approval. Add limits like "don't send anything" when you want a draft only.

Does Autopilot work outside the US?

Not yet. Microsoft says availability "is initially limited to eligible customers in the United States," with new Frontier tenants rolling out over time, and that Autopilot is optimised for English. No dates for other countries have been published.

What is the difference between Autopilot and Cowork?

Cowork does one multi-step task and returns a finished result. Autopilot handles ongoing work over days and weeks for one person, including routines it runs on its own. Cowork is generally available; Autopilot is a Frontier preview. Both are billed in Copilot Credits.

What is the difference between Autopilot and Workspaces?

Autopilot is your personal agent. Workspaces is a shared place where a group of people and agents work on one workstream with shared files and context. In a Workspace, your Autopilot can take a task and check with you before anything is final.

Is Copilot Autopilot the same as Windows Autopilot?

No. Windows Autopilot is an older IT tool for setting up and resetting Windows PCs. Copilot Autopilot is the new AI agent in Microsoft Copilot. Same maker, same word, different products.

Can Autopilot make phone calls?

The Microsoft pages we read describe Autopilot's email address, mailbox, calendar and Teams presence, but not phone calls from a number of its own. In Vestra, your chief has its own number and email address today, with phone calls on the Pro plan and up.

Is Copilot Autopilot an AI team?

No. It is one personal agent per person. Several people's Autopilots can join a Workspace, but each still works for its own person, on Microsoft's one model. An AI team, like the one Vestra's chief runs, has agents hired for different jobs, on different engines, working under one set of rules.

What is the best alternative to Copilot Autopilot for a small business?

If you are on Microsoft 365 but not in Frontier, start with Copilot Cowork. If you want agents that work together as a team, on more than one maker's engine, with caps, approvals and receipts and nothing for IT to set up, Vestra starts at $19 a month. Compare in detail at Vestra vs Microsoft Copilot.

Glossary

  • Agent: AI software that does work, not just answers questions: it plans steps, uses tools and acts.
  • Always-on agent: an agent that keeps working in the background, on a schedule or when something happens, without being prompted each time.
  • Autopilot (Frontier): Microsoft's name for Copilot Autopilot in its preview documentation.
  • Approval card: a message asking you to Allow or Deny an action an agent wants to take.
  • Accountable user: Microsoft's term for the person an Autopilot is assigned to.
  • Agent 365: Microsoft's control plane for governing, securing and managing the cost of agents.
  • Chief: in Vestra, the agent that runs your other agents, speaks for you and answers to you. You name it.
  • Copilot Credits: the unit Microsoft uses to bill advanced AI such as Cowork, Code and Autopilot. Pay-as-you-go credits cost $0.01 each.
  • Cowork: Microsoft's Copilot experience for delegated, multi-step tasks that return a finished result.
  • Engine: what an agent runs on: the part that thinks and works. In Vestra, you pick one per teammate.
  • Entra Agent ID: an identity in Microsoft Entra for an agent, so IT can manage its access like a user's.
  • Frontier: Microsoft's program for early access to preview features in Microsoft 365.
  • Prompt injection: hidden instructions in an email, file or web page that try to make an agent do something its owner didn't ask for.
  • Receipt: in Vestra, the record of what a piece of work cost, who did it and who said yes.
  • Routine: work an agent runs on a schedule or when something happens, such as a morning brief.
  • Spending policy: in Microsoft 365, the admin setting that grants access to usage-billed AI and sets monthly limits and alerts.
  • Teammate: in Vestra, an agent hired for a whole job with goals, shaped by you, on the engine you pick.
  • USL (user subscription license): Microsoft's per-user Copilot licence, which covers "everyday AI" and is required for usage-billed features.
  • Windows Autopilot: a separate, older Microsoft tool for setting up Windows devices. Not an AI agent.
  • Work IQ: Microsoft's layer that gives Copilot context about your people, projects and Microsoft 365 content.
  • Workspaces: Microsoft's shared place for group work with Copilot, where people and agents work on one workstream. Formerly Copilot Notebooks.

The bottom line

Copilot Autopilot is a serious step for Microsoft: a personal agent with its own identity, inbox, calendar and memory, working in the background across Microsoft 365, with a second model checking its actions. If your business runs on Microsoft 365 Copilot, is in the US and is in Frontier, it is worth a careful pilot.

But it is still one agent per person, on one maker's model, inside one maker's tenant. That is true of almost every big agent this year, from Claude Code and Codex to Grok Bot, Meta Muse and OpenAI Dots. Each is built for one person. Run several and you become the glue.

Vestra is built for the other shape: one office where a chief runs a team of teammates, each on the engine that suits its job, under one set of rules, with a receipt for every piece of work. It starts at $19 a month, with 1,000 free credits to try real work first.

Sources