Microsoft has 30 million paid Microsoft 365 Copilot seats. That sounds huge until you notice it is only about 7% of the company’s 450 million commercial Microsoft 365 seats. On September 25, 2026, Microsoft made its biggest move yet to change that number: a rebuilt Copilot app with a new feature called Microsoft Copilot Autopilot, which lets anyone create a named, always-on AI agent that keeps working after they log off.

This is not another chatbot upgrade. Each Autopilot gets its own identity, memory and email address, runs in the cloud inside a company’s Microsoft 365 tenant, and can be tagged in Teams or Outlook like a coworker. Satya Nadella called the new Copilot a “new OS for work.”

In this article, you will learn what Microsoft actually shipped, how Autopilot agents work in practice, what the new usage-based pricing means for your budget, and how to prepare your team before the private preview opens at the end of September.

What Is Microsoft Copilot Autopilot? Inside the Copilot Super App

The new Copilot app, widely described as a “super app,” is split into three sections. Home combines chat with Microsoft’s Cowork agent and now runs full versions of Word, Excel and PowerPoint inside Copilot, not a simplified viewer. Code lets non-developers describe an app, dashboard or internal tool and have Copilot build it, using the same underlying technology as GitHub Copilot. Autopilot is where people set up their own persistent, proactive agents.

Autopilot has a history. In June, Microsoft introduced Scout, an always-on work agent that we covered in our look at always-on AI agents and ambient automation. Nadella later told investors Scout was powered by OpenClaw, the open-source personal agent he had earlier compared to “a virus” from a security standpoint. The new Copilot drops the Scout name entirely. Instead, users name their own Autopilots, assign each one a role and a goal, and let it run.

The launch came with less than a week to spare on Nadella’s promise that the super app would arrive in the quarter ending September 30. Most features are still in early access: Code is rolling out more widely to Microsoft’s Frontier program, Office in Copilot reaches consumers and Frontier customers in the coming weeks, and Autopilot expands to private preview at the end of the month.

How Always-On AI Agents Like Autopilot Work in Practice

The clearest picture of Autopilot came from Microsoft’s own demo, reported by GeekWire. An employee playing a regional retail manager created an Autopilot named Dot and gave it a single goal: make sure stores had inventory in time for Black Friday. Dot decided on its own which Teams channels, email threads and inventory data to monitor. It sent a daily summary and flagged a shipment problem that could affect 18 stores.

That pattern is what separates always-on AI agents from assistants. You do not prompt them task by task. You hand them an outcome and they choose what to watch. According to Fortune’s coverage of the launch, typical Autopilot jobs include project updates, tracking decisions and deadlines, reminding colleagues when commitments are due, gathering prep materials and reaching out to partners.

Microsoft also leaned hard on model choice. A demo menu let users pick OpenAI’s GPT, Anthropic’s Opus, or an Auto mode that routes each request to a model automatically. Nadella framed this as independence: companies should keep a “private eval” they can use to test and swap models, instead of handing their accumulated know-how to a single AI lab.

Jacob Andreou, the executive vice president who leads Copilot, summed up the tension neatly: the autonomy that makes a long-running agent useful is “the same thing that makes them absolutely terrifying to an IT admin.”

How Businesses Should Prepare for Microsoft Copilot Autopilot

Microsoft’s answer to that fear is governance built into the platform. Autopilots run in the cloud inside the customer’s own Microsoft 365 environment rather than on a local machine, carry explicit permissions, and produce full audit logs. Humans decide how much autonomy each agent gets. Apps built with Code run in isolated local or cloud environments.

Pricing is the other big shift. Copilot now has two ways to pay. A per-user monthly license covers “everyday AI” such as chat and Copilot inside Word, Excel, Outlook and Teams. Cowork, Code, Autopilot and the newest frontier models are billed on usage. Usage-based services are off by default, and admins can set spending limits for the whole company, specific groups or individuals through Agent 365.

Here is a practical checklist before the preview reaches you:

1. Pick one outcome, not ten tasks. The Dot demo worked because the goal was concrete and measurable. Start with a single recurring outcome such as weekly vendor follow-ups or deadline tracking.
2. Scope data access first. Decide which channels, mailboxes and files an Autopilot may read before anyone creates one. Our guide to AI agent governance strategy covers proportional controls by autonomy level.
3. Set spending caps on day one. Always-on agents consume tokens around the clock. Read our breakdown of why AI agent costs keep rising and configure group-level limits before you switch usage billing on.
4. Build your private eval. Collect 20 to 50 real tasks with known good answers so you can compare GPT, Opus and Auto mode on your own work.

The Bigger Picture: Usage-Based Pricing and the Agent Platform War

Microsoft is not launching into empty space. OpenAI unveiled its own super app in July, and Anthropic has been selling Claude Cowork and coding tools directly to the businesses Microsoft has served for decades. Andreou took a pointed swipe at approaches that send “your most sensitive files down to a local machine into a VM,” calling that a nonstarter for enterprises. Expect that security argument to be Microsoft’s core pitch to IT buyers.

There is a reasonable counterpoint, though. Every major platform now claims permissions, audit logs and identity controls. What will actually differentiate them is whether agents like Autopilot deliver measurable value without surprise bills. Moving from per-seat to “per-seat-plus-consumption,” as Nadella put it, shifts cost risk onto customers. Teams that deploy dozens of always-on agents without clear goals could see budgets move quickly.

Broader availability details are expected at Microsoft Ignite, November 17 to 20 in San Francisco. Until then, treat the preview as a pilot, not a rollout.

Conclusion: Is Your Team Ready for Autopilot?

Three takeaways stand out. First, Microsoft Copilot Autopilot turns agents into named digital coworkers with their own identity, memory and inbox, working toward goals rather than prompts. Second, the new pricing splits everyday AI (per seat) from agentic work (per use), so spending controls matter as much as features. Third, model choice plus a private eval gives businesses leverage against lock-in to any single AI lab.

For more practical guides on AI agent tools, platforms and strategy, explore BigAIAgent and stay ahead of what ships next.

If you could name one Autopilot tomorrow and give it a single goal, what would you have it watch for you?