At Microsoft Ignite in Chicago on November 17, 2026, Satya Nadella told the keynote crowd that “Copilot is the new operating system for work.” The line was a pitch, but the product behind it is specific: Autopilot, an always-on Copilot agent that gets its own Microsoft Entra Agent ID, its own memory, and a named human manager, and that began rolling out that same day to Microsoft’s “Frontier” early-access customers. It’s the clearest sign yet that Microsoft wants Copilot’s 30 million paid seats to stop being a chat window people type into and start being a roster of digital coworkers that run tasks on their own.
What actually shipped on November 17
Autopilot isn’t brand new — Microsoft previewed it in June under the name “Scout” inside Microsoft 365 Copilot, then renamed and repositioned it on September 25, 2026 as one of three modes in a redesigned Copilot app, alongside Home (a merged chat-and-Cowork view) and Code (an app- and dashboard-builder). What’s new as of Ignite is that Autopilot is rolling out — still restricted, still Frontier-only, with no general-availability date committed — and that Microsoft used the keynote to spell out exactly how it’s supposed to work: a user names the agent, gives it a role and a goal, and lets it run continuously on things like tracking project deadlines, drafting status updates, nudging colleagues, prepping meeting materials, and following up with partners by email.
Alongside Autopilot, Microsoft announced that Copilot Home and a “Copilot Managed Runtime” for hosting agent-built apps are now generally available, that Copilot Code reaches general availability “in the coming weeks,” and that Copilot Workspaces — a shared project space where employees and multiple agents collaborate side by side — is live for Frontier customers now. A new “Microsoft IQ” layer (bundling Work IQ, Fabric IQ, Foundry IQ, and Web IQ) is meant to decide, task by task, which of a company’s data actually matters, and an “Auto” routing feature picks which underlying model handles a given request. Microsoft also said employees can manually choose OpenAI or Anthropic models inside Copilot where their organization’s policy allows it — an unusually blunt admission, from a company selling its own agent platform, that customers still want rival models on tap.
The part that’s actually new: agents with a manager
The detail worth sitting with is identity. Microsoft says every Autopilot instance is issued its own Entra Agent ID — the same identity infrastructure Microsoft uses for human employee accounts — separate from the permissions of whoever created it, with its own audit trail. Each one is also assigned a named human manager inside the organization, and IT administrators can scope what it’s allowed to touch or block it outright through Agent 365, Microsoft’s governance layer for AI agents. Sensitive actions can be configured to require a human sign-off before they execute. Nadella framed the underlying goal at Ignite as governability: “Everything that you observe should be governable by policy,” he told the keynote audience, arguing that the real constraint on enterprise AI adoption isn’t model capability but whether a company can track and control what its agents are doing.
That’s a different problem than the one OpenAI tackled on November 14, when it let companies build and publish branded support agents inside ChatGPT for customers to talk to. Microsoft’s Autopilot isn’t customer-facing — it’s an internal employee’s assistant, built to be treated organizationally like a junior staffer: given a name, a job, a manager, and a paper trail. Annie Pearl, Microsoft’s corporate vice president for Copilot, put the ambition plainly in comments to reporters covering the Ignite drop: “We do want this to become an all-in-one productivity suite.” Omar Shahine, who runs the Scout team that built Autopilot’s predecessor, described the bet behind it back in June: “Most systems still stop at answering the question. The real unlock is in the follow-through.”
Numbers Microsoft is leaning on
Microsoft’s pitch rests heavily on scale and cost. The company says Microsoft 365 Copilot has passed 30 million paid seats, with net new seat additions reportedly doubling quarter over quarter, and that nearly 80% of the Fortune 500 now use Copilot Cowork, its agent-assisted collaboration mode — Eli Lilly alone, Microsoft said, has more than 5,000 employees using Cowork monthly. On pricing, Microsoft cited a GigaOm report claiming that for comparable output quality, Copilot runs 84% cheaper than Claude and 63% cheaper than ChatGPT for an “everyday” user, narrowing to 68% and 35% cheaper, respectively, for “advanced” workloads — figures Microsoft is using to argue its per-seat-plus-usage pricing beats rivals’ consumption-based models, though it’s Microsoft’s own commissioned comparison, not an independent benchmark. The company also disclosed new plugin counts for its Dynamics 365 and Salesforce integrations inside Copilot: more than 100 prebuilt skills across sales, customer service, and finance, including a new 38-skill Salesforce plugin.
Why this is landing now
Microsoft is moving into a market that’s already skeptical of agent follow-through: Gartner has forecast that 40% of enterprise applications will integrate AI agents by the end of 2026, while research cited alongside Ignite coverage from IDC and Lenovo put the failure rate of custom-built enterprise agents at 88%. That gap — widespread intent to deploy agents, against a poor track record of the ones companies build themselves — is the opening Microsoft is aiming Autopilot at: a pre-built, identity-governed agent that IT can audit and switch off, rather than a bespoke bot a team cobbled together and then lost track of. Whether a named Entra ID and an assigned manager actually solve the reliability problem, rather than just making it easier to see when an agent goes wrong, is the thing Frontier customers will be testing over the next few months — Microsoft has given itself no public deadline to find out.

