Microsoft Ships Scout, Its First "Always-On" Agent for Microsoft 365
At Build on June 2, Microsoft unveiled Scout, an experimental Autopilot agent that runs continuously in the background of Teams, Outlook, OneDrive, and SharePoint without needing to be prompted. It is built on Microsoft's OpenClaw framework and requires Frontier-program enrollment plus both a Microsoft 365 Copilot and a GitHub Copilot license. The pitch is a category shift from assistant to operator. The reality is a narrow private preview with real governance questions.
Microsoft used the opening day of Build 2026 to introduce Scout, which it calls its first "Autopilot agent" — a new tier above Copilot that operates with its own Entra identity, stays resident across cloud and desktop, and takes action without a prompt. Scout connects to Teams, Outlook, OneDrive, SharePoint, calendar and contacts, and uses a "Work IQ" context layer to learn recurring patterns. Microsoft frames it as the jump from AI-as-assistant to AI-as-operator. The access list is narrow: Frontier-program enrollment, Intune policy, an opt-in attestation, plus both a Microsoft 365 Copilot and a GitHub Copilot license to install.
The framework underneath matters more than the demo. Scout runs on OpenClaw, Microsoft's open-source agent runtime (previously known as Clawdbot), with Entra identity governance and Purview compliance wrapped around it. Microsoft says it is contributing policy-conformance features upstream. The honest read is that Microsoft is doing what AWS, Google, and IBM are also doing — turning the agent runtime into a platform layer the enterprise has to buy into, then competing on governance rather than model quality. That is the same play NVIDIA is making at the silicon layer, and it is the move ServiceNow and NVIDIA telegraphed last month with Project Arc, a sandboxed desktop agent running inside NVIDIA's OpenShell runtime under ServiceNow's AI Control Tower.
The skepticism is warranted. Forrester's Jeff Pollard told Computerworld that Scout "amplifies whatever data governance problems already exist" — an agent acting on stale permissions does damage a chatbot answering a question never could. LLM agents still drift on multi-step goals and remain exposed to prompt injection, and OpenClaw itself has already drawn security scrutiny. Microsoft's answer is the human-in-the-loop sign-off for sensitive actions, which is the right answer and also the answer that quietly admits the autonomy is conditional. Pricing inside or alongside the $30/month Copilot SKU is still undecided.
What actually shipped on June 2 is a private preview with a license stack, not a general-availability agent. What shipped conceptually is bigger: Microsoft has formally declared that the unit of automation is no longer a chat turn but a persistent identity with goals. Everyone selling enterprise software now has to answer the same question — what is the governance surface for an entity that works while you sleep, and who is liable when it ships the wrong file to the wrong tenant. Scout is the first product to force that question into a procurement meeting.
Microsoft 365 Blog — Introducing Microsoft Scout → · Computerworld — Microsoft unveils Scout, an autonomous AI agent built on OpenClaw → · NVIDIA Blog — NVIDIA and ServiceNow Partner on New Autonomous AI Agents for Enterprises →