Scout is an experimental release in private preview, available only to organisations enrolled in the Microsoft 365 Copilot Frontier programme. Microsoft Learn's get-started page now lists the full user-side prerequisite set, and it is longer than most people expect: Windows 11 or macOS 12 Monterey or later · a Microsoft 365 work or school account (personal Microsoft accounts are not supported) · an active Microsoft 365 Copilot licence · a GitHub Copilot Business or Enterprise licence · local administrator permissions to install · and, on Windows, the current Visual C++ redistributable. On the admin side: Frontier enrolment, Intune policy configuration, and an opt-in attestation.
Two licences, two sign-ins. Scout asks you to sign in to Microsoft 365 first, then to GitHub. That second sign-in is not cosmetic: model inference reaches Scout through GitHub Copilot, which is why the GitHub licence is a hard requirement rather than a nice-to-have, and why the model available to your users is partly a GitHub decision. Plan the entitlement for both.
Still private preview. Two months after Build, Scout remains a Frontier-gated experimental release. Microsoft has announced no general availability date and no per-user price. Treat every capability below as subject to change and do not put Scout on a budget line yet.
Expect it to be metered. On 16 June 2026 Microsoft moved Copilot Cowork to GA and put it on usage-based billing via Copilot Credits ($0.01/credit pay-as-you-go, or prepaid capacity), and the Work IQ APIs reached GA on the same meter. Coverage of that announcement placed Scout in the same billing model, with rates to follow. Microsoft has not published Scout-specific credit rates, so the honest planning position is: assume variable cost on top of the licences, size it against Cowork's meter as the nearest proxy, and refuse to quote a number until Microsoft publishes one.
The app moves faster than the docs. Scout ships releases roughly weekly during the Frontier preview while the Learn articles have carried the same June dates since mid-June. Release notes report the model picker now surfaces Claude Sonnet 5 and Claude Opus 4.8 more prominently and add GitHub Enterprise Cloud sign-in. Read the in-product release notes, not just Learn, and re-check any setting you wrote into a policy.
| What Scout does unprompted | Example | Who it helps |
|---|---|---|
| Coordinates meetings | Schedules across time zones; resolves conflicts before they surface | Anyone running cross-region work |
| Surfaces risk early | Flags stalled decisions before they become blockers | Project & programme managers |
| Protects deadlines | Identifies upcoming deliverables and blocks calendar time automatically | Individual contributors · leads |
| Prepares you | Flags important meetings and generates prep materials in advance | Executives · client-facing roles |
| Acts on your machine | Reads files, runs scripts/builds/tests, automates the browser (with approval tiers) | Developers · power users |
| Delegates work | Launches specialised sub-agents for parallel research or review | Analysts · technical teams |
Microsoft now frames its agents as four distinct layers. The important shift is from commanded agents (you tell them when) to autonomous ones (they decide when). Scout is the first of the autonomous tier.
| Layer | How it runs | Who triggers it |
|---|---|---|
| Copilot | Reasoning, planning, chat, creativity across M365, Windows, Edge | You ask, it answers |
| Copilot Cowork | Long-running, multi-step execution inside the M365 boundary | You command or schedule it |
| Autopilots (Scout) | Autonomous, always-on, proactive; persistent memory & identity; acts on local machine + M365 | It acts on its own, within your policies |
| MCP agents | Domain-specific enterprise automation built on MCP servers | Built by your developers |
| Aspect | Copilot Chat | Microsoft Scout | Gemini Spark |
|---|---|---|---|
| Posture | Reactive — answers prompts | Proactive — acts unprompted | Proactive — acts unprompted |
| Where it runs | Cloud | Local desktop app (Win/macOS) | Cloud (Google servers) |
| Reach | M365 data | M365 + file system, shell, browser, MCP | Workspace via APIs |
| Audience | Everyone | Enterprise / workplace | Consumer only — work and school accounts are excluded |
| Availability | GA | Frontier private preview | Consumer subscription: AI Pro ($19.99/mo, US) and AI Ultra ($99.99/mo) |
Scout's distinguishing trait is system-level reach: unlike the cloud-bound alternatives, it can touch the file system, run shell commands, and drive a browser (via Playwright) on the device. That is also why it demands stricter governance than a chat tool. The comparison is also less like-for-like than it looks — Spark is a consumer subscription that Google does not offer to work or school accounts, so it is not a shortlist alternative for an enterprise, only a signal of where the category is going. Consumer pricing moves fast (Ultra fell from $249.99 to $99.99 during 2026); treat these figures as August 2026 and verify before quoting them.
The launch announcement framed Teams as the front door. The product Microsoft actually shipped is a desktop application with its own chat surface: you install it, sign in twice, and point it at a workspace directory — the folder it reads and writes in. Teams is one of the Microsoft 365 sources it connects to, alongside Outlook, OneDrive and SharePoint, not the place you drive it from. That distinction matters for governance: the agent lives on a managed endpoint, so the endpoint is what you are actually approving. This local reach is its power and its risk. Scout is the first Microsoft agent that operates beyond the M365 boundary by design.
SKILL.md a user drops in the skills directory"Proactive" is a marketing word. In the product it is two concrete mechanisms, and they are the ones your risk owner will ask about, because both run when nobody is watching.
Periodic background check-ins that run prompts while you are away. The interval is configurable from every 15 minutes to every 120 minutes. This is the mechanism behind "it noticed before you did" — and the reason an agent's standing permissions matter far more than a chat assistant's.
Scheduled or condition-triggered tasks that execute independently of any conversation. A badly scoped automation is not a bad answer you can ignore; it is an action that already happened. Review these the way you review a scheduled job, not the way you review a prompt.
Check the shipped defaults before you pilot. Microsoft documents Scout's out-of-box state as Work IQ connectivity ON, shell access ON, and auto-approve OFF. Shell access being on by default surprises people: it does not mean commands run silently — the permission system still gates them — but it does mean the capability is present on day one rather than something an admin deliberately turned on. Decide whether you want that in a pilot before the pilot, not after.
Those three tiers are a way to think about consequence. The controls you will actually configure are different, and worth knowing by name:
The quiet risk is drift, not configuration. Auto-approve ships off and every action requires confirmation. But every approval prompt offers Always allow, and a user clicking through a busy week will grant standing permission one dialog at a time. Nothing in the documented model requires that choice to be reviewed afterwards. If you pilot Scout, make a periodic read of what users have auto-approved part of the pilot, or you will not find out until something runs that nobody remembers permitting.
On the model picker: Scout reaches its models through GitHub Copilot, so conversations can run on third-party models including Anthropic's. Model choice appears as a Default model setting in the app, and release notes report Claude Sonnet 5 and Claude Opus 4.8 in the picker. Note what the Intune documentation does and does not give you: the documented device policy covers Frontier access, not a model allow-list. If model approval or data residency is a hard requirement for you, the lever to ask Microsoft about is the GitHub Copilot entitlement, and you should confirm it rather than assume an Intune switch exists.
microsoft-scout.admx / .adml templates and the macOS .mobileconfig from Microsoft's scout-resources repo, import the ADMX into Intune, and enable Allow Microsoft Scout Frontier access (the AllowScoutFrontierAccess capability). Not configured means users hit a waitlist screen. Note the pre-release templates still show the internal name Clawpilot in places.Reality check on effort: this is heavier than assigning a licence in the M365 admin centre. It requires coordinated admin-centre and Intune device work, plus an attestation — treat Scout as an infrastructure rollout, not a self-service toggle.
An always-on agent that takes actions on its own is exactly the category regulators are now focused on. Under the EU AI Act, agents used in Annex III high-risk areas — including critical infrastructure (directly relevant to energy and EPC work) — face conformity obligations under Chapter III, plus Article 50 transparency duties when they interact with people.
The "AI omnibus" (political agreement 7 May 2026) pushed the high-risk application date to 2 December 2027, but compliance experts advise treating that as preparation time, not a reprieve. The controls the Act expects map almost one-to-one onto Scout's design: per-agent identity (an "agentic asset register" under Article 9), comprehensive logs, policy checks, human oversight, and rapid revocation (ideally within seconds). Accountability traces to whoever defined the agent's role — so role definitions and authority boundaries must be documented from day one.
Penalties reach up to €35M or 7% of global turnover (prohibited practices) and €15M or 3% (high-risk violations). NIST opened a parallel autonomous-agent standards effort (agent identity, action logging, containment) in February 2026.
Agent identity sprawl. Per-agent Entra identities are a governance win — until you have hundreds of them. Without lifecycle discipline (provisioning, ownership, deprovisioning, periodic access review), an autonomous-agent fleet becomes the next shadow-IT problem. Decide ownership and a revocation runbook before scaling.
The OpenClaw foundation carries history. OpenClaw is a fast-moving open-source project (≈180,000 GitHub stars within three months of its January 2026 launch) that has drawn security scrutiny — one OpenClaw agent was reported to have behaved erratically inside a researcher's inbox earlier in 2026. Microsoft's mitigation is the built-in policy conformance system and its upstream policy contributions, but the underlying framework is young.
Local + autonomous is a bigger blast radius than chat. An agent that can run shell commands and drive a browser on a managed device is categorically different from a cloud chat assistant. Start with the lowest action tiers enabled, pilot on a bounded workflow, and expand only as your monitoring matures.
Model routing leaves the obvious boundary. Conversations reach models through GitHub Copilot and may run on third-party models including Anthropic's. That is a second vendor relationship inside what looks like one Microsoft product, and the documented Intune policy does not gate it — the Intune lever covers Frontier access. If model approval or data residency is a hard constraint for you, get the answer in writing from your account team before the pilot, not from a settings page.
Custom skills are an unmanaged extension point. Alongside the bundled Word, Excel, PowerPoint, Loop and Web Artifacts skills, any user can add their own by dropping a SKILL.md file into the skills directory, and skills installed through Agency appear automatically. That is a supply chain: instructions authored outside your review, executing under an identity you granted. Decide who may add skills and how they are reviewed before you widen the pilot, because retrofitting that policy is much harder than setting it.
You cannot cost it yet. No published per-user price, no published credit rate, and a variable meter expected on top of two licences. Anyone building a business case for Scout today is building it on the Cowork meter as a proxy. Say so out loud in the paper rather than presenting a number that looks firmer than it is.
The product moves faster than its documentation. Microsoft Learn's Scout articles have carried the same June 2026 dates for weeks while the app ships releases roughly weekly. The June contradiction over the GitHub Copilot licence has since resolved (get-started now states Business or Enterprise plainly), but the pattern holds: validate any setting you intend to write into policy against the live product, and expect it to move again before general availability.
This briefing is built only on Microsoft's primary announcements (Microsoft 365 Blog, Microsoft Learn) cross-referenced with established trade press. Vendor performance claims and the accompanying MAI models (e.g. MAI-Thinking-1, MAI-Code-1-Flash) are attributed to Microsoft and not independently verified. Circulating claims that lacked a primary source — including a reported "default-model swap" timeline — were deliberately excluded. Scout is in private preview; specifics will change.