The headline is bigger than the reality right now. Nobody can buy any of these as a generally available enterprise product. Availability, billing and admin controls differ a lot from one vendor to the next. For Windows and Microsoft 365 shops, the useful questions are practical: who can turn these agents on, what can they reach, and who is accountable when one sends the wrong email?
Chat vs. persistent agents
Nivinskus explains the difference simply. In a chat, you ask and the model answers. An agent takes an assignment and uses tools to carry out the steps. A persistent agent also keeps the assignment's context and picks the work back up without another prompt. Depending on the product and its permissions, it can check for updates, follow up with people, or resume a task days later.
Her example comes from communications work. A product announcement is waiting on comments from legal, an executive and a subject-matter expert. With approved access and instructions, an agent could:
- track feedback as it comes in
- flag edits that conflict with each other
- send reminders to reviewers it has been authorized to contact
- propose a new review schedule if the launch date moves
The human team still decides what gets published. That's a possible workflow, not a tested result, and she frames it that way.
Section summary: A persistent agent works like a chat session that remembers the assignment and keeps acting on it. The value depends on what the agent can reach, and the risk does too.
Microsoft: Copilot Autopilot (not the Intune one)
Microsoft announced a redesigned Copilot on September 25, 2026, built around three areas: Home, Code and Autopilot. In its official blog post, Microsoft calls Autopilot a "persistent, proactive and personal agent." It says Autopilot was formerly called Scout and is set up with a name, a role and a goal. From there it watches channels, follows up on threads, runs recurring work and resumes projects days later without waiting for a prompt. Microsoft's own example is running a full supplier review, from building the schedule and workback plan through prep, meetings and chasing stakeholders for updates.
Microsoft says Autopilot is cloud-hosted and lives in the customer's tenant with its own identity, memory, computer and workspace. It appears in Teams, Outlook, chats, channels and documents, where people can @mention it, with "permissions, audit and governance" behind it. These are Microsoft's descriptions of its own product. A WindowsForum megathread notes that these controls are Microsoft's claims about its own product, and no independent assessment of Autopilot's containment has been published.
Availability: Autopilot is a private preview. No general availability date was given. Home and Code are rolling out through Microsoft's Frontier early-access program. Nivinskus says Autopilot is limited to selected Copilot customers for testing, which matches what Microsoft announced.
A naming trap for Windows admins: WindowsForum's earlier coverage cites IT-Connect on this point: Copilot's Autopilot has nothing to do with Windows Autopilot, the device-provisioning service many of them already use through Intune. Expect confused help-desk tickets and a few awkward change-advisory meetings.
The identity question: Autopilot isn't completely new. Scout was presented in June as a desktop app for Windows and macOS, while Autopilot is a cloud-hosted digital teammate with its own identity, memory and workspace inside your tenant. The June Scout preview required Frontier enrollment, Intune policy configuration, an opt-in attestation, and a GitHub Copilot license to install the desktop experience. Microsoft hasn't said whether the Autopilot preview keeps those requirements, so confirm eligibility before planning a pilot. Microsoft also does not spell out a mailbox, calendar, or license assignment for that particular private-preview experience.
The billing question: This is the part finance will notice first. Microsoft's post separates two kinds of spend. Everyday Copilot use falls under a fixed user subscription license (USL). Cowork, Code, Autopilot and other long-running agentic work run on usage-based billing (UBB). Microsoft also announced FinOps for AI controls. Admins can manage spending policies through an API, send end users' credit requests into existing approval workflows, and choose which model families each user group can access. End users can see their own credit balance and usage history inside Copilot.
Section summary: Autopilot is a tenant-resident, metered, private-preview agent with its own identity. Treat it as a new kind of directory object with a usage bill attached, not as a Copilot feature you flip on.
OpenAI: dots
OpenAI announced dots on September 29, 2026. It describes them as always-on agents with their own cloud computer and browser, connected to apps through plugins. Users can reach a dot through ChatGPT, Slack or Teams. OpenAI says dots run on GPT‑6 Astra and connect to more than 4,000 apps. Those are company claims.
Rollout starts with Pro and Business Premium users in eligible markets. Enterprise customers, including Edu and Healthcare, can try the beta only after a workspace admin enables it. That matches Nivinskus's point that admins have to switch access on.
OpenAI's Help Center has more detail for admins than most launch posts:
- Dots are off by default for Enterprise. Workspace owners go to Workspace settings > Permissions & roles, then Permissions > Workspace default > Workspace capabilities, and enable "Use dots (Beta)." Then they check any custom roles that apply to the target members.
- Each capability has its own switch. Separate settings control adding dots to Slack and Microsoft Teams, local computer access, and custom rules. Local access and custom rules are also off by default for Enterprise. Cloud browser use, cloud network access, cloud computer use and the password manager are controlled separately again.
- Turning on the Slack/Teams permission doesn't install anything. Members still have to complete setup, and a Slack workspace may require approval from its owner or an app manager.
- Turning off access takes more than one click. Admins have to remove every grant of "Use dots (Beta)," including grants that come through group-assigned roles. If any one role still grants access, the member keeps it.
- Workspace model controls don't apply. OpenAI says Enterprise model controls and default model settings do not govern dots.
Two more details matter for Windows shops. A dot's cloud computer does not inherit the user's local VPN, browser sign-ins or device policies, so your conditional-access setup doesn't automatically carry over. Local computer access requires the member to connect their PC, keep it online and keep the ChatGPT app open. That access is unavailable in workspaces where Codex or ChatGPT Work policies target a specific operating system.
OpenAI also says it is working with Microsoft to bring "specialist dots" under Microsoft Agent 365 governance and security controls. That integration is planned, not shipped.
Section summary: Admins switch dots on, and the defaults are cautious. The permission model is spread across a lot of separate settings, and some of the usual Enterprise guardrails don't apply.
Meta: Muse
Meta introduced Muse on September 8, 2026 as a personal AI agent for consumers. On September 29 it followed with Muse for Small Business. In that post, Meta describes Muse as available in the US and Canada and lists connectors including Asana, Box, Canva, Dropbox, QuickBooks, Shopify, Slack, Stripe and Zoom, plus Facebook and Instagram business accounts. Meta's promise: "nothing publishes, sends, or spends without your approval." Muse is free for most uses, with subscription plans for heavier use.
As Nivinskus points out, a consumer launch doesn't make Muse an enterprise product. A small-business connector list isn't tenant governance either. It's a useful comparison, not an option for your Entra-managed fleet.
Is "keeps working while you sleep" a selling point or a risk?
Both. The appeal is clear. Follow-up work is tedious and humans are bad at it. But every vendor here is selling autonomy and promising control in the same pitch, and only one of those is easy to verify. OpenAI's own launch post says dots can still make mistakes and tells users to review consequential work. An agent with its own identity, persistent memory and permission to message your stakeholders is effectively a new employee, and nobody has checked its references yet.
How to scope a pilot with IT
Nivinskus's guidance is sensible and doesn't depend on the vendor. Here it is with the admin details from Microsoft's and OpenAI's documentation added:
- Fix the workflow first. Find where a recurring task stalls and who reviews and approves it. Automating a broken process just gets you to the broken part faster.
- Separate tracking from authority. Monitoring approvals does not authorize the agent to accept edits, change deadlines or publish. Write that down.
- Inventory the agent's identity and reach. For Autopilot, confirm preview eligibility and how the agent identity shows up in your tenant. For dots, review every capability switch, connected-app authorization and custom-role grant.
- Budget for the meter. Autopilot runs on usage-based billing, so set spending policies before the pilot starts, not after the first invoice.
- Name an owner. One communications or business owner supervises the agent's work. IT owns access and handles incident response.
- Measure the supervision cost. Track accuracy, follow-through and missed escalations, plus the staff hours spent checking and correcting the agent's output. If supervising it takes as long as doing the work, you've hired a very expensive intern.
Bottom line
September's launches mark a real shift from asking AI questions to handing AI assignments. For Microsoft 365 admins, Copilot Autopilot is the one to watch: tenant-resident, metered and still in private preview. OpenAI's dots are the most configurable today, and Meta's Muse shows where consumer agents are heading. The technology works in the vendors' demos. The open questions are governance, cost and accountability.