Guide Personal Agents · Chapter 4

    Which personal agent fits you and your company.

    Where you work, where your data should stay and whether the agent must reach you by messenger decide the agent. The finder gets there in a few questions.

    As of: October 2026 · Editor: Marcel Haas

    • Agent finder
    • Copilot licence
    • ChatGPT Business Premium
    • SuperGrok
    • OpenClaw
    • Hermes Agent

    Agent finder

    The questions that lead to the agent.

    Every answer points at one or two agents. The agent with the most hits is the starting point; its status and what to check before the start are listed below.

    1. 01

      Where do you work?

      • In Outlook and Teams, in a tenant with Copilot licencesleads toCopilot Autopilot
      • In the ChatGPT workspace, on Business Premium or Enterpriseleads toChatGPT dots
      • On a SuperGrok planleads toGrok Bot
      • On none of these plans, or across several worldsleads toOpenClawHermes Agent
    2. 02

      Where should mail and memory stay?

      • Within the frame of our Microsoft 365 tenantleads toCopilot Autopilot
      • At the vendor is fine, under its rulesleads toChatGPT dotsGrok Bot
      • On a vendor machine that keeps everything togetherleads toMuse
      • On our own hardwareleads toOpenClawHermes Agent
    3. 03

      How should it reach you?

      • In Teams and Outlook, where I work anywayleads toCopilot Autopilot
      • In ChatGPT, by voice, in Slack or in Teamsleads toChatGPT dots
      • At the vendor, in its app or in the browserleads toGrok BotMuse
      • Over WhatsApp, Signal or Telegramleads toOpenClaw
    4. 04

      How does it reach your company systems?

      • Through Microsoft Graph and the connectors in the tenantleads toCopilot Autopilot
      • Through plugins and connectors of the workspaceleads toChatGPT dots
      • It signs in itself with stored accessleads toGrok BotMuse
      • Through MCP servers of our own that we runleads toOpenClawHermes Agent
    5. 05

      Does it need a local model?

      • No, the vendor's model is enoughleads toCopilot AutopilotChatGPT dotsGrok BotMuse
      • Yes, for private notes and drafts through Ollamaleads toOpenClawHermes Agent
      • Yes, and it should derive skills from completed tasksleads toHermes Agent
    6. 06

      Who maintains it?

      • Our IT through Entra and policiesleads toCopilot Autopilot
      • The workspace adminleads toChatGPT dots
      • The person herselfleads toGrok BotMuse
      • Someone who runs a server, with us or at CNEXTleads toOpenClawHermes Agent

    Where the answers lead.

    Copilot Autopilot

    What that means: A persistent agent in the new Copilot, with its own identity, memory, computer and workspace within the frame of your tenant.

    Status: Announced as a private preview on 25 September 2026, through the Frontier programme for tenants with a Copilot licence

    What you check: Whether the preview has reached your tenant, the Copilot licence per person, the access in Entra, the directive; Cowork until then.

    Open page

    ChatGPT dots

    What that means: A dot keeps working towards goals in ChatGPT on its own cloud computer and reports by chat, voice, Slack or Teams.

    Status: Available since 29 September 2026; in Switzerland through Business Premium and as an Enterprise beta, not for Pro accounts at launch

    What you check: Business Premium or the Enterprise beta once the admin enables it, the plugins and connectors per dot, the check-backs before actions, the directive.

    Open page

    Grok Bot

    What that means: An agent on its own cloud computer at SpaceXAI that signs into your tools and completes multi-step work up to the approval.

    Status: Beta since 11 August 2026 on the SuperGrok plans, enterprises on a waitlist

    What you check: The person's SuperGrok plan or the waitlist for enterprises, bounded test accounts before real access, the directive.

    Open page

    Muse

    What that means: An agent on a dedicated virtual machine at Meta that keeps agent, data and stored credentials together, as an app and in the browser.

    Status: US only so far; no dated status with a source on our pages

    What you check: Whether Meta offers the agent in Switzerland, the small-business variant, where the machine runs and what it stores, the directive.

    Open page

    OpenClaw

    What that means: A self-hosted agent with a gateway for Teams, WhatsApp, Signal and Telegram, a sandbox, tool rules and the model of your choice; memory and credentials stay with you.

    Status: Open source under the MIT licence, developed by the OpenClaw Foundation, without a paid tier; October 2026

    What you check: A machine, server or container and the person who owns it, the model access, the approval of the Teams app by your IT, the directive.

    Open page

    Hermes Agent

    What that means: A self-hosted agent from Nous Research that derives skills from completed tasks, remembers across sessions and runs with any model, including local ones.

    Status: Open source under the MIT licence; no dated status with a source on our pages

    What you check: A machine or container and the person who owns it, a bounded toolset, stop conditions, the evaluation of what it has learned, the directive.

    Open page
    01

    Where you work decides first.

    A managing director works in Outlook and Teams, and her tenant has Copilot licences. A development team lives in the ChatGPT workspace. A consultant uses a SuperGrok plan privately and wants to reach the agent from the phone. For each of these people there is a different agent today, and the first question is where the mail, appointments and documents live that it should read. Autopilot works with the person's access in the Microsoft 365 tenant, a dot with the connectors of the workspace, Grok Bot signs into tools with stored access, OpenClaw and Hermes reach mail and calendar through Microsoft Graph and company systems through MCP servers you run yourself. An agent that cannot reach the person's tools remains a chat.

    02

    Where the data should stay draws the second line.

    The data protection officer of a fiduciary firm asks first where the mail goes before talking about an agent at all. Autopilot stays within the frame of the tenant IT already knows; a dot and Grok Bot work on the vendor's cloud computers under its rules; Muse keeps everything on a machine at Meta; OpenClaw and Hermes keep memory and credentials on your hardware, and only what the model gets to see goes to its provider, unless a local model through Ollama takes over. Whoever has third parties' personal data, contracts or professional secrets in the mailbox decides on this line first; chapter 5 shows per agent what it may read.

    03

    Messengers, company systems and a local model narrow the circle.

    If the agent should be reachable from the phone over WhatsApp, Signal or Telegram, today that leaves an OpenClaw gateway that serves those channels; Teams and Slack are also served by a dot and Autopilot, and Grok Bot and Muse you reach at the vendor, in its app or in the browser. If it should reach line-of-business systems through MCP servers of your own, the route leads to OpenClaw or Hermes, with a dot to the plugins and connectors of the workspace, with Autopilot to Microsoft Graph and the connectors in the tenant. If a local model must read private notes, the self-hosted agents with Ollama remain. And who maintains it decides in the end: your IT through Entra and policies, the workspace admin, the person herself, or someone who runs a server.

    04

    The finder names an agent and its status.

    Every answer in the finder points at one or two agents; the agent with the most hits is the starting point. Its card names the status from the overview and what you check before the start: availability at the vendor, the licence or plan, your company's directive. The finder promises no more than that, because the vendors run their agents in preview and beta and availability and plans change. What holds up with you is shown by the pilot with one task and one pilot group; after that it is settled whether the agent goes to more people.

    05

    What we recommend for the typical situations.

    Our recommendation follows the place where the person works and the data route the data protection officer approves.

    • Microsoft 365 tenant with a Copilot licence: Cowork as the first step, the sign-up for the Autopilot preview through the Frontier programme, directive and identity in the same period. Result: the agent appears in the tenant IT already governs.
    • ChatGPT workspace: one dot per person with one workflow, approvals and a cost limit, in Switzerland through Business Premium or the Enterprise beta once the admin enables it. Result: the dot works in the workspace IT evaluates.
    • SuperGrok plan: Grok Bot with bounded test accounts before real access follows; enterprises are on a waitlist. Result: a trial that shows whether the agent suits the roles.
    • Own hardware: OpenClaw when messengers, MCP servers of your own or a local model are required, with sandbox, minimal permissions, pairing and log; Hermes Agent when the agent should derive skills from completed tasks. Result: memory and credentials stay with you.
    • What we advise against: an agent that reads a mailbox with third parties' personal data before the directive exists, and an agent on a person's private account for company mail.

    What you decide.

    Which agent for a person in the Microsoft 365 tenant?

    Options

    1. 1Cowork as the first step, with the same approvals as Autopilot later
    2. 2Copilot Autopilot as soon as the preview reaches the tenant
    3. 3OpenClaw alongside for messengers and MCP servers of your own

    Our recommendation: Cowork first, because it stays in the programs IT governs and gets described tasks done across mail, meetings and files; in parallel the sign-up for the Autopilot preview, the directive and the agent's own identity in Entra. OpenClaw joins when the person has to reach the agent over WhatsApp, Signal or Telegram or line-of-business systems are connected through MCP servers of your own.

    What we advise against: A second agent at a vendor for people whose mail lives in the tenant, only because it is available earlier.

    Which agent for a team in the ChatGPT workspace?

    Options

    1. 1One dot per person with one workflow, approvals and a cost limit
    2. 2Microsoft Copilot for the people whose mail lives in Microsoft 365
    3. 3Both, separated by group

    Our recommendation: One dot per person, with a workflow that begins and ends in the workspace, approvals before actions with consequences and a cost limit; the admin enables Business Premium or the Enterprise beta. Whoever has mail and documents in Microsoft 365 gets Copilot, and both groups follow the same directive. Which platform the team itself needs is settled by the platform choice in the OpenAI guide.

    What we advise against: Dots for Pro accounts or private accounts. The workspace the admin administers is the prerequisite.

    When is your own hardware worth it?

    Options

    1. 1When memory and credentials have to stay with you
    2. 2When the agent should be reachable over WhatsApp, Signal or Telegram
    3. 3When a local model should read private notes

    Our recommendation: In all three cases, and as soon as one person owns the machine, the updates, the security audit and the backups of the workspace. OpenClaw brings gateway, sandbox, tool rules and pairing and serves the messengers; a strong cloud model takes mail and tools, a local model through Ollama private notes and drafts. Hermes Agent fits when the agent should derive skills from completed tasks and remember across sessions.

    What we advise against: Your own hardware without a person who owns updates, audit and backups. An agent nobody maintains eventually reads with outdated rules.

    When does the agent read the first real mail?

    Options

    1. 1After the directive, with a task without third parties' personal data first
    2. 2During the pilot, with test data
    3. 3Right away, the directive follows the pilot

    Our recommendation: After the directive. A mailbox contains third parties' personal data, and before an agent at a vendor reads it, purpose, recipients and retention are settled with your data protection officer. The pilot starts with test data and a task without third parties' personal data, such as the weekly report from your own notes; the directive takes shape alongside, and with it the agent reads the first real mail, reading only at first.

    What we advise against: An agent at a vendor that reads the mailbox before the directive exists. What it has read cannot be recalled from the vendor.

    Status and sources.

    As of: October 2026

    • Since 25 September 2026

      Copilot gets Autopilot.

      Microsoft introduced the new Copilot with Home, Code and Autopilot: a persistent personal agent with its own identity, memory and computer, announced as a private preview for the end of September.

      Source: Microsoft: Introducing the new Copilot
    • Since 29 September 2026

      OpenAI brings dots.

      Dots keep working towards goals in ChatGPT on their own cloud computer and are reachable by chat, voice, Slack or Teams. In Switzerland through Business Premium and as an Enterprise beta, not for Pro accounts at launch.

      Source: OpenAI Help Center: Getting started with your dot
    • Since 11 August 2026

      SpaceXAI launches Grok Bot.

      Agents with their own cloud computer that sign into your tools, complete multi-step work and report back when an approval is needed. As a beta on the SuperGrok plans, enterprises on a waitlist.

      Source: SpaceXAI: Introducing Grok Bot
    • Licensing as of 25 September 2026

      Web chat is included, chat with work data needs the licence.

      Copilot Chat with internet results comes at no extra cost with eligible Microsoft 365 subscriptions. Answers from your work data need a Copilot licence. For smaller organisations, Microsoft offers a "Copilot Business" variant for the Business plans. We clarify up front what your plans include.

      Source: Microsoft Learn: License options for Microsoft Copilot
    • As of October 2026

      Open source, backed by a foundation.

      OpenClaw is MIT-licensed and developed by the OpenClaw Foundation, an independent US non-profit. There is no paid tier, and telemetry is limited to a version check that can be switched off.

      Source: OpenClaw Docs

    Frequently asked questions.

    Does the finder really name an agent?

    Yes, a starting point: the agent most answers point at, with its status from the overview and what you check before the start, namely availability at the vendor, licence or plan and your company's directive. Whether it holds up with you is shown by the pilot with one task. The finder promises nothing the vendor does not yet deliver.

    Which licence does Autopilot need compared with Copilot?

    The same Copilot licence per person, plus the approval for the private preview through the Frontier programme. Copilot Chat with internet results comes at no extra cost with eligible Microsoft 365 subscriptions, answers from mail, files and meetings need the licence, and an agent that works with that data needs it just the same. What your plans include, we read off up front.

    Can we trial two agents in parallel?

    Yes, with separate groups and the same task, so that the evaluation stays comparable: Cowork in the tenant for one group and OpenClaw on a test machine for the other, for instance. Both work with test data until the directive stands, and both groups record what the agent does reliably and where it gets on their nerves.

    What if we have neither Copilot nor ChatGPT nor a SuperGrok plan?

    Then the pilot starts with what is available without a plan: OpenClaw on a machine at your place with model access in the cloud or a local model through Ollama, or Copilot Chat on the web, which comes with eligible Microsoft 365 subscriptions. Which plan pays off later is shown by the pilot on one task.

    Can the agent be changed later?

    Yes, if what the person has taught it stays with her. We keep notes, templates and rules in your own storage, the directive and the identity apply to every agent, and the evaluation from the pilot describes the tasks a successor has to take over. We expect that over time one agent will carry most of the work; which one is open.

    Who decides on the agent in the end?

    The person, your IT and your data protection officer, with distinct roles: the person describes her tasks and channels, IT sets the frame of approved agents, identities and access, the data protection officer approves the data route. The finder supplies the starting point with its status, the pilot the figures.

    Next chapter · Chapter 5

    Data boundaries: where your agent's mail and memory end up.

    Where do my mails and the agent's memory go, and what may it read?