Guide Personal Agents · Chapter 6

    Cost: which levers move your agent's bill.

    Autopilot hangs on the Copilot licence, dots and Grok Bot on a plan per person, OpenClaw and Hermes on model access, a machine and operating time.

    As of: October 2026 · Editor: Marcel Haas

    • Copilot licence
    • ChatGPT plan
    • SuperGrok
    • Tokens
    • Ollama
    • Operations

    Cost levers

    Which meter, which lever.

    For every meter that moves the bill: what drives it and which lever you hold. Mechanics without figures; the amounts are in the proposal.

    1. 01

      Copilot licence per person

      What moves the meter: The number of people with a licence; Autopilot joins for tenants with a licence as a private preview, and what the preview costs beyond that is open.

      Your lever: Assign role by role after the pilot; people who rarely work with mail, documents and meetings often manage with Copilot Chat.

    2. 02

      ChatGPT plan per person

      What moves the meter: The plan tier, Business Premium or Enterprise, the number of seats and the plan's usage limits, which a dot consumes on its cloud computer.

      Your lever: One dot per person with one workflow and a cost limit; the admin enables it and reads usage in the workspace.

    3. 03

      SuperGrok plan

      What moves the meter: The plan per person on which Grok Bot runs as a beta; for enterprises the waitlist and after it Grok Business or Enterprise.

      Your lever: Bounded test accounts until the waitlist opens; the plan only for the roles where the pilot proved research with sources.

    4. 04

      Muse at Meta

      What moves the meter: The plans at the vendor and the variant for small businesses; US only so far, without a dated status with a source on our pages.

      Your lever: Wait until Switzerland is included, then review the plans with the sheet per agent.

    5. 05

      Model access in tokens

      What moves the meter: Calls per day from rhythm and schedules, the size of the context per task, the model per task and tools stuck in a loop.

      Your lever: Working hours for the rhythm, a strong model only where tools are, a spending cap per API key and project at the vendor.

    6. 06

      Local model in hardware

      What moves the meter: The model size, the graphics memory and the number of parallel requests; hardware replaces tokens for the tasks the local model takes on.

      Your lever: A small model for private notes and drafts, the hardware measured on the test set before it is bought.

    7. 07

      Machine, server or container

      What moves the meter: A virtual machine at the hosting provider or a computer in the office that runs continuously, plus the host for the sandbox with Docker or Podman.

      Your lever: One machine per gateway, that is per trust boundary; the size according to the model that computes locally.

    8. 08

      Operating time

      What moves the meter: Updates, the security audit after every change, the review of new skills and MCP servers, backups of the workspace and the person behind SOUL.md.

      Your lever: A named person with you or our support and operations team; count the hours in the pilot.

    9. 09

      Second account and identity

      What moves the meter: The agent's own identity in Entra, depending on the setup a mailbox or a licence for it, and a separate profile for the private agent.

      Your lever: One identity per company agent with the person as its owner, private matters apart; what the identity costs is in the proposal before the start.

    As of October 2026, derived from the dated facts of this chapter. Muse and Hermes Agent carry no dated status with a source; their meters name the mechanism. The amounts are in the proposal and in the dated cost articles, never on this page.

    01

    Every agent family has a different meter.

    Three people, three bills. The managing director's agent comes with the Copilot licence her tenant pays per person, and the approval for the preview. The developer's dot hangs on the ChatGPT plan of his workspace, likewise per person and month. The consultant's Grok Bot runs on his SuperGrok plan. And the self-hosted agent of the head of IT costs no licence, but model access that ticks with every call, a machine that runs continuously, and the time of the person who maintains it. A licence or a plan is a fixed amount per person, tokens move with use, a machine costs once and then in hours. Whoever wants to understand an agent's bill therefore first asks which of these meters is running, and only then how high it stands.

    02

    At Microsoft: the Copilot licence carries Autopilot.

    Copilot Chat with internet results comes at no extra cost with eligible Microsoft 365 subscriptions; answers from mail, files and meetings need a Copilot licence, and for smaller organisations Microsoft offers the Copilot Business variant. Autopilot arrives in the new Copilot for tenants with that licence, as a private preview through the Frontier programme; whether the preview costs anything beyond the licence is on none of the vendor pages we cite, and it belongs to the things we check before the start. The lever lies in the assignment: licences apply per person, and people who rarely work with mail, documents and meetings often manage with Copilot Chat. The group is enough for the pilot; afterwards you decide role by role with the figures from the evaluation, and until the preview Cowork gets described tasks done on the same licence.

    03

    At OpenAI, SpaceXAI and Meta: a plan per person, with limits.

    A dot belongs to the person's ChatGPT plan: in Switzerland through Business Premium and as an Enterprise beta once the admin enables it, not for Pro accounts at launch. Business and Enterprise give teams a workspace with sign-in through SAML SSO, Codex is included in both, and Enterprise adds SCIM, role-based access and the Compliance API. A plan comes with usage limits, and a dot that keeps working towards goals on its cloud computer consumes them, even when nobody is in the chat; what one task on the cloud computer consumes, the vendor does not say on the pages we cite, so the pilot measures it. Grok Bot runs as a beta on the SuperGrok plans, with enterprises on a waitlist; next to it there are Grok Business with central administration and usage analytics and Grok Enterprise with single sign-on, SCIM and advanced audit controls, and whether Grok Bot joins them is decided by the waitlist. Muse is offered by Meta with a variant for small businesses, US only so far; we check the plans as soon as Switzerland is included.

    04

    Self-hosted: no licence, but model access, a machine and operations.

    OpenClaw is MIT-licensed without a paid tier, Hermes Agent likewise, and yet this route has three meters. Model access: a cloud model bills per call by input and output, through an API account of your company with a project of its own and a spending cap at the vendor; a local model through Ollama replaces those tokens with hardware that has enough graphics memory. The machine: a server, a virtual machine at your hosting provider or a small computer in the office that runs continuously, plus the host for the sandbox with Docker or Podman. Operations: updates, the security audit after every change, the review of new skills and MCP servers, backups of the workspace and a person who owns all of that, with you or in our support and operations team. Two models keep the first meter small: a strong cloud model for mail, planning and everything with tools, a local one for private notes and drafts, so the token meter runs only for the tasks that need it.

    05

    Credits, limits and what one task on the cloud computer costs.

    An agent that checks every thirty minutes whether anything is due consumes tokens even when nothing is due; that is why the rhythm rests outside the working hours you set. A mailbox with two hundred mails since Friday is a long context, and every attachment it reads makes it longer; so does a research task with twenty sources and a tool stuck in a loop. Plans come with usage limits that reset at fixed intervals, and extra capacity a vendor releases for a while stays extra and never becomes a line in the budget. The levers: the working hours of the rhythm, the size of the context per task, the model per task, a spending cap per API key and per project at the vendor, a cost limit per dot, and check-backs that stop a loop before it gets expensive. What a cloud computer consumes per task we read off in the pilot on real tasks, per agent and per week.

    06

    The hidden costs: a second account, a sandbox host, the person behind SOUL.md.

    No pricing page lists the items that make the difference. The second account: an agent with an identity of its own needs that identity in Entra, depending on the setup a mailbox or a licence of its own, and a private agent next to it a separate profile. The sandbox host: tools run in a sandbox, and that runs on a machine that is maintained. The person behind SOUL.md: someone writes the instructions, maintains skills and memory, reviews logs and answers approvals, and that time belongs to the bill even though it appears on no invoice. Add the directive, the sheet per agent with the data protection officer, the training of the pilot group and the review when a vendor changes its terms. We name these items in the proposal before the first agent runs, so the bill after the pilot holds no surprise.

    07

    What the pilot shows before the agent goes to more people.

    One task, one pilot group, a few weeks, and at the end a report with three columns: what the agent does reliably, where it gets on people's nerves, and what it cost per person and week, in licences, plans, tokens, machine time and operating hours, read off real usage. From that follows the decision role by role, who gets a licence or a plan and who manages with Copilot Chat, and the caps per agent before it is opened to more people. The proposal names the frame for build and operation; the amounts are there, not on this page. What the pilot records per agent:

    • Licence or plan: which plan the person has, what it includes and which limits it sets, with the status at the vendor.
    • Usage: tokens or usage limits per week, split by rhythm, schedules and tasks a person started.
    • Machine and operations: hours of the machine, hours of the person who maintains it, and incidents that cost time.
    • Check-backs: how often the agent stopped, how often the yes came and how often the no, with the reasons.
    • Outcome: which tasks stay, which are narrowed and whether the agent goes to more people.

    What you decide.

    Licence, plan or your own hardware?

    Options

    1. 1A licence or a plan per person at the vendor
    2. 2Your own hardware with model access and operations
    3. 3Both: an agent at the vendor for mail, a local model for private notes

    Our recommendation: The meter follows the agent, and the agent follows the place where the person works; the agent finder in chapter 4 settles the place. Whoever works in a tenant with a Copilot licence pays that licence; whoever lives in the ChatGPT workspace pays the plan; whoever needs messengers, MCP servers of their own or a local model pays model access, a machine and operations. What that means per person and week the pilot proves with real usage, and only then do we compare the routes for your task.

    What we advise against: An agent chosen by the lowest list price before the pilot has shown what it gets done with you and what it consumes.

    Who gets a licence or a plan after the pilot?

    Options

    1. 1Every employee
    2. 2Role by role, with the figures from the evaluation
    3. 3Only the pilot group

    Our recommendation: Role by role, with the figures from the evaluation: which tasks the agent got done reliably for which role and what it cost there per week. People who rarely work with mail, documents and meetings often manage with Copilot Chat, which comes with eligible Microsoft 365 subscriptions; whoever spends the day in the mailbox gets the licence or the plan first. From that comes the plan for the next group.

    What we advise against: Licences or plans for everyone before the measurement. Whatever stays unused costs every month, and the pilot would have shown it.

    Where do you set the cap?

    Options

    1. 1At the vendor: a spending cap per API key, project or workspace
    2. 2At the agent: a cost limit per dot, a purchase limit per the directive, working hours for the rhythm
    3. 3Both

    Our recommendation: Both. At the vendor a spending cap for every API key and every project and the usage analytics of the workspace; at the agent a cost limit per dot, a purchase limit per the directive that it does not exceed, working hours outside which the rhythm rests, and a check-back that stops a loop. API keys sit apart from the configuration and belong to the company, never to the person.

    What we advise against: An agent with the company card and no limit, or an API key paid privately to speed things up.

    Cloud model or local model, when it comes to the bill?

    Options

    1. 1One cloud model for everything, paid per call
    2. 2Two models: cloud for mail and tools, local for private notes
    3. 3A local model only, on your own hardware

    Our recommendation: Two models, for two reasons: a local model replaces tokens with hardware that is bought once and then counted in hours, and it keeps private notes on the machine. For mail, planning and everything with tools the strongest current model stays, as the OpenClaw documentation recommends, with a spending cap at the vendor. Whether the hardware pays off against the tokens depends on the number of tasks, and the pilot measures that on your test set.

    What we advise against: A small local model for mail with tools to save tokens; security decides here before the bill.

    Status and sources.

    As of: October 2026

    • 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
    • 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
    • As of 12 October 2026

      ChatGPT Enterprise brings SCIM, role-based access and the Compliance API.

      Business and Enterprise give teams a workspace with sign-in through SAML SSO; Codex is included in both plans. Enterprise adds user provisioning through SCIM, role-based access, the Compliance API for audit and eDiscovery, and the Connector Registry.

      Source: OpenAI: Business plans
    • 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
    • Since 30 December 2025

      Grok Business and Grok Enterprise.

      Grok Business brings central administration and usage analytics, Grok Enterprise adds single sign-on, SCIM provisioning and advanced audit controls. According to the vendor, customer data is not used for training. Enterprise Vault adds a dedicated data plane with customer-managed keys.

      Source: SpaceXAI: Grok Business and Enterprise
    • 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 Autopilot cost more than the Copilot licence?

    The Microsoft pages we cite say nothing about it: Autopilot arrives in the new Copilot for tenants with a Copilot licence, as a private preview through the Frontier programme. Whether the preview has a line of its own and what applies after the preview we check with the vendor before the start and record in the proposal. The licence itself applies per person, and until then Cowork gets described tasks done on the same licence.

    Does a dot pay per task or per person?

    Per person, through the ChatGPT plan of the workspace, in Switzerland Business Premium or the Enterprise beta. The plan comes with usage limits, and a dot that keeps working towards goals on its cloud computer consumes them, even when nobody is in the chat. What one task on the cloud computer consumes, the vendor pages we cite do not say; the pilot reads it off, and a cost limit per dot holds the frame.

    What does an agent cost that does nothing at night?

    With a self-hosted agent on a cloud model: every call of the rhythm, even when nothing is due, because the model still reads instructions, memory and mailbox. That is why the rhythm rests outside the working hours you set, and fixed schedules run only when a task is due. With a licence or a plan the night costs nothing extra, but it consumes usage limits if the agent keeps working. The machine itself runs continuously, and its hours belong to the bill.

    Is a self-hosted agent cheaper?

    That depends on the number of tasks, the model choice and the person who maintains it, which is why the pilot answers it and no pricing page does. Without a licence, three other meters run: tokens or hardware for the model, a machine that runs continuously, and the operating time for updates, audit, skills, MCP servers and backups. For one person with many tasks and a local model that can be inexpensive, for a pilot group of twenty people on a cloud model not necessarily. We work out both routes in the proposal with your figures.

    Who pays for model access with OpenClaw?

    Your company, through an API account of its own with a project per agent or pilot group, a spending cap at the vendor and keys that sit apart from the configuration and are issued by IT. A person's private API key does not belong in a company agent, because usage, data route and log then sit in the wrong place. Which data the cloud model may see is on the sheet per agent.

    How do we avoid a surprise on the bill?

    With caps in two places and a measurement: a spending cap per API key and project at the vendor and the usage analytics of the workspace; at the agent a cost limit per dot, a purchase limit per the directive, working hours for the rhythm and check-backs that stop a loop. The pilot reads off the usage per agent and week before the agent is opened to more people, and the proposal names the frame for build and operation.

    Can we end the plan again after the pilot?

    Licences and plans run per person, and what applies to termination and term is in the vendor's terms, which we read before the pilot. We set the pilot up so that it ends with the group that used it: licences are then assigned or withdrawn role by role, an API project is closed, a machine is switched off. What the person taught the agent stays in your storage, even when the plan ends.

    Next chapter · Chapter 7

    How a personal agent stays governed inside the company.

    How does a personal agent stay governed inside the company?