Guide Personal Agents · Chapter 9

    Who does what so that your agents stay in operation.

    A pilot group starts with one task, a one-page directive says what is allowed, and a list shows which agents run in the company and who owns each.

    As of: October 2026 · Editor: Marcel Haas

    • Pilot group
    • Directive
    • Agent list
    • Training
    • Measurement
    • Handover
    01

    One task, one pilot group, a few weeks.

    At Muster AG six people from sales and management start with a single task: the inbox, sorted by urgency, with replies as drafts. Research, expenses and appointments come in the second round. One task can be measured, its access can be kept narrow, and it shows within a few days whether the agent holds. The agent finder from chapter 4 chose the agent, licence, enablement or machine are ready, every person has an identity for her agent, and one person from IT runs the pilot: she keeps the list, collects the requests and reads along in the logs. The pilot has an end date, a few weeks after the start, and at the end stands a decision and not an extension out of habit.

    02

    The directive is written with the pilot, not after it.

    In the first week the pilot group, IT and the data protection officer sit together for an hour and write the page from chapter 7: which agents are allowed, what none of them reads, what waits for a yes, where the memory lives, who keeps the list. The pilot group supplies the sentences, because it knows what comes up in daily work; IT the accounts and limits; the data protection officer the check of the duties for third parties' personal data in the mailbox, which goes through the same legal check as any other processing activity. Every week the page comes briefly back on the table: a request answered with yes three times becomes a permitted action, a manipulated mail the agent reported becomes a sentence in the reading limits. When the pilot ends, the directive is finished because it grew with it, and it stands on the intranet, linked in every notice a stopped agent shows.

    03

    The agent list records who runs which agent.

    On the first day the list has six rows: person, agent, vendor, where its data lives, in the tenant, at the vendor or on your own hardware, its identity, the accounts it signs into, its approvals, the review date and the responsible person. Agents in the tenant IT pulls from the Agent Registry, a dot, a Grok Bot or an OpenClaw the person enters when registering, and whoever reports a private agent gets an entry and no reprimand. The list answers the questions that come later: which agents read mail with third parties' personal data? Whose agent is still running although the person has been away for three months? Which vendor changed its terms, and whom does that affect? Once a quarter the operations owner goes through it, agents without an owner pass to a new person or are blocked, and a departure closes the identity together with the entry.

    04

    Training means briefing, checking and saying where it annoys.

    The training takes an hour per group and runs on real mailboxes, with test mails we have placed there beforehand. Briefing: how to give the agent a goal, with limits and an example, and why six lines of SOUL.md achieve more than six pages; in Autopilot name, role and goal in the interface, for a dot goal and instructions in ChatGPT. Checking: where the daily log is, how to read a draft against its source before it goes out, and what a yes means. Correcting: one sentence in MEMORY.md or a reply in the chat, and the agent does it differently next time. And saying where it annoys: too many requests, a wrong tone, a task it misunderstands every week; for that there is a channel where the pilot group writes and the operations owner answers. In each area one person starts ahead of the others and answers the questions in her own team; the quick guide with the five moves sits on the learning site, and a drop-in session in the first two weeks catches the rest.

    05

    The measurement from the pilot: reliable, wrong, cost per week.

    Every Friday the operations owner reads off three columns. Reliable: which tasks the agent got done without correction, counted by the drafts that went out unchanged and the overviews nobody had to revise. Wrong: corrections, refused approvals with their reason, requests nobody understood, and the manipulated test mails it reported or missed. Cost: per person and week in licences, plans, tokens, machine hours and operating hours, read off real usage, as chapter 6 records it. Plus the time the person estimates she gained, in one sentence and not in a table. Evaluation is per group and without names, by the rule your data protection officer sets; whoever does not use the agent appears in no list but in a conversation with the champion. At the end of the pilot stands a two-page report, and it carries the decision: which tasks stay, which are narrowed, which agent goes to more people.

    06

    The handover: notes, templates and rules stay in your storage.

    We expect that over time one agent will carry most of a person's work, and which one that will be is open. That is why everything the person has taught the agent lives in her own storage: the notes it reads as memory, the templates for quotes and overviews, the rules from SOUL.md, the skills for the Friday routines. With OpenClaw and Hermes Agent those are the files of the workspace, which we mirror to OneDrive or SharePoint; with Autopilot, a dot or Grok Bot the person additionally writes agreements, templates and rules into her storage, because the memory there sits at the vendor and not as a file with you. When the agent changes, the folder moves along, and the new one reads it on day one. When the person leaves the company, IT disables the identity, the folder passes to a named successor or is deleted according to the retention rule, and the row in the list closes. That is how the foundation stays, identity, approvals, data rules, logs and storage, when the agent becomes a different one.

    07

    Support after the pilot and the step to more people.

    After the pilot every move has an address. The person maintains SOUL.md, skills and memory of her agent and answers its requests. The champion in the area answers the questions in the team and collects where the agent annoys. IT keeps the agent list, issues and withdraws identities, keeps updates, audit, backups and monitoring for self-hosted agents, or hands that to our support and operations team. The data protection officer reads the list and the sheet per agent once a quarter. And management decides with the figures from the pilot who gets an agent next: role by role, with the licence or plan that fits the role, with a training that is repeated, and with the directive that was updated beforehand. Where personal agents belong in the company's AI strategy, which roles come first and what the target picture looks like in two years, KA Partners prepares as our strategy partner; CNEXT puts it into practice, from the agent finder to operations. The roles after the pilot:

    • Person: brief, check, correct; maintain SOUL.md, skills and memory; answer requests.
    • Champion: answer questions in the own area, collect feedback, keep the quick guide current.
    • IT: agent list, identities and accounts; with own hardware updates, audit, backups, monitoring and on-call duty.
    • Data protection officer: the sheet per agent, the reading limits and the list once a quarter.
    • Management: the decision role by role with the figures from the pilot, and the directive as its word.

    What you decide.

    Which task does the pilot start with?

    Options

    1. 1The inbox with replies as drafts
    2. 2Research with sources
    3. 3The Friday routines
    4. 4All tasks at once

    Our recommendation: With a task that comes up often, is narrow, can be counted and has an approval before anything goes outside: the inbox with drafts or research with sources. Both show within a few days whether the agent holds, and both need only the access chapter 7 provides for anyway. The Friday routines come in the second round, once skills and memory have grown.

    What we advise against: A pilot with all tasks and all access on day one that after three weeks cannot tell anyone what worked.

    Who belongs in the pilot group?

    Options

    1. 1Management
    2. 2Volunteers from one or two areas with a champion
    3. 3Everyone who has a licence

    Our recommendation: Six to eight volunteers from one or two areas who spend the day in the mailbox, plus the person who will later keep the agent list, and one person from management who backs the pilot and gets an agent herself. IT and the data protection officer read along without piloting themselves. Volunteers say where the agent annoys; conscripts say nothing.

    What we advise against: A pilot group of people who were ordered to take part, or one that consists of IT alone.

    Who owns the agents after the pilot?

    Options

    1. 1Each person her own, nobody else
    2. 2A named person in IT keeps the list, every person maintains her agent
    3. 3Our support and operations team

    Our recommendation: A named person in IT keeps the agent list, issues and withdraws identities and keeps operations for own hardware, or hands that to our support and operations team; every person maintains her own agent, with SOUL.md, skills and memory. One champion per area catches the questions. Without the named person there are agents after a year whose owner has left the company.

    What we advise against: Agents without an owner that keep running when the person changes, and a list that nobody keeps.

    When do more people get an agent?

    Options

    1. 1Right after the pilot, for everyone
    2. 2Role by role with the figures from the pilot
    3. 3Only once the vendor has ended the preview

    Our recommendation: Role by role with the figures from the pilot: which tasks the agent got done reliably for which role and what it cost there per week. Every new group gets the training, the updated directive and a champion, and the list grows with it. Where an agent is still in preview, that is on the sheet per agent, and the pilot shows whether the preview is enough for the task.

    What we advise against: Everyone at once, without directive, list and training, because the licences are already there.

    Status and sources.

    As of: October 2026

    • Since September 2026

      Copilot Cowork is generally available.

      Cowork takes a described task from start to finish: it gathers context from emails, meetings, chats and files, works across several tools and delivers a finished result. Who reviews the result before it is passed on is set by your directive.

      Source: Microsoft Learn: Release notes for Microsoft Copilot
    • 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
    • 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
    • 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 April 2026

      Entra Agent ID, one identity for every agent.

      Microsoft Entra Agent ID is generally available: an identity and authorisation framework for AI agents based on OAuth 2.0, MCP and A2A, for agents from Microsoft Foundry, Copilot Studio and third-party platforms. Access, conditions and logs therefore apply per agent as they do for people.

      Source: Microsoft Learn: What is Microsoft Entra Agent ID

    Frequently asked questions.

    How long does the pilot take?

    A few weeks with an end date fixed on day one: long enough for the Friday routines to have run twice and for the agent to have seen a month end, short enough that the decision is not skipped out of habit. Every week the operations owner reads off the three columns, and at the end stands a two-page report with the decision which tasks stay and who gets an agent next.

    What if the pilot group stops using the agent after two weeks?

    Then that is a finding and no failure. We ask in the channel where it annoys: too many requests, a wrong tone, a task that was too broad. Usually a narrower task, a different briefing or one approval fewer helps; if not, the pilot ends with the report, the licences or plans end with the group, and what the people learned stays in their storage. A pilot that ends cleanly has done its job.

    Who writes SOUL.md for twenty people?

    Each person her own six lines, in the training, from a template that already contains the company's hard rules; those sit in the configuration as approvals anyway. The champion helps with the first sentences, and whatever a person tells the agent twice she adds herself. Nobody writes twenty files alone, and a SOUL.md someone else wrote the agent reads, but the person does not.

    What happens to the agent when someone leaves the company?

    IT disables the identity, and with it every access ends at once; for a dot SCIM removes the account in the workspace, for a Grok Bot or Muse the person or IT closes the account at the vendor. Open requests stay unanswered, so nothing happens. The workspace passes to a named successor or is deleted according to the retention rule, and the row in the agent list closes with the departure.

    Do we already need the data protection officer for the pilot?

    Yes, in the first week, because the agent reads mail with third parties' personal data from day one. The data protection officer co-writes the directive, checks the duties under the data protection act for the mailbox and sets the rule for how usage is evaluated in the pilot: per group, without names. The sheet per agent from chapter 5 is her tool, and with it the check takes an hour, not a month.

    What does KA Partners do, and what does CNEXT do?

    KA Partners prepares the strategy as our strategy partner: where personal agents belong in the company's AI strategy, which roles come first, what the target picture looks like in two years and what management decides for it. CNEXT puts it into practice: the agent finder, the identities, the directive with the pilot group, the training, the measurement, operations with support and operations. You have one contact person, and strategy and pilot talk to each other because we know the target picture.

    How do we measure whether the agent saves time?

    By what can be counted, and by one sentence from the person. Countable are drafts that went out unchanged, overviews nobody had to revise, and requests answered with yes; plus the cost per person and week from chapter 6. The person's sentence says what she did with the time, and that is often more important than the number. There is no stopwatch, and evaluation is per group without names; management takes the decision role by role with both.

    To the guide

    Personal agents explained, from Autopilot to OpenClaw.

    The guide shows what a personal agent gets done for you, where your mail and its memory end up and what your company allows, with date and source.