Data boundaries: where your agent's mail and memory end up.
Autopilot stays within the frame of your tenant, dots, Grok Bot and Muse work at the vendor under its rules, OpenClaw and Hermes keep everything on your own hardware.
Every cell says whether an agent may read a data class, needs a documented decision, or should not see it.
Data class
Copilot AutopilotWhere it processes: Within the frame of your Microsoft 365 tenant, with the person's access; in preview
ChatGPT dotsWhere it processes: Cloud computer at OpenAI, under the terms of the workspace
Grok BotWhere it processes: Cloud computer at SpaceXAI, under the terms of the plan; as a beta
MuseWhere it processes: Virtual machine at Meta, with data and credentials; US only so far
OpenClawWhere it processes: Your hardware; what the model sees goes to its provider, locally through Ollama it does not
Hermes AgentWhere it processes: Your hardware; with a local model the text stays with you
Your own notes and appointmentsCalendar, your own notes, weekly report, internal templates
appliesWith the person's access in the tenant
appliesIn the workspace the admin has enabled
appliesWith bounded test accounts while the beta runs
documented decisionOnly once Meta offers the agent in Switzerland
appliesLocal model or cloud model under its terms
appliesLocal model or cloud model under its terms
Company mail with third parties' personal dataCustomer enquiries, job applications, supplier mail, invoices
documented decisionIn the tenant with register and directive; preview terms and memory on the sheet
documented decisionOnly in a workspace with business terms, decision of the data protection officer
documented decisionOnly with business terms; test data only in the beta
noNot available in Switzerland; data and credentials at Meta
documented decisionWith a cloud model under its terms, sandbox and directive; reading only at first
documented decisionAs OpenClaw; with a local model the content stays with you
Contracts and confidential documentsQuotes, contracts, management minutes
documented decisionPurview labels stay in the company profile; check the extract right
documented decisionOnly documents from approved connectors, after a decision
noNo confidential documents in test accounts while the beta runs
no
documented decisionReading only from the sandbox, document labels respected
documented decisionReading only, bounded toolset
CredentialsPasswords, tokens, keys for line-of-business systems
noNo passwords in the chat; the agent uses its own identity
noAccess runs through connectors, never as text in the chat
documented decisionSeparate accounts with minimal rights, never the main account
noStored credentials on the machine at Meta, as long as there is no approval
documented decisionAs secrets with restricted file permissions, never in SOUL.md or in the chat
documented decisionAs secrets with restricted file permissions, never in the chat
Professional secretsMandates, patient records, health details
documented decisionDecision of the data protection officer, a dedicated label and DLP policy; usually a separate folder it does not see
no
no
no
documented decisionOnly with a local model, without a cloud model, decision of the data protection officer
documented decisionOnly with a local model, without a cloud model, decision of the data protection officer
Legend:appliesdocumented decisionno
As of October 2026, derived from the dated facts of this chapter and the vendor pages that back them. Our frame for the start, not legal advice; the decision per data class and agent is taken with your data protection officer.
01
Every agent reads in a different place.
Monday morning, 06:45: the agent of a fiduciary reads the mail since Friday evening, among it a job application, an invoice with bank details and a client's question about his tax return. Where these mails are processed now depends on the agent. Copilot Autopilot works within the frame of the Microsoft 365 tenant with the person's access, and Enterprise Data Protection applies to Copilot: prompts and answers do not serve to train the foundation models, and the audit log in Purview records what was read. A dot and Grok Bot read on the vendor's cloud computers under its rules; Muse keeps agent, data and stored credentials on a virtual machine at Meta. OpenClaw and Hermes Agent read on your hardware, and only what the model gets to see goes to its provider, unless a local model through Ollama takes over.
02
What «no training» covers differs per vendor.
For the OpenAI API, data has not been used for training since 1 March 2023, abuse monitoring logs are kept for up to 30 days, and per project the Europe region, which OpenAI labels EEA and Switzerland, can be chosen for the listed endpoints; a self-hosted agent that calls a cloud model through this API moves within that frame. A dot lives in the ChatGPT Business or Enterprise workspace, to which the business terms apply; Enterprise adds SCIM, role-based access and the Compliance API for audit and eDiscovery. At SpaceXAI the vendor says for Grok Business and Enterprise that customer data is not used for training; Grok Bot runs today on the SuperGrok plans, whose terms you read before the start. What Meta promises for Muse we check as soon as the agent is available in Switzerland. For every agent, the sheet records before the first mail which of these commitments applies.
03
The memory is a data store with rules of its own.
What the agent remembers lasts longer than any mail: preferences, agreements, earlier decisions, and with them the names and circumstances of third parties. Microsoft Copilot stores its memory, in preview, in a hidden folder of the Exchange mailbox; Purview retention policies do not apply to it, but IT finds entries through eDiscovery. A dot keeps goals and history in the workspace at OpenAI, Grok Bot on its cloud computer, Muse on the machine at Meta. OpenClaw and Hermes Agent write MEMORY.md and USER.md as files into the workspace that we put in your own storage; the person reads, changes and deletes every entry herself. The directive therefore also says what the agent may remember: agreements with clients yes, health details and bank details of third parties no.
04
Third parties' personal data in the mailbox is a matter of the data protection act.
The job application, the invoice with bank details and the client's tax question are third parties' personal data, and whoever sent them did not know that an agent reads them. The register of processing activities therefore names the agent with purpose, data categories and recipient, that is the vendor where it reads; the information to the people concerned, the contract with the vendor and retention belong to it. Professional secrets in law firms, practices and fiduciary firms need their own review before a mailbox with mandates reaches an agent at a vendor. Your data protection officer answers these questions; we supply the technical details per agent and never say that a tool fulfils the duty. The company rule on which data may go into which tool applies to agents as it does to chat.
05
The sheet per agent, before it reads the first mail.
One page per agent, signed by the person, IT and the data protection officer: which agent at which vendor, where it computes, what it stores and for how long, which data classes it may read, which mailboxes and folders, which training commitment applies, who approves exceptions and when the details are checked again. This sheet belongs in the register of processing activities, and the data boundaries matrix above is its frame for the start. When the vendor changes its terms, the sheet changes, and the chronicle of this guide shows the change with its date.
06
At Microsoft, OpenAI, SpaceXAI, Meta and with you.
Where mail and memory end up per agent, in one sentence per vendor.
At Microsoft: Autopilot stays within the frame of your tenant, with Enterprise Data Protection, the audit log in Purview and a memory in the person's mailbox that eDiscovery reaches; the preview terms go on the sheet.
At OpenAI: a dot works in the ChatGPT Business or Enterprise workspace under its business terms, with SCIM and the Compliance API in Enterprise; if a self-hosted agent calls the API, its data rules apply, with the Europe region per project.
At SpaceXAI: Grok Bot works on a vendor cloud computer on the SuperGrok plans; the published commitments cover Grok Business and Enterprise and the API, and you read the terms of the plan before the start.
At Meta: Muse keeps agent, data and stored credentials on a virtual machine at Meta; as long as the agent is available in the US only, it stays in the matrix without an approval.
With you: OpenClaw and Hermes Agent read on your hardware, memory and credentials stay there as files, and the model sees only what you give it; with a local model through Ollama the text never leaves the machine.
What you decide.
Which data class may which agent read?
Options
1Your own notes, appointments and internal content without third parties' personal data
2Company mail with third parties' personal data
3Contracts, credentials and professional secrets
Our recommendation: Your own notes, appointments and internal content may be read by every approved agent, at the vendor too. Company mail with third parties' personal data an agent reads only after the documented decision per agent: Autopilot in the tenant with register and directive, a dot only in a workspace with business terms, Grok Bot only with business terms, OpenClaw and Hermes with a cloud model under its terms or with a local model. Contracts, credentials and professional secrets stay with the agents whose data route your data protection officer has approved one by one, and credentials never belong in the chat.
What we advise against: A mailbox with mandates or health details handed to an agent at a vendor because the pilot should start quickly.
Cloud agent or your own hardware for the memory?
Options
1At the vendor, with its retention and access rules
2On your hardware, as files in your storage
3At the vendor for appointments and agreements, locally for everything private
Our recommendation: On your hardware when the memory contains names and circumstances of third parties that you must be able to read, change and delete, or when it should outlast the agent. At the vendor when the agent works there anyway and the sheet records where the memory lives, who finds it through eDiscovery and how long it stays. For Copilot no Purview retention policy applies to the memory, and that goes on the sheet.
What we advise against: A memory at the vendor of which nobody knows where it lives and who can read it.
Who reads the agent's memory?
Options
1Only the person
2The person and IT through eDiscovery
3The person, IT and the data protection officer during a review
Our recommendation: The person in daily work, IT through eDiscovery in a proceeding, the data protection officer during a review, each with a log. With OpenClaw and Hermes the memory sits as a file in the workspace with restricted file permissions, with Copilot in the person's mailbox, with a dot and with Grok Bot at the vendor. Who can read there is in the vendor's terms, and the sheet records it.
What we advise against: A memory that a team shares. A personal agent belongs to one person, and OpenClaw assumes one trust boundary per gateway.
Who signs the sheet per agent?
Options
1The person alone
2The person and IT
3The person, IT and the data protection officer
Our recommendation: All three: the person describes which mailboxes, folders and tasks the agent gets; IT records with which identity and which access it works and what is logged; the data protection officer approves the data route, the data classes and the training commitment. The sheet belongs in the register of processing activities and is signed again when the vendor changes its terms.
What we advise against: One decision for every agent in the company. The answers differ per vendor, plan and mailbox, so per agent.
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.
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.
API data: no training, a Europe region per project.
OpenAI has not used API data for training since 1 March 2023, and abuse monitoring logs are kept for up to 30 days. The Europe region, which OpenAI labels EEA and Switzerland, stores and processes in the region for the listed endpoints and requires a retention option approved by OpenAI. The Agents API is not listed there.
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.
OpenClaw assumes one trust boundary per gateway: one person or a team that trusts each other. A shared gateway is not meant for people with opposing interests. On a regular host the gateway binds to the local address, and unknown senders receive a pairing code.
The documentation states that smaller models are more susceptible to tool misuse and hijacked instructions. For agents with tools it recommends the strongest current model, a sandbox for sensitive tools and web search switched off where it is not needed.
This is how we put it into practice: data protection under the FADP for every agent in the company, with register, directive and the sheet per agent before it reads the first mail. Compliance & data protection
Does the vendor read along when the agent sorts my mail?
At the vendor the agent processes your mail on its computers, and the terms of the plan say who has access there, how long logs are kept and whether content goes into training. Microsoft names Enterprise Data Protection for Copilot, OpenAI the exclusion from training for the business plans and the API, SpaceXAI the same for Grok Business and Enterprise. What applies to your agent's plan is on the sheet that is signed before the first mail.
May the agent read mail with a client's health details?
Your data protection officer decides that, and with professional secrets the decision usually goes against an agent at a vendor. We prepare the groundwork: which mailboxes and folders contain such details, which label they carry and which rule keeps them out of the agent's reading scope, such as a separate folder it does not see. Many pilot groups start with appointments, their own notes and routines without third parties' personal data.
What happens to the memory when I change agents?
With OpenClaw and Hermes you take the files MEMORY.md and USER.md with you from your storage, because we create the workspace there. With Copilot the memory stays in the person's mailbox, with a dot and with Grok Bot at the vendor, and whether you can export or delete it is in its terms and on the sheet. That is why we also keep agreements, templates and rules in your own storage.
Does everything stay on the machine with a local model?
For the tasks the local model takes on, yes: private notes and drafts through Ollama never leave the machine. For mail, planning and everything with tools, the OpenClaw documentation recommends the strongest current model, because smaller models are easier to steer with manipulated content, and that model usually runs in the cloud. What it sees goes to its provider; that is why we split the tasks across two models and record the line on the sheet.
Do we have to inform the senders of our mail?
Whether and how, your data protection officer settles; the duty to inform and the register belong to the data protection act, and an agent that reads the mailbox is a new processing with a new recipient. We supply the details it needs: which agent, at which vendor, with which purpose, which data categories and which retention. One line in the privacy notice and in the register is often the result.
Can IT see what the agent has read?
With Copilot, Purview writes prompts, answers and the sources used into the audit log. With OpenClaw every session has a transcript on your machine, and every approval is in it with its time. With a dot the admin sees the workspace, in Enterprise with the Compliance API; with Grok Bot and Muse it depends on the plan, and what the vendor logs we check before the start. What your IT can read is on the sheet per agent.
Next chapter · Chapter 6
Cost: which levers move your agent's bill.
What does a personal agent cost, and what moves the bill?