AI Assistants and Privacy: Where Does Your Data Go?

• Where does the data sent to an AI assistant actually go?
• Who has access to the content of my exchanges with my agent?
• What does one instance per client actually change?
• Does GDPR apply to a craft business owner processing their own clients' data through their agent?
• What never goes through the AI?
• What's true today, and what could change?
• Can I request deletion of my data?
• What happens to my history if I stop my subscription without deleting my account?

Where does the data sent to an AI assistant actually go?¶

Every message you send to your agent passes through a small, named set of recipients — never a vague circle of "partners." The content of your exchanges goes to the provider of the language model generating the responses; billing data goes to our payment processor; the application itself runs on identified servers. None of these recipients exists for any reason beyond running the service — no data resale, no advertising recipient.

This article walks through that circuit precisely, with real names — not generic categories like "our technology partners." It's written for anyone wondering where their data goes when handing a task to an AI agent, whether or not they use our service.

Who has access to the content of my exchanges with my agent?¶

Two parties, and only two, see the actual content of your messages. We do, strictly to the extent necessary for technical support if you contact us. And the provider of the language model generating your agent's responses, via OpenRouter — the platform that routes the request to the chosen model.

What most people wrongly assume: that the publisher of the underlying software, Nous Research, has access to these exchanges. It doesn't. We run their open-source software on our own infrastructure; the content of your conversations is never sent to them. The only piece of data that reaches them is your agent's display name — built from your first and last name — at the moment it connects to Telegram, needed for the automated creation of the bot dedicated to your account. Nothing more.

What does one instance per client actually change?¶

In practice, every client has their own container, isolated from others, with its own credentials, its own mailbox, and its own storage. It isn't a shared database with rows filtered by client ID — a difference that matters in the event of a breach: a vulnerability on one instance doesn't mechanically grant access to others.

This isolation has an accepted trade-off: no configuration transfers automatically from one instance to another, including on a plan change or a new deployment — a point covered in our article on migrating a self-hosted setup.

Does GDPR apply to a craft business owner processing their own clients' data through their agent?¶

Yes, and it's a point often overlooked. If you forward your agent an email containing the details or invoice of one of your own clients, you remain the data controller for that data under GDPR — being a Hermes Agent France customer changes nothing about your own obligations toward your clients. In that chain, we become an indirect subprocessor for your client data, in the same way your accounting software or your email host already are.

In practice, this changes nothing about your day-to-day use — there's no extra formality to complete specifically for this. But if one of your clients exercises their right of access or erasure with you, that may include data that passed through your agent — worth keeping in mind, not something to fear.

What never goes through the AI?¶

Two categories of data are excluded by design, not by mere recommendation. Passwords and login credentials: the agent is set up to never ask for them, and to refuse to treat them as information worth remembering if handed over anyway, systematically pointing instead to the secure login in your client account. Bank details: managed exclusively by Stripe, our payment processor — we never store them, and they never pass through the content of a conversation with the agent.

What's true today, and what could change?¶

One point deserves to be stated plainly rather than buried in vague wording: the model used by default today is offered free of charge by its publisher, which logs exchanges and may reuse them to improve its own products — a less protective policy than a typical paid model. This isn't a hidden exception: it's explicitly stated in our privacy policy, section 5, and the model actually active on your account is visible in your client account.

We plan to offer a paid model that doesn't train its systems on your data, though no date is set as of this writing. If you handle sensitive data for your own clients, factor that in today rather than waiting for that change.

Can I request deletion of my data?¶

Yes, at any time, no justification needed. Two ways to do it: deleting your account directly from the Profile section of your client account, with immediate effect — Stripe subscription cancelled, instance and mailbox deleted, account row erased, all in a single action — or a request sent to contact@hermesagentfrance.com for any special case.

What happens to my history if I stop my subscription without deleting my account?¶

It isn't erased automatically — worth knowing before cancelling rather than discovering after. Cancelling, or an unpaid invoice not resolved in time, suspends your agent: the container stops, your data stays as it was but is no longer accessible or active. Permanent deletion only happens if you explicitly request it, through account deletion or a written request — there is, as of today, no automatic timer after which your data would be purged on its own.

If you want your data to genuinely disappear when leaving the service, deleting your account remains something you need to do yourself, not an automatic consequence of stopping payment.