Legal
AI transparency
Last updated: May 15, 2026
AIXEL builds virtual employees: software that prepares marketing work and hands it to a person before anything leaves the company. This page explains what that means in practice, because a claim about AI is only useful if you can check it.
You always know when you are talking to AI
Every conversation with a virtual employee carries a visible line saying the replies are generated by an AI. In the portal it is always on screen. In email, Telegram or Teams it appears in the first message of every new thread. There is no setting that removes it. A client can change the wording to match their own voice and language; an empty wording is not accepted, because a field for tone must not become a way to switch disclosure off.
A person approves before anything goes out
Nothing is published, sent or answered in public without a named person approving that exact text. This is not a default that can be relaxed with a setting: the approval is recorded with the identity of the person who gave it and the route it came through, and outbound delivery refuses to run without it. Under Article 50(4) of the EU AI Act, content that has gone through human review and carries editorial responsibility is treated differently from unreviewed generated output - and that review is a real one here, not a checkbox.
What the employee knows about you
A virtual employee remembers what your team tells it: your company, your products, your audience, your voice, the corrections you made to its drafts. That memory is visible and editable by you at any time. It is not shared between clients, and it is not used to train anyone's model.
Labelling generated content
Everything the employee produces carries metadata saying it was generated, by which employee, with which skill, and whether a person approved it. Where a platform offers its own AI label, we set it. Two honest limits: most social networks strip metadata when they re-encode an upload, and we do not sign content with C2PA - signing needs a key infrastructure we do not operate, and claiming it without one would be worse than not having it.
What we do not do
We do not train models on client data. We do not read a client's material or their employee's memory unless they open support access themselves, for a limited time and with a reason - and every read then appears in their own audit log, visible to them without asking us. We do not sell or share data with anyone outside the list of subprocessors on this site.
Human oversight and getting it wrong
A generated draft can be wrong, and the system is built on that assumption. It refuses to state numbers it did not compute, it blocks material that breaks a rule the client set, and it stops rather than guessing when a delivery outcome is ambiguous. If something still goes wrong, the client can pause the employee instantly, and every action it took is in an audit log they own.