AI transparency

1. Anyone who writes knows it is an assistant

In its first reply, when someone comes back after 12 hours of silence and when a team member hands the conversation back, the assistant says that it is the business’s virtual assistant and that the person can ask for someone from the team whenever they like. That text is not generated by the model: the server adds it, so it always reads the same.

This is the transparency duty set by the EU AI act, Regulation (EU) 2024/1689, in its article 50: anyone talking to an AI has to know it.

2. What the assistant does

• Answers with what the business publishes: what it offers, its opening hours, its rates and its rules.
• Offers free calendar slots and drafts bookings, reschedules and cancellations, which the person confirms before they are saved.
• Notes down the name and language the person mentions; any other change to their details it proposes before saving.
• Hands the chat over to the team when asked, when it cannot resolve something or when it receives a voice note or an image, and notes the reason and the gist of what was said. In a healthcare centre, faced with an emergency, it tells the person to call 112 before handing it over.

Its scope is that business alone: it is no all-round assistant.

3. The model and the data

Aipiter uses Google’s Gemini models through Vertex AI, run on EU infrastructure, and the system configuration rejects any region outside Europe.

The model receives the customer’s text or the button they pressed, never the team’s private notes, attachments or the inbox’s internal fields. Google undertakes not to use what its Cloud customers send to train or adjust its models without prior permission or instruction (Google Cloud service terms, training restriction), and Aipiter does not train models on conversations.

4. What the model does not decide

The model grants no permissions, does not touch stored data and changes nothing on its own: it proposes, and the server checks the business, the permissions, availability and the confirmation before making a change. A confirmation is valid for 15 minutes, and only for that conversation and that appointment.

Before it is sent, every draft goes through a check that stops it if it treats an unconfirmed step as done, asks for payment or identity details or shows internal text. Whatever a customer writes or a document contains is treated as information, never as an order that changes permissions.

Aipiter is not designed to decide on its own about health, employment, credit, insurance, education or access to essential services. Those uses need their own assessment and a qualified person to make the decision.

5. Talking to a person

Asking in the chat is enough. The case goes to the team, who see it among the conversations that need attention, and the customer is told that their request is waiting for the team. From then on a person replies, and how long that takes depends on each business.

The team can also take over a conversation whenever they want or pause the assistant in it; when it is switched back on, the assistant picks up only from new messages.

6. Errors and complaints

If an answer looks wrong to you, tell the business in the same chat or write to us at hola@aipiter.com with the business’s name, the channel and the approximate date, without passwords or details that are not needed. What else is done with your data, and on what basis, is covered by Aipiter’s privacy policy.