You are talking to an assistant
A message linked you here because it was written by software, not by a person. That is the whole disclosure. What follows is what the software is allowed to do, what it cannot do, and how to get a human instead.
Yes, it is a machine. Here is a human.
Reply with anything that asks for a person — “are you a bot?”, “can I speak to someone?”, or just “human” — and the assistant stops writing in that thread and hands it to the cardholder. It will not argue, and it will not answer the question ambiguously. Under EU AI Act Article 50 it must not; more practically, pretending is the fastest way to lose the conversation.
Why it wrote to you at all
Because you tapped an NFC card, filled in the form and ticked a box for that channel. There is no other route in. First contact with someone who has no consent record is not a setting that can be switched on — the code has no branch that permits it. What was recorded when you ticked is listed under consent and withdrawal.
What it may do without a person
- Sends itself
- A material you asked for. A fact that has a source. Offering, confirming or reminding you of a meeting slot.
- Waits for a person
- Any price or discount. Any objection. Any attempt to re-engage you after you have gone quiet.
- Never happens
- A message to anyone who did not consent. A claim with no source behind it. A reply that hides what it is.
These are decided by the class of the action, in code, before the message exists. They are not a confidence threshold and not a tone setting, so a better model does not widen them.
Where its facts come from
The assistant answers only from material the cardholder uploaded — a rate card, a specification, a one-page explainer — and from your own messages in the thread. Every factual claim it makes carries the id of the chunk it came from.
The prompt forbids inventing a commercial fact, but a prompt is advice rather than enforcement, so the check runs on the generated text as well: a claim that does not map to a stored source never reaches a channel. If the assistant does not know, it says so and asks the cardholder.
Every message is checked before it is sent
A deterministic gate — plain code reading plain rules, no model involved — runs at the moment of sending rather than at the moment of drafting. The first failure wins and the full result is stored on the message row.
- Suppression: are you on the do-not-contact list.
- Consent: is there a valid record for this exact channel.
- Channel window and jurisdiction: is this channel lawful for you, here, now.
- AI disclosure: is the line present in your language. If it is missing, the gate inserts it and re-evaluates rather than sending without it.
- Frequency and quiet hours: not too often, and not in the middle of your night.
- Grounding: does every claim map to a source.
- Action class: is this something a machine may do at all.
What it does with what you write
Your replies are used to continue that conversation and are stored with the lead record. A language model provider processes them in order to draft; it is never asked whether a message may be sent. Your messages are not used to train a model.
Anything the assistant retrieves — a web page, an uploaded document — is fenced and labelled as data before a model sees it, never merged into its instructions. Text that arrives inside retrieved content cannot give the assistant orders.
Stopping it
Reply STOP on a messaging channel, or use the unsubscribe link in an email. It takes effect on that channel immediately and is applied at send time, so a draft written before you replied still does not go out. Asking for a person stops the assistant without stopping the conversation.