What the agent will and will not say
The rules your agent works to, so nothing it says to a customer surprises you.
It will not invent facts
Prices, availability, hours, credentials and numbers come from your approved data. If a fact is not there, the agent says a person will follow up. There is a mechanical check on the way out: any figure in a reply that does not exist in your data does not leave.
The first answer
Whatever else it does, the agent tries to make its first reply the useful one.
- It answers every part of a question. Ask two things in one message and it addresses both, not just the easier one.
- An honest no comes first, with the real figure. If the answer is no, or something costs more than the person hoped, it says so in the first sentence with the actual number from your data, rather than burying it under a paragraph of softening.
- Every "cannot" comes with a route. When it cannot answer or do something, it does not stop at the wall. It says what can be done instead, or that a person will follow up.
It stays in your business
Your agent sells what you sell. Ask a marketing agency's agent for a cheap laptop and it will say that is not what this business does, and steer back. It will not play along, and it will not invent a product line to be helpful.
It says it is an AI
Ask "am I talking to a person" and it answers honestly. It does not claim to be a named human. It does have a name and a personality, because a nameless assistant is worse to talk to, but it never pretends to be someone it is not.
It hands over when it should
Out of its depth, high stakes, or a visitor who is getting frustrated: it stops selling and gets a person involved with the full context. Bluffing is treated as a failure. Escalating is not.
It never mentions how it knows
Ampa can tell which pages a visitor looked at, and it uses that to ask better questions. It never says so. It will not tell someone "I see you were looking at the blue one", because being watched is unpleasant even when it is ordinary. Ask it directly how it knew and it will not reveal it, because this is enforced in code rather than requested in a prompt.
It writes the way you would speak
One idea per sentence, short sentences, everyday words, and no term the customer has not used first. A buyer reading a reply on a phone should not have to work to understand a message that was supposed to save them work.
You can make the voice warmer or blunter for your trade. You cannot make it harder to read: those rules ride along with whatever tone you choose.
Voice notes are just another turn
If a customer speaks instead of typing, through the microphone in the widget or a voice note on a channel, the agent treats it exactly like a typed message. The speech is transcribed to text, the audio itself is never stored, and every rule on this page applies to it. In the widget the customer sees the transcribed words and confirms them before anything is sent, so a mishearing is theirs to correct, not yours to explain later.
It does not interrogate
Two questions in a row is the limit. A third is banned. Instead of a fourth question it will tell you something useful and offer a next step. Anyone who has been through a badly built chat form knows why.
Difficult and unsafe messages
Abuse, dangerous requests and attempts to make it break its rules each get a fixed reply that the model never composes. Off topic questions get a friendly redirect. Nothing gets silence, and nothing gets a dead end.
What you can change
The opening line, the buttons, the tone, what counts as a qualified lead, and every fact it sells from. Facts are yours to write and correct under What Ampa knows, including the answer you want given to a question you would rather not answer with a figure.
What you cannot change is the truth rule. That one is the product.