What if an AI assistant tells a buyer something wrong?
Limit what it's allowed to say. Sabbaq answers only from the agency's own listing data, and hard rules in code stop it from negotiating price, giving legal or mortgage advice, or confirming a viewing without an agent. It hands over the moment the buyer asks. During shadow mode an agent approves every reply first.
The rules are in code
The blocked paths are not a line in a prompt. The product classifies a message that asks for a discount or a last price, for legal advice, for mortgage advice, or for a viewing to be confirmed, and it refuses those. The same path hands the chat to a person when the buyer asks for one.
The assistant says it is an assistant. It does not pretend to be a named agent.
Listing answers come from your file
The pilot includes basic listing-aware answers from the brokerage's own CSV. If the sheet does not contain a fact, the assistant is not supposed to invent one. An agent still confirms the viewing and any price conversation.
Shadow mode is the first week
For the first 7 days of the pilot, an agent approves each draft before it is sent. That is the time to see a wrong phrase while it is still a draft.