What it does
A customer sends a message at 23:00 asking a price. Within about four seconds they have the real number, in Iraqi dialect, followed by the next question that moves them toward buying. The conversation is written to your lead log with an intent and an urgency score. If the customer asks something the system should not answer alone, it stops, marks the thread for a human, and pushes a Telegram alert to your staff.
What it does not do
It does not pretend to be a person. Every conversation opens with a disclosure that the customer is talking to an AI assistant, and that line cannot be removed. Clients ask for this occasionally and the answer is always no. It is written into the service agreement.
For clinics and pharmacies, replies that carry clinical content route to a human for review before they reach the patient. That is a safety requirement, not a feature toggle, and we sell it as one.
Where the limits are
The generic assistant Meta ships for free can already answer a simple question. What it structurally cannot do is know your actual price list, your actual stock, your actual booking calendar, speak Iraqi dialect and Kurdish Sorani properly, handle a voice note, or put a named person in Erbil on the other end when it breaks. That gap is the product.