A chatbot and an AI sales assistant are not the same thing

On a website they look identical: a round button in the corner, a greeting, a box to type in. The difference only appears when a buyer asks something that has an actual answer — is residence 1204 still available, and what does it cost.

Three things sold under the same name

• A scripted chatbot. Buttons and decision trees somebody drew. It can route a person to a form. Ask it anything the tree does not contain and it apologises.
• A chatbot over your documents. A language model with your brochure and website loaded. It talks well, and it answers from a document — so it will happily quote a price from the PDF you uploaded in June, and tell a buyer about a unit you sold in July.
• An assistant with tools. Same conversation, except that when the unit question comes up it queries your inventory, and when the buyer wants Saturday at ten it looks at your team’s calendar and books it. The answer comes from the system of record, not from a copy of it.

For a development selling real units at real prices, the third is the only one that can be let near a buyer unsupervised.

The questions to ask before you buy either

• Where does an availability answer come from? If the answer is “the documents you upload”, the product will be wrong within a month of your first closing. You want a live source and a way to replace the price list in one step.
• What does it do when it does not know? The only acceptable answer is that it says so and offers a person. An assistant that invents an amenity, a deposit structure or a delivery date will do it in writing, to a buyer, in a transcript you will later read.
• Can you control what it says about price? Exact list prices, rounded ranges, or nothing until a person is involved — and the same rule on every channel, not just the website.
• Does it work anywhere other than the website? Most inquiries arrive on WhatsApp, the phone line and by text. A website-only assistant answers the smallest part of the problem.
• Can it take a phone call? Genuinely: answer, speak, understand an accent, check inventory mid-sentence, book the visit, hang up.
• Does the lead get written down? With name, number, email, broker or buyer, what they wanted, and what was said — in your CRM, not in a dashboard nobody opens.
• When does it hand over? You should be able to name the moments: ready to reserve, budget over a number, a buyer with their own broker, anyone who asks for a person.
• Which languages, and how well? Answering a Portuguese message in English is worse than not answering. Ask to hear a call in the language your buyers actually use.
• Is there a record? Every call and chat, readable and playable, with the hand-offs flagged for review.
• Who is accountable for what it says? Read the contract. If the answer is “you are”, then the controls in the first four questions are not optional.

A test that takes five minutes

Whatever you are shown a demo of, ask it these four things in a row:

• “Do you have a two-bedroom facing the water under 1.2 million?”
• “Is 1204 still available?”
• “Can I see it Saturday at ten?”
• “Are pets allowed on the terrace?” — something the brochure does not cover.

The first two show you whether it reads inventory or a document. The third shows whether it can actually do anything or only talk. The fourth shows what it does when it does not know — which, over a year of conversations, is the one that decides whether you can leave it alone with your buyers.

Ask Sofia is the third kind: she queries your live inventory for every question about a unit, follows the price rule you set, books visits with your team, answers on WhatsApp, phone, SMS and web chat in seven languages, and says she does not know rather than guessing. Try her on your own project, or see why answering in seconds changes the outcome.