A buyer asks a question in the evening. Nobody answers before the next day, and the sale is lost. That is the problem chatbots promise to solve. Three families compete on the Moroccan market, and they are not remotely equivalent. Here is how to tell them apart before you pay.
Key point
- Family 1: the scripted chatbot
- Family 2: an assistant plugged into a language model, with no guardrails
- Family 3: the assistant constrained by the catalogue
- The six questions to ask a supplier
Family 1: the scripted chatbot
It is the most common and the cheapest. Questions and answers are written in advance, sometimes as buttons: “Track my order”, “See sizes”, “Talk to an adviser”. As long as the buyer stays inside the script, it works.
The trouble is that nobody writes like a script. “What material is it?”, “do you deliver to Kenitra?”, “chi haja rkhisa?”: the moment the question leaves the planned frame, the bot answers beside the point, or pushes a form. The buyer realises they are talking to a dumb machine, and leaves.
Family 2: an assistant plugged into a language model, with no guardrails
For two years now it has been enough to plug a language model into a shop to get an assistant that speaks well. It understands everything, rephrases, reassures. It is impressive in a demo.
It is also the most dangerous. A language model produces plausible text, not true text. Unconstrained, it announces a price that does not exist, promises 24-hour delivery nobody planned, or offers a reference that is not in the catalogue. In Morocco, where sales are settled on delivery, that mistake is paid for twice: the parcel goes out, the buyer refuses it, and the merchant pays for the round trip.
Family 3: the assistant constrained by the catalogue
The third family uses a language model for the phrasing, but checks every answer before sending it. In practice, the application reads back the sentence produced and verifies that every number it contains exists in the context provided: catalogue prices, product-page figures, the computed total, or what the buyer wrote themselves. A number from anywhere else gets the answer rejected.
The result is less spectacular than an assistant that talks about anything, and far more useful: it advises like a salesperson who knows the stock, and it says “I do not know” when the answer is not in the catalogue.
The six questions to ask a supplier
- 1How do you stop the assistant inventing a price or a delivery time? Ask for the mechanism, not the intention.
- 2Does the assistant read my catalogue live, and what happens when I remove a product?
- 3Does it understand Darija written in Latin script, and typos?
- 4What does it do when it does not know? Where can I see the list of those questions?
- 5What exactly does it collect for an order, and in what form does it hand it to me?
- 6How many conversations are included, and what happens if I go over?
What a chatbot will not fix
An assistant does not make up for an incomplete catalogue. If your product pages say nothing about material, dimensions or delivery areas, no bot will invent them — and a good assistant will refuse to. The first week often serves exactly that purpose: seeing, in the unanswered questions, what your catalogue does not yet say.
Nor does an assistant replace the merchant. It minds the shop while the merchant sleeps, it qualifies, it hands over. The call-back, the delivery and the payment stay human.
Where Cosensible fits in
Conseilor
is the sales assistant for online shops: it answers your buyers from your real catalogue, in English, French and Darija alike, fills the basket and hands you the order. It cannot invent a price. Fourteen-day trial, no credit card.
Discover Conseilor
