Blog5 min read
Chatbots that actually convert
What separates a chatbot that sells from one that annoys: real business context, a clean handover to a human, and presence where customers already write.
Why most chatbots get in the way
The average chatbot opens by itself after three seconds, asks for a name and an email before helping with anything, and answers every specific question with a link to the FAQ. It is not a poor implementation of a good idea: it is the wrong idea, executed faithfully.
Visitors do not want a conversation. They want an answer — what it costs, whether you deliver to their city, whether it fits their case — and then they want to move on. Anything standing between the question and the answer works against you.
A useful chatbot knows your business
The difference between an assistant that helps and a decorative one is not the language model: it is what information sits in front of it when it answers. Connected to your catalogue, your prices, your delivery windows and your policies, an agent answers with verifiable facts. Without that connection, it improvises.
That distinction matters because a model without context does not say “I don’t know”. It invents something plausible, and in customer service a plausible falsehood is worse than nothing: it creates a commitment somebody will have to undo later.
Knowing when to stop answering
A good agent recognises its limits and hands the conversation to a person, with everything already said loaded in, without making the customer repeat it. The handover is the most important function in the system and almost always the worst built.
It is worth defining explicitly what triggers it:
→ A complaint, a cancellation, or anything emotionally loaded
→ Off-list pricing, special terms, or negotiation
→ Two consecutive answers where the agent could not find the information
→ The customer asks for a human, in those words or close to them
Where the agent lives
Across Latin America commercial conversations happen on WhatsApp, not in a website widget. An agent that answers on the site but not on WhatsApp solves the smaller half of the problem, and usually the half with less traffic.
The sensible shape is one agent, with one body of knowledge, answering on the channels your customers already use. Keeping separate bots per channel guarantees they contradict each other within weeks.
What to measure
The figure most people report — number of conversations — says nothing. Three matter instead: what share resolved without human intervention, how many ended in a concrete action (a quote, a booking, a purchase), and how many were abandoned midway.
That last one is the most valuable. The conversation where the customer stopped replying is showing you precisely what the agent could not answer.
Let’s Connect
Got a project in mind?
Let’s make something happen together
