A chatbot should not close the sale. It should recognize the moment when a human needs to be called.
I say this as someone who builds AI agents for companies and sells them myself.
Most conversations in a website chat are questions about delivery, returns, invoices, opening hours. The agent answers them better than an FAQ and nobody has to sit at the keyboard. But some conversations look different.
A message at 11:10 pm: "do you have this in a 12-person version? the site only shows up to 8 and I have to order in the morning".
This is a customer with a budget and a deadline who has just written to you what they want to buy. A bot that replies "our offer includes versions up to 8 people" and closes the topic has just lost the order. In the statistics that conversation will be "resolved", because the customer didn't write a second time.
I checked how it looks elsewhere. Detecting a buying signal in a conversation appeared this year at the biggest vendors, so it is not our invention. The difference is in what happens over the next minute.
At Nexly Agents we do it like this.
First: during the conversation the agent recognizes a buying signal: a question about a quote, a deadline, a version, a quantity, availability. It doesn't try to sell on its own. It asks the customer: "I can connect you with a consultant who will handle this. Would you like that?". The customer clicks yes or no.
Second: a human from your company takes over the conversation in the same window. They see the whole history, write in Polish, and the customer reads in their own language. The agent stays silent until the end of the conversation. The lock sits on the server, so it doesn't depend on whether the model follows instructions. One person handles several conversations at once, also from a phone.
Third: when nobody is there at 11:10 pm, the agent doesn't pretend someone will reply any moment. It collects what the consultant needs in the morning: what, how many, by when, whom to call. The request lands in the team's e-mail or Slack.
Every such conversation gets an outcome: handed to a human, lead, order, unresolved. Silence on the customer's side counts as "unknown", not "resolved". Independent research shows why this matters: chatbot vendors claim the bot handles 30 to 60 percent of cases on its own, while independent surveys measure 10 to 25 percent. The 2026 median is 41 percent. The rest goes to a human anyway; the only question is whether with the whole conversation or from scratch.
The handover only works when someone at your company is actually at the panel. If a consultant doesn't accept the conversation within 60 seconds, the agent goes back to collecting contact details. The agent also can't tell a customer with a budget from a student writing a paper: it recognizes that the question is about a purchase, and a human judges the rest.
A question for you: how many times in the last month did someone in your chat write "I want to order", and a bot answered them?
Want to see it on your own site rather than on slides? Enter your website address and e-mail in the demo form. In 20 minutes you get an agent trained on your website's offer, ready to be quizzed.
Sources
- Zendesk CX Trends, Salesforce, Forrester 2025, independent deflection benchmarks