1. The rules are missing or live in one person’s head
The agent answers “by feel”, because nobody gave it the price list, the terms and the list of things it must not promise. The fix takes a day: we write the rules the way your strongest salespeople answer, and give them to the agent.
2. The prices and terms are out of date
The prices changed in February, and nobody told the agent. The fix is to connect the agent to the source of truth: the price in your ERP or CRM, not in a file from three months ago. And support, where changes are made as they happen.
3. The agent was trained on invented examples
The scripts were written “as it should be”, but customers write “as it is”: with typos, with no item numbers, in three messages. The fix is to train on a month of real correspondence.
4. There is no handover to a person
The agent answers everything, including complaints and disputed cases, and fails there. The fix is a boundary rule: what the agent closes by itself and what it passes to a person with the history. The right share of handovers is not zero.
5. Nobody reads the dialogs
In the first two weeks the agent is learning. Without a person who reads the dialogs every day and corrects them, it locks in its errors. The fix is a responsible person with half an hour a day during the pilot.
6. There is no support

After six months all the causes above add up. The fix is support: monitoring, corrections and a monthly report on the metrics, and with us you can cancel it in any month.
How to test an agent in one hour
Take twenty real customer messages from the last week and send them to the agent. Rate each answer on three points: is it true, does it follow the rules, and did the agent pass to a person what it had to pass. Twenty answers show which of the six causes you have.
How we protect against this in a pilot
We write the rules in the first three days and give them to the agent in the form your strongest salespeople use. The agent reads prices and terms from the CRM or ERP, not from a file. It learns from a month of real correspondence. We set the handover boundary before the first customer. The responsible person reads dialogs every day. We discuss support before the end of the pilot, not after the first error.
Errors will still appear in the first week, and that is normal: the agent is learning. What matters is that they are visible and fixed within a day.
If someone else built the agent
If your chatbotchatbot. A chatbot is a program that exchanges messages with a person in a chat window: on a website, in a messenger or in an app. answers wrongly, send us twenty dialogs where it failed. They usually show which of the six causes applies, and often the fix needs no rebuild: the rules, the source of prices and the handover boundary. If the agent lives on a subscription platform and cannot write to your CRM, we say honestly where the limit of repair is and whether to move the process.
In short
Six causes: no rules, an old price list, invented examples, no handover to a person, nobody reading dialogs and no support. Each is fixable in days. A one-hour test: twenty real messages and three questions for each answer.


