How to train an AI chatbot without prompts when it answers wrong

The most common complaint about a bot is “it answers wrong”, and that raises the question of how to train an AI chatbot. It quotes the wrong price, argues with the client, doesn't handle “too expensive” and forgets to offer a booking. The client won't complain about it. They simply stop replying, and the lead is lost.

“Training AI” is a test chat where you fix this before the bot ever sees clients. You write to the bot as a client, and when a reply goes wrong, you explain in plain words how it should be. The AI rewrites the instruction itself, and the dialog is replayed right away. You don't need to know anything about prompts, and manual training is free. One fix immediately improves replies on your website, in Telegram and in MAX.

Watch the two-minute video, and below you will find step-by-step setup.

Why train your AI chatbot

• Online school: write “too expensive” and “do you offer installments?” as a client. If the bot argues or mixes up the price, fix the reply. Auto-training on “skeptics” will teach it to win back those who left a request and went quiet.
• Clinic: teach the bot not to give medical advice but to offer an appointment with the right doctor. That way the bot won't say anything you'd have to apologize for later.
• Real estate: the rule “talk about properties only based on the knowledge base data” stops the bot from making things up. If there is no suitable property, the bot offers a selection by a manager.
• Online store: a review of real dialogs with “Only WITHOUT a target action” shows the question at which clients left without ordering, and fixes the instruction itself.
• Sales team: a lead from a test is created in the CRM for real. You can check the fields, the funnel step and duplicate protection before launch, not on live clients.

How long to train a chatbot before letting it talk to clients

Training a bot well on all questions usually takes a couple of days, a week at most. It learns like a child — through practice: in tests, you won't anticipate about a tenth of the ways real people write. That's why training continues with live clients after launch: find and fix new misses in “Dialogs”, where the whole chatbot conversation history is visible. The goal is for the bot to confidently handle about 95% of questions.

How the test chat works

The test chat works exactly like the real bot: from the same AI chatbot prompt, with the knowledge base, triggers and quick reply buttons. Target actions are performed for real — a lead from a test will appear in the CRM, so you can check your chatbot CRM integration before launch. The exception is the “Bot stop/start trigger”: it doesn't fire in the test.

Edits made with the “Fix” button are first applied to a copy of the bot. Clients get them only after you click “Save” in the “You have unsaved training changes” bar.

How to fix a chatbot reply

1. 1
   Open training

   In the list of bots, click the gear icon → “Training AI”, choose “Manual training” and click “Start training”.
2. 2
   Talk to the bot

   Wait for the greeting and write the questions your clients ask: price, timelines, “too expensive”, “I'll think about it”.
3. 3
   “Fix”

   Click it under the bot's bad reply. A hint with a three-part structure opens. Click “Start training” at the bottom of the hint, and a comment field appears.
4. 4
   Comment

   Write the situation, what not to do and what to do. For example: “If the client asks about installments, don't say there are none; describe the terms and offer to submit a request.” Write in the third person, without direct speech.
5. 5
   “Apply”

   The AI rewrites the instruction, and the dialog is replayed with your lines.
6. 6
   “Save”

   Click it in the bar at the bottom so the edit works for clients. Don't like it? Click “Discard”.

When the problem is bigger than one reply

• “Fix prompt” changes the bot's behavior as a whole: tone, reply length, pushiness. Complete the task after the words “Rewrite the whole Main instruction completely so that when…”. The result is saved immediately.
• “Generate”, after the “Attention! Full instruction reset” window, restarts the bot setup wizard. All previous training is wiped out.
• “Variables” set a test name, phone number and CRM fields if the instruction uses lead data.
• “Fix quick replies” teaches the AI to suggest the right buttons, hides them after final replies and adds permanent buttons like “Book now”.
• “Fix trigger” and “Fix target action” are available for triggers and target actions of the “By AI” type. On the “Training by launch” tab, explain when to send a file or create a lead.

Automatic AI chatbot training

When you have no time to test the bot by hand, choose “Automatic training”. There are two sources:

• “Synthetic data” — the AI plays four types of clients: skeptical, neutral, warm and difficult. Tick “Outgoing dialogs” and/or “Incoming dialogs” and set the number of dialogs for each type.
• “Real dialogs” — a reviewer goes through real chats from “Dialogs”, the way a manager listens to call recordings. For example, “Only WITHOUT a target action”.

The reviewer, acting as head of sales, rates the bot's replies from 1 to 10 and edits the instruction and triggers. “Supervisor strictness” is 7 by default. To set what to look at, tick “Set a role” and describe the criteria in your own words. Before the start, the window shows the “Estimated time” and “Estimated cost”.

The default mode is “Save manually”: edits pile up in the bar, and without “Save” they are lost when you exit. For the first run, “Confirm saving” is handy: after each dialog with edits, training pauses and waits for your decision.

Read also: AI chatbot prompt: how to write instructions and add a knowledge base, Chatbot conversation history: how to review chats and take over, Chatbot CRM integration: creates leads, sends files and follows up