Dmytro Kot
Most chatbots don't fail because of the technology. They fail because nobody decided what the bot is for, it answers from thin or outdated information, it sounds like a form, and it traps people who need a person. The fixes are simple, and most of them take an afternoon.
These eight practices apply to any chatbot. For each one, we show how to set it up in Chatfuel, where an AI Agent answers your customers on WhatsApp, Instagram, Facebook Messenger, TikTok and your website:
- Give your chatbot one clear job per channel
- Feed it a clean knowledge base
- Write instructions for voice, tone and language
- Keep the conversation human
- Say it's an AI and be honest about its limits
- Hand off to a human at the right moment
- Follow up inside the messaging window
- Test before launch, then review every week
1. Give your chatbot one clear job per channel
Before you write a single answer, decide what the bot is for. A bot that tries to do everything ends up doing nothing well. Pick one main job and the number that tells you it works:
- Answer questions about prices, open hours and policies. Track how many chats end without your team.
- Qualify leads: collect what your sales team needs before they call. Track qualified leads per week.
- Book appointments: offer free slots and confirm the visit. Track bookings made in chat.
Then look at where your customers actually write. A salon may get booking requests on Instagram and WhatsApp, while a store's website chat mostly gets delivery questions. The job can differ from channel to channel.
In Chatfuel, the Default automation holds the general rules for the AI Agent, and each channel has its own automations under it. Every setting is either inherited from Default or overridden for that channel, so you keep one voice everywhere and change only what a channel needs.
A focused bot pays off. Nissan Israel used a Messenger chatbot built with Chatfuel for one job, qualifying car buyers. In the first month, it brought $380,000 in sales and 3.5x more qualified leads from Facebook.
2. Feed it a clean knowledge base
An AI chatbot is only as good as the information you give it. If the price list is from last year, the bot will quote last year's prices with full confidence. Good knowledge is short, current and written the way customers ask.
In Chatfuel, the AI Knowledge Base has these sections:
- General information: what the business does, address, open hours and policies
- Catalog: your products or services with prices. Keep it under about 50 items
- FAQ: 10–25 of the questions customers ask most, with answers of 1–3 lines
- Specialists, if the AI Agent books appointments with your team
Write each FAQ question the way a customer would type it ("Can I cancel my appointment?"), not as an internal heading ("Cancellation policy"). The AI Agent can't guess, so when a price or a policy changes, update the knowledge base the same day.
See how the AI Agent answers FAQs on every channel.
3. Write instructions for voice, tone and language
The knowledge base is what your bot knows. Instructions are how it behaves: its name, tone, reply length, emoji, what it should never promise and when to stop. Without them, an AI chatbot sounds generic, too formal or too chatty.
A simple structure works well:
- Role: "You are Mia, the assistant of Nuvia Skin, a beauty salon."
- Task: "Answer questions about services and prices, and help clients book a visit."
- Output format: "Reply in 1–3 short sentences. Use one emoji at most. Ask one question at a time."
- Example: one or two sample replies in your brand voice.
Add the language rule here too: one language for everyone, or the language the customer writes in if you serve several markets.
In Chatfuel, all of this goes into AI instructions in the Default automation. A channel automation can override it, for example with a more casual tone on Instagram than on your website.
See also: How to create a multilingual chatbot
4. Keep the conversation human
People write to a business the way they write to a friend: short messages, typos, several messages in a row. A good chatbot meets them there.
- One question per message. A list of questions reads like a form. If you qualify leads, start with three questions and add more only if your team uses the answers.
- Short replies. Two or three sentences fit a phone screen. Long paragraphs get skipped.
- A few options, not the whole catalog. In a classic study, people were more likely to buy when offered 6 jams than 24. In chat, suggest two or three options that fit what the customer asked.
- A natural pace. A wall of text a split second after the question feels robotic.
In Chatfuel, the AI Agent understands free text, so customers don't have to pick from button menus. Message delays ("Makes chatting feel more human") adds a typing animation: the AI Agent waits 8–12 seconds after a message, so when someone sends three messages in a row, it answers them in one reply. Lead Qualification picks up the details you need from the conversation itself instead of a form. See how AI lead qualification works.
See also: 5 chatbot UI examples of user interface design
5. Say it's an AI and be honest about its limits
People don't mind talking to a bot. They mind being fooled by one, or getting a confident wrong answer.
Disclosure is also becoming a legal requirement:
- EU: from 2 August 2026, the AI Act (Article 50) requires AI systems that talk to people to tell them they're interacting with AI, unless that's obvious.
- California: the B.O.T. Act makes it unlawful to use a bot that misleads people about its artificial identity to sell something, unless you disclose that it's a bot.
- Messenger: Meta's policy asks automated chats to disclose that people are talking to an automated service where the law requires it.
This isn't legal advice, so check the rules for your market. The practical part is easy:
- Introduce the bot as an AI assistant in its first reply.
- When it doesn't know, it says so and offers a person: "I'm not sure about that. Let me pass you to our team."
- Don't let it promise what only your team can decide, such as refunds, discounts or exceptions.
In Chatfuel, you write these rules into AI instructions, for example: "Introduce yourself as Nuvia Skin's AI assistant in your first message. If the answer isn't in the knowledge base, don't guess: say a teammate will help."
6. Hand off to a human at the right moment
No bot can handle everything, and customers know it. In a 2026 Gartner survey, 87% of customers said it's essential to have an option to reach a human agent when a company uses AI in customer service. WhatsApp's Business Policy requires it too: you may use automation, but you must also have "prompt, clear, and direct escalation paths."
The handoff itself matters as much. In Zendesk's CX Trends 2026 research, 74% of consumers said they're frustrated when they have to repeat information. The person who takes over should see the whole conversation.
Decide in advance when the bot steps aside:
- the customer asks for a person
- refunds, complaints and other sensitive topics
- the bot can't find the answer
- a lead is ready to buy
In Chatfuel, these are Switch to human agents rules. You can add up to 20, each with its own condition, instructions for the message the AI Agent sends before stepping aside, and the teammate who gets the chat (or your shared inbox).
The teammate sees the whole conversation in Chats, so the customer doesn't repeat anything. When a teammate takes over, the AI Agent stops replying in that chat. With Chat timeout in Settings, the chat goes back to the AI Agent after a period without replies from your team.
See how live chat and AI handoff work.
See also: Customer support automation: choosing the right tools
7. Follow up inside the messaging window
Many chats stall: the customer asks for a price and goes quiet. A short, friendly reminder brings many of them back. But messaging apps limit when a business can write:
- WhatsApp: each customer message opens a 24-hour customer service window. After it closes, you can only send pre-approved template messages.
- Messenger and Instagram: the standard window is also 24 hours. A teammate can still reply manually for 7 days with the Human Agent tag.
In Chatfuel, Send follow-ups ("When contact goes silent") has the AI Agent send a reminder after a set pause, and only inside the allowed window. To reach people after the window closes, send an approved WhatsApp template in a broadcast. One reminder is usually enough: three in a row feels like spam.
8. Test before launch, then review every week
Most chatbot problems show up in the first real conversations. Catch them before your customers do. Before launch, play the customer:
- ask your top FAQ questions and compare the answers with the knowledge base
- try to book, change and cancel an appointment
- write "I want to talk to a real person"
- make typos, use slang and send three messages in a row
In Chatfuel, Test automation in the Preview panel runs a chat with that automation's settings and your real knowledge base, without messaging customers or adding anyone to your leads.
After launch, review real chats once a week. An AI chatbot gets better when you fix its knowledge base and instructions based on what customers actually ask. AI Co-pilot does the reading for you. Ask it, for example:
- "What did clients ask most this week?"
- "Find AI mistakes in conversations"
- "Show conversion funnel"
Meet AI Co-pilot, the assistant that sets up and checks your AI Agent by chat.
Rule-based chatbots vs. AI agents
Much of the older chatbot advice, like button menus and "Back to start" options, comes from rule-based bots. They follow a script: the user taps a button and gets the next step. AI agents read free text and answer from your knowledge base.
| Rule-based chatbot | AI agent | |
|---|---|---|
| Understands | Buttons and keywords | Free text, typos, several messages at once |
| Setup | Build every branch of the flow | Fill in the knowledge base and write instructions |
| Best for | One fixed path, like a quiz or a menu | Questions, lead qualification, booking and support |
| Weak spot | Breaks on anything outside the script | Gives vague answers when knowledge is missing or outdated |
Chatfuel started with flow-based bots. Today the AI Agent is at the center of the panel, and flows are the legacy option.
See also: How Much Does a Chatbot Cost? Build vs. Buy
Summary
- Scope: give the bot one job per channel and one number to track. In Chatfuel, the Default automation sets the rules and each channel adapts them.
- Knowledge: keep the AI Knowledge Base short and current, with a catalog under about 50 items and 10–25 FAQ answers.
- Voice: write AI instructions for name, tone, length, emoji and language.
- Pace: one question per message, short replies and Message delays.
- Honesty: say it's an AI and admit what it doesn't know.
- Handoff: Switch to human agents rules for requests for a person, sensitive topics and hot leads.
- Follow-ups: inside the 24-hour window, templates after it.
- Review: test with Test automation, then check real chats every week with AI Co-pilot.
Chatfuel plans start at $20 a month with $20 in AI credits. Start a 7-day free trial and set up your AI Agent with these practices from day one.
FAQs
What are chatbot best practices?
They're the habits that make a chatbot useful instead of annoying: one clear job, a current knowledge base, a consistent voice, short replies at a human pace, honesty about being an AI, a quick path to a person, follow-ups that respect messaging rules and a regular review of real chats.
Should a chatbot tell people it's a bot?
Yes. People trust a bot that's upfront, and in some places it's the law: the EU AI Act requires it from 2 August 2026 unless it's obvious, and California bans bots that hide their artificial identity to sell. In Chatfuel, add a line to the AI instructions so the AI Agent introduces itself as an AI assistant.
When should a chatbot hand off to a human?
When the customer asks for a person, when the topic is sensitive (refunds, complaints, health or money), when the bot can't find the answer and when a lead is ready to buy. Pass the whole conversation along so the customer doesn't repeat anything. In Chatfuel, these are Switch to human agents rules.
Does an AI chatbot get better over time?
Yes, if you keep teaching it. Chatfuel's AI Agent answers from your knowledge base and instructions, so it improves when you review real conversations and update them. AI Co-pilot helps by finding the most common questions and the chats where the AI Agent got it wrong.
How do I measure chatbot success?
Track the number tied to the bot's job: chats resolved without your team for support, qualified leads for sales, bookings for appointments. In Chatfuel, leads move through the Kanban board stages (Captured, Sorting, Qualified, Selling, Closed), and you can ask AI Co-pilot for a business report or a conversion funnel.
