TL;DR
- Rule-based chatbots answer exactly what was written but miss unexpected phrasing; LLM chatbots understand free text but can hallucinate prices or terms unless limited to a knowledge base (RAG).
- On LINE OA, reply messages sent with the user's reply token do not count toward the monthly message quota, while push messages the business sends later do.
- Hand the chat to a human when the user asks, after two failed understandings, for complaints or refunds, and for judgement calls like special discounts, and pass the chat history along.
- Separate abandoned chats from resolved ones: in a hypothetical 1,000-chat month with 150 abandoned, a 75% 'not handed over' rate is really a 60% resolved rate.
A chatbot is a program that replies to users automatically in a chat channel such as a LINE Official Account, Facebook Messenger or a chat window on a website. Some chatbots follow rules and buttons set up in advance. Others use AI or a large language model (LLM) to understand freely typed questions. Businesses use chatbots to answer repeat questions, take orders, book appointments or screen customers before passing them to a human admin.
This article explains how chatbots work, how rule-based and AI chatbots differ, how Thai businesses use chatbots on LINE OA and Facebook, when to hand a chat to a human, what to watch with customer data, and which numbers to measure.
What a chatbot is and how it works
A chatbot system has four main parts. The first is the channel where users type, such as LINE or Messenger. The second is the message receiver, usually a webhook the platform calls on the business's server each time someone sends a message. The third is the bot's brain, which decides what to reply. The fourth is the back-end the bot pulls data from, such as stock levels, an appointment calendar or order status.
The usual flow: a user sends a message, the platform forwards it to the webhook, the bot works out what the user wants, fetches any data it needs, and sends a reply through the platform's API. If the bot does not understand, or the matter needs a person, a well-built system hands the conversation to an admin along with the chat history.
"Understand" means very different things depending on the bot type. Some bots only match keywords. Some classify the intent of a message. Some use an LLM to generate a fresh answer every time. That difference shapes cost, risk and the upkeep needed after launch.
How a rule-based chatbot differs from an AI chatbot
Rule-based chatbots
A rule-based chatbot follows conditions a person wrote: if the user taps "Check prices", send the price table; if the message contains "opening hours", reply with the hours. Every answer is predictable, easy to audit, and the bot will never say something the business did not write. The downside is that it handles unexpected phrasing badly. A user who types "what time do you close" may get nothing if the rule only matches "opening hours". So these bots lean on buttons, menus and quick replies to steer users along designed paths.
AI chatbots built on an LLM
An AI chatbot built on a large language model understands free text far better, including typos and casual language, and writes natural replies. It also brings risks to manage. First, an LLM can give answers that sound right but are wrong (hallucination), such as prices, promotions or terms that do not exist. Second, cost scales with usage. Third, answers vary from one time to the next, which makes them harder to audit.
The standard ways to reduce risk: have the bot answer only from the business's own knowledge base (retrieval-augmented generation, or RAG), define topics it must not answer on its own, such as special prices, refunds or medical advice, and hand over to a human when it is not confident. Data that must always be right, like prices and stock, should come straight from back-end systems, not from the model's memory.
Hybrid bots
In practice many businesses combine the two: rule-based buttons and menus for fixed processes like booking or order tracking, and AI for general free-text questions. That gives precision where precision matters and flexibility where users ask in many different ways.
| Comparison | Rule-based chatbot | AI chatbot (LLM) |
|---|---|---|
| How it decides what to reply | Buttons, keywords and written conditions | Interprets free text and generates a new reply |
| Accuracy of information | Exactly as written, every time | Must be limited by a knowledge base and rules, or it may answer wrongly |
| Unexpected phrasing | Handles it poorly, often falls back to a default message | Handles it better, including typos and casual language |
| Upkeep after launch | Add new rules when questions are not covered | Update the knowledge base, review conversations and adjust answer limits |
How Thai businesses use chatbots on LINE OA and Facebook
LINE Official Account already includes basic tools in LINE Official Account Manager: a greeting message when someone adds you as a friend, keyword-based auto-response messages, and a rich menu of buttons at the bottom of the chat. For a small business with a handful of repeat questions, these may be enough without building a bot.
When you need a bot that pulls data from other systems or uses AI, you connect LINE OA to your own server through the Messaging API by enabling the webhook in settings. One thing worth knowing: LINE splits outgoing messages into two kinds. Reply messages use the reply token that arrives with the user's message, must be sent within a limited time, and do not count toward the monthly message quota. Push messages, which the business sends on its own initiative, do count toward the plan's quota. So a bot that answers instantly adds no messaging cost, while a bot that follows up later uses quota.
The LINE OA response settings also need to agree with each other. If both LINE's built-in auto-responses and a webhook bot are on, users can get two replies to one message. Decide which one answers. For account and rich menu setup, see LINE Official Account services.
Facebook Messenger works on similar principles, but Meta has messaging-window rules. Businesses can reply normally within 24 hours of the user's last message, and sending outside that window has specific conditions. A bot designed to follow up later has to be planned around this rule. Meta also has a handover mechanism for passing a conversation between a bot and the admin inbox.
Tasks chatbots handle well on these two channels in Thailand include answering repeat questions like price, opening hours, directions and shipping fees, taking and confirming bookings, checking order status, and collecting basic details before passing a lead to sales. Once the bot collects data, keeping tags and customer history in one system lets you send better-targeted messages later, which is the job of LINE CRM.
When to hand the chat to a human
A good chatbot knows when to stop. A chatbot handoff to a human should happen at least in these cases:
- The user asks for a person, for example types "talk to admin". Hand over at once without asking again.
- The bot fails to understand twice in a row. Looping the same question frustrates users and they give up.
- Complaints, refunds, or any conversation where the customer is getting upset.
- Matters that need judgement, such as special discounts, corporate offers, or medical and financial questions.
- Strong buying signals, like a request for a large quote. Sales should pick it up quickly.
A good handoff carries the chat history so the customer does not have to repeat themselves, and tells the customer plainly whether they are talking to a person or a bot and when an admin will reply. Outside office hours, say when someone will answer instead of leaving them waiting with no idea. Designing the handoff path between bot and team is part of customer service automation.
Caution with customer data
Chatbots collect personal data very easily, because users can type anything. Thailand has the Personal Data Protection Act B.E. 2562 (2019), known as PDPA. This article is not legal advice, but these general practices lower the risk:
- Ask only for data the task needs. A booking bot may need just a name and phone number, not a national ID number.
- Avoid having the bot ask for sensitive data such as health details or card and bank account numbers. If truly needed, send users to a system designed for that data.
- Tell users they are talking to a bot and what their data will be used for, with a link to the business's privacy policy.
- If an AI chatbot sends messages to an outside provider for processing, check that provider's data terms to see how messages are stored or used.
- Limit who on the team can read chat history and set how long data is kept.
For what your business specifically must do under the law, consult a qualified data protection legal adviser directly.
How to measure a chatbot
An unmeasured chatbot tends to be left running with nobody knowing whether it helps or annoys customers. The numbers to track:
- Containment rate: the share of conversations finished without a human. A high number is not always good; users who give up midway also count as "not handed over".
- Handoff rate and the reasons for handoff, which show what the bot should learn next.
- Fallback rate: the share of messages the bot did not understand. Use it as a to-do list for new rules or knowledge base entries.
- Post-chat satisfaction, for example one question at the end asking whether the user got what they needed.
- Business outcomes, such as bookings made, leads passed to sales, or orders that started in chat.
Hypothetical example: a bot handles 1,000 conversations in a month. 600 finish in the bot, 250 go to a human, and in 150 the user gives up midway. Counted crudely, the "not handed over" rate is 75%. Separate out the abandoned conversations and the genuinely resolved rate is 60%. The second figure is the honest one, and those 150 conversations are where to read real transcripts and find the cause.
Beyond the numbers, reading real conversations regularly helps a lot, especially in the first weeks after launch. The questions customers actually ask often differ from what the team expected at design time.
Steps to start a chatbot
- Collect real questions from past chats, group them by topic, and start with topics asked often that have clear answers.
- Choose the bot type by task. For fixed processes, start rule-based. If questions vary widely and a knowledge base is ready, consider AI.
- Write the answers and the path to a human, including the message for when the bot does not understand.
- Test internally with many phrasings, including typos and casual language.
- Launch to part of your audience first, track the numbers above, then expand.
If the bot is part of a marketing journey, such as sending follow-ups after a booking or segmenting customers by what they asked, connecting it to marketing automation lets you reuse chat data downstream.
Common mistakes
- Hiding the way to a human, so customers who want an admin get stuck in a loop.
- Letting AI answer prices or promotions from the model's memory instead of live data.
- LINE auto-responses and a webhook bot running at the same time, so users get two answers.
- Launching a bot and then nobody maintains it, so its information falls behind the website.
- Measuring only message volume, not whether customers got answers or bought.
Chatbots in the Thai market: extra things to consider
Thai is written without spaces between words, so a simple keyword-matching bot can segment words wrongly or match a keyword hidden inside a longer word. Testing with real messages matters more than in English. Thai users also mix in English, abbreviations, slang and stickers, which rule-based bots often cannot handle. Have a fallback message that returns the user to the menu or passes them to a person.
Tone matters too. A bot that uses polite particles matching the brand's personality reads more naturally than answers translated straight from English. And because many Thai customers are used to messaging a shop before buying, fast replies in the evening and on holidays are where a bot clearly helps, as long as it tells customers plainly which matters must wait for an admin.
Frequently asked questions about chatbots
Is a chatbot the same as an AI chatbot?
Not entirely: a chatbot is any automated chat responder, while an AI chatbot is one that uses a language model to understand and generate answers. Many chatbots still run on buttons and rules with no AI at all.
Can I build a LINE OA chatbot without code?
Yes, at a basic level: LINE Official Account Manager lets you set a greeting message, keyword auto-responses and a rich menu yourself. Pulling data from other systems or using AI requires connecting through the Messaging API and a webhook.
Can a chatbot fully replace admins?
It should not fully replace them. Chatbots suit repeat questions and clear processes, while complaints, negotiation and judgement calls still need a person, so there should be an easy path to one.
What should I use to measure chatbot success?
Measure the genuinely resolved rate, the handoff rate, the fallback rate, satisfaction, and business outcomes such as bookings or orders, and look at conversations users abandoned midway as a separate group.
Summary
A chatbot is an automated chat tool that helps most on the right jobs: repeat questions, clear processes, and screening before a human takes over. Choosing between a rule-based chatbot and AI depends on how much the task needs accuracy versus flexibility, and every version needs a route to a person, care with customer data, and honest measurement. If you want a chatbot on LINE OA that works together with your team's replies, see customer service automation from Relevant Audience.
