When a book or article says “chatbot”, it may mean quite different technologies. Knowing the main families helps you read technical chapters with confidence and choose the right kind of system for a job.
Rule-based bots
A rule-based bot follows a script. The designer maps out possible paths and writes the replies. Often the user picks from buttons or types a keyword that triggers a path.
Strengths: predictable, easy to audit, cheap to run and well suited to short, structured tasks such as booking a slot or checking a status.
Weaknesses: brittle when users phrase things unexpectedly, and the number of paths grows quickly as the scope expands.
Intent-based bots
These add a language understanding layer. The system is trained on example phrases and learns to assign a user’s message to an “intent”, such as “reset password”, and to pick out details called entities, such as a date or product name. Designers still write the responses and the flow.
Strengths: handles many phrasings, controlled output and a clear record of what the bot can do.
Weaknesses: needs training examples and ongoing tuning, and can still fail on unfamiliar requests.
Language-model bots
Systems built on large language models generate replies instead of selecting prewritten ones. They cope with open-ended questions and varied wording, and can summarise and rewrite.
Strengths: flexibility, natural tone and fast setup for broad topics.
Weaknesses: they can state incorrect things with confidence, may drift off topic and are harder to test exhaustively. Many teams therefore ground the model in approved documents and add guardrails.
Hybrid designs
A common pattern combines the approaches. Rules or intents handle sensitive transactions, while a language model handles general questions or phrases friendly replies. This keeps control where it matters and flexibility where it helps.
A worked example
Imagine an online shop that wants a bot for customer questions. Order status and returns are structured and sensitive, so they suit a scripted or intent-based flow that looks up the order and never improvises. Questions about care instructions or sizing are open-ended, so a language model grounded in the shop’s own help pages can answer them. If the model cannot find an answer in those pages, the bot should say so and offer a person, instead of guessing. This division of labour is typical of mature deployments.
Cost and maintenance
Rule-based systems cost most in design time. Intent-based systems need continuing review of misunderstood messages. Language-model systems add usage costs and the work of monitoring output. When reading vendor claims or book chapters, ask what the total effort looks like a year after launch, not only at the demo.
Which should you choose?
- Narrow, regulated or high-stakes tasks usually favour rules or intents.
- Wide knowledge bases and exploratory questions suit language models with grounding.
- If unsure, start small with the simplest option that meets the need.
Questions for any technical book
Does it explain how the approach fails? Does it discuss maintenance cost? Does it cover privacy, such as what data leaves your organisation? Answers to these separate solid guides from enthusiastic ones. The design principles in conversation design basics apply across all three families. Next, see testing and measuring chatbots.