Types of Chatbots: 8 Kinds and How to Pick One

There are eight main types of chatbots: rule-based (menu) bots, keyword bots, NLU intent-based bots, LLM (generative AI) chatbots, RAG knowledge-base bots, agentic chatbots that take actions, voice bots, and hybrid bots that hand off to a human. The first three follow paths you design in advance. The next three use a large language model to understand and write replies. The last two describe the channel and the escalation path, so they combine with any of the others.

A real product is usually a mix: an LLM bot that answers from your help docs (RAG), can process a refund (agentic), works on the phone (voice) and passes hard cases to a person (hybrid).

The short version:

• Need predictable answers to a handful of questions? A rule-based or intent-based bot.
• Need answers from your own documents? A RAG bot.
• Need the bot to actually do things, like book or refund? An agentic bot, with a human handoff.
• Don't want to build anything? A chatbot-as-a-service platform you configure and embed.

Below: how each type works, what it's good at, where it breaks, and how to pick one.

The 8 types of chatbots, by how they work

1. Rule-based (menu) chatbots

How it works: you draw a decision tree. The bot shows buttons ("Track my order", "Talk to sales"), and each tap leads to the next fixed message. No language understanding is involved.

Best for: narrow, high-volume flows where every path is known: opening hours, order status lookups, lead capture forms, appointment slots.

Limits: the moment someone types a free-text question the tree didn't anticipate, the bot has nothing to say. Big trees also get painful to maintain.

Example product type: the button menus in website chat widgets and in messaging apps like WhatsApp and Messenger.

2. Keyword chatbots

How it works: the bot scans the message for trigger words and returns the reply mapped to them. This is the oldest design there is. ELIZA, built by Joseph Weizenbaum at MIT in the mid-1960s, ranked keywords by priority, matched the sentence against patterns and reassembled the fragments into a reply. When it found no keyword, it fell back to something like "Please go on."

Best for: simple FAQ routing and auto-replies where a word like "refund" or "password" reliably signals the topic.

Limits: keywords miss synonyms and misfire on context. "I don't want a refund, I want an exchange" contains the word "refund."

Example product type: auto-responders in social inboxes and older help-desk tools.

3. NLU intent-based chatbots

How it works: instead of matching words, the bot classifies the meaning of a message into an intent you defined. In Google's Dialogflow intents model, you give each intent example training phrases; the built-in machine learning expands on them, matches each user message to the best intent, and pulls out parameters such as a date or an order number as structured data.

Best for: structured tasks with a known set of requests, where you want tight control over every answer: banking menus, booking flows, IT requests.

Limits: you have to anticipate every intent and write training phrases for each. Anything outside the list lands in a fallback.

Example product type: Dialogflow, Amazon Lex and the open-source Rasa framework. Two signs of where this category is heading: Amazon now offers an "Assisted NLU" option that uses LLMs to improve intent classification while staying within the intents you configured, and Rasa's own README says Rasa Open Source is in maintenance mode, pointing to its newer approach as the future.

4. LLM (generative AI) chatbots

How it works: a large language model reads the whole conversation and writes each reply from scratch. There's no decision tree and no intent list, only instructions (a system prompt) that shape tone and scope.

Best for: open-ended questions, drafting, explanations and messy phrasing.

Limits: the model answers from what it learned in training, so it doesn't know your prices, policies or stock. It can also state wrong things confidently, which is the main risk in any customer-facing deployment.

Example product type: general-purpose assistants like ChatGPT, Claude and Gemini.

5. RAG knowledge-base chatbots

How it works: before answering, the bot searches your own content (help center articles, PDFs, product docs) and hands the most relevant passages to the LLM. The model then answers from those passages instead of from memory. We cover the mechanics in what is RAG.

Best for: support and internal Q&A where the answers already exist in writing and you want the bot to stay inside them.

Limits: the bot is only as good as the documents. Stale or contradictory articles produce stale or contradictory answers, and retrieval sometimes pulls the wrong passage.

Example product type: "train a chatbot on your website" services. Chatbase, for example, describes training an AI agent on your documents, websites or databases and embedding it with copy-paste code.

6. Agentic chatbots that take actions

How it works: the LLM is given tools: functions that look up an order, change a booking or create a ticket. In Anthropic's tool use documentation, Claude decides when to call a tool based on the request and the tool's description, then returns a structured call that your application runs. The result goes back to the model, which replies or takes the next step.

Best for: requests that end in a change, not just an answer: "move my delivery to Friday," "reset my password," "cancel the duplicate charge."

Limits: every tool is something the bot can now get wrong for real. You need permissions, confirmations for anything irreversible, and logs. If you're unsure where the line between an AI that writes and an AI that acts sits, see agentic AI vs generative AI.

Example product type: customer service agents such as Intercom's Fin, which Intercom says can update accounts, process refunds and troubleshoot through connections to your systems.

7. Voice bots

How it works: speech recognition turns the caller's words into text, one of the engines above decides what to say, and text-to-speech reads the reply aloud. Amazon describes Lex as combining automatic speech recognition (ASR) with natural language understanding for both voice and text bots.

Best for: phone lines and call center chatbots: routing callers, confirming an appointment, answering a balance or delivery question without a hold queue.

Limits: accents, background noise and interruptions all cost accuracy, and callers are less patient than chatters. Outbound calls carry legal weight too: in the US, the FCC ruled in February 2024 that AI-generated voices count as "artificial or prerecorded voice" under the TCPA, so those calls need the called party's prior express consent unless an exemption applies.

Example product type: contact-center voice platforms and phone-enabled bot builders.

8. Hybrid chatbots with human handoff

How it works: the bot handles what it can and passes the conversation, with its history, to a person when it gets stuck, when the customer asks, or when a rule says so (a large refund, an angry customer, a legal question).

Best for: almost any customer-facing deployment. Handoff is what makes the other types safe to launch.

Limits: handoff only helps if someone is on the other end. Escalating into an empty queue at 2 a.m. is worse than "we'll reply in the morning."

Example product type: most modern support bots. Microsoft's Copilot Studio is a good illustration of how the types blend: its standard harness matches requests to designed topics using natural language understanding, its customer conversation harness can escalate to a human, and its GitHub Copilot harness reasons through multi-step work and chooses tools.

Types of chatbots compared

Read the bottom two rows as add-ons. "Voice" and "hybrid" change the channel and the safety net, not the brain.

Chatbot types by deployment

The same engine can sit in several of these places at once.

• Website chatbots. A chat widget on your site, usually the first bot a business adds. Most are now RAG bots with a lead-capture or handoff step.
• Messaging app bots. Bots inside WhatsApp, Messenger, Slack or Teams. Button menus work well here because people tap on phones more readily than they type. Amazon lists all four among the platforms Lex bots can be deployed to.
• Internal and IT bots. Employee-facing bots for HR questions, password resets and ticket creation; Amazon's Lex docs list an IT helpdesk bot as a common use case. These need single sign-on and permission checks more than a friendly tone.
• Phone and contact center bots. Voice bots on your support line, often paired with human agents.

Chatbot as a service vs building your own

Chatbot as a service means a hosted platform: you sign up, point it at your content, set the tone and handoff rules, and paste an embed snippet onto your site. The vendor runs the model, the retrieval and the hosting. Chatbase is one example of this shape; it lists website, email, Slack, Zendesk and phone among its deployment channels, plus pre-built actions including human escalation. We compare options in this category in Chatbase alternatives, and voice-first builders in Voiceflow alternatives.

Building your own means wiring the pieces yourself: an NLU service like Lex or Dialogflow, or an LLM API with tool use, plus your own retrieval, hosting, logging and handoff.

On billing models, vendors differ a lot. Amazon says Lex charges only for the text or speech requests you make, while Intercom describes Fin's pricing as outcome-based. Check each vendor's own pricing page before you compare, because the unit being billed often matters more than the headline rate.

How to choose the right type of chatbot

1. Do people need answers, or need something done? Answers point to a RAG bot. Changes to an account point to an agentic bot.
2. Do the answers already exist in writing? If yes, a RAG bot can launch on them. If not, write the docs first.
3. How much variation is in what people ask? A small, stable set of requests suits a rule-based or intent-based bot, which is cheaper to run and easier to audit. Wide, messy questions need an LLM.
4. What's the cost of a wrong answer? For refunds, medical, legal or account security, keep scripted flows or require confirmation steps, and always offer a human.
5. Where are your users? Website visitors, WhatsApp customers, employees in Teams and phone callers each point to a different deployment.
6. Who will maintain it? If nobody on your team can own an integration, a chatbot-as-a-service platform is the realistic choice.

A practical pattern for most businesses: start with a RAG bot on your website, add human handoff on day one, then add actions one at a time once you've read real transcripts and know which requests repeat.

FAQ

What are the main types of chatbots?

By how they work: rule-based, keyword, NLU intent-based, LLM (generative), RAG knowledge-base, agentic, voice and hybrid. By where they're deployed: website, messaging apps, internal or IT, and phone. Most real products combine several.

What is the difference between a chatbot and an AI agent?

A chatbot talks; an AI agent acts. Once a chatbot is given tools to look things up, change records or trigger workflows, it's an agentic chatbot. In practice the two terms now overlap, and many vendors call their support bots "AI agents."

What is chatbot as a service?

A hosted chatbot platform you configure instead of build. You connect your content, customize behavior and embed the bot on your site or in messaging channels, while the vendor runs the model and infrastructure.

Are rule-based chatbots still worth using?

Yes, for flows where every path is known and mistakes are expensive, like payments and identity checks. An LLM can handle the open-ended questions around them.

Key takeaways

• Six types describe the engine; voice and hybrid describe the channel and the safety net.
• Scripted bots are predictable and limited. LLM-based bots are flexible and need guardrails.
• Pick by the job: answers from docs means RAG, changes to accounts means agentic, and anything customer-facing needs human handoff.
• Chatbot as a service gets you live fastest; building your own buys control at the cost of maintenance.

A chatbot answers questions for your customers. If what you actually want is to run AI tools yourself, like a research agent or a document workflow someone else already got working, that's a different job. Taku mirrors a working AI app or workflow into your own desktop workspace, so you can run it, adjust it and keep it without setting it up from scratch. Taku is in Beta, and the Mac app is available now.