A chatbot is a software program that holds a conversation with a person through text or voice, on a website, in an app, or inside a messaging platform. Simple chatbots follow pre-written rules; AI chatbots use language models to interpret free-form questions and generate answers. Businesses use them to answer questions, capture leads and book appointments 24/7.
KEY TAKEAWAYS
- A chatbot is software that holds a conversation in text or voice and answers or acts on what the person says.
- Rule-based bots follow a decision tree; AI bots interpret free-form language with a language model and compose an answer.
- Most production business chatbots are hybrids: scripted flows for structured jobs, an AI layer for everything nobody scripted.
- The idea is nearly sixty years old - ELIZA dates to 1966 - but large language models are what made unscripted answers workable.
- Free plans exist, small-business plans run about $19 to $99 a month, and enterprise suites commonly reach four figures.
- A chatbot is only as good as the content behind it and the escalation path in front of it.
Free plan, 600 conversations a month, no credit card. Live in about 10 minutes.
What Is a Chatbot, Exactly?
At its core, a chatbot is a program that takes one side of a conversation. The interaction looks like any messaging thread: a person types a message, the chatbot interprets it, and it responds with text, buttons, images, or an action such as booking a slot or opening a support ticket. It can live in a chat widget in the corner of a website, inside WhatsApp or Instagram, in Slack, or behind a phone line.
Three words get used interchangeably and should not be. A bot is any program that performs automated work - a search-engine crawler is a bot and never speaks to anyone. A chatbot is the subset whose job is conversation. An AI chatbot is one whose decisions come from a machine-learning model rather than from rules a person wrote in advance. All three sit under chatbot in our glossary.
The distinction that matters is not what a chatbot looks like but how it decides what to say. That one question - scripted branch, trained intent, or generated sentence - determines what it can handle, what it costs, how it fails and how much work it is to maintain. Everything below follows from it.
From ELIZA to LLMs: A Short History of Chatbots
Chatbots are not a product of the current AI wave. They are one of the oldest ideas in computing, and the arc explains why so much chatbot advice still assumes constraints that no longer apply.
Turing proposes the imitation game
Alan Turing's paper Computing Machinery and Intelligence asks whether a machine can converse indistinguishably from a human. It made conversation the test of machine intelligence, which is why chatbots still carry outsized symbolic weight. Turing, 1950 (PDF).
ELIZA
Joseph Weizenbaum builds ELIZA at MIT. Its DOCTOR script imitates a Rogerian psychotherapist by pattern-matching and reflecting the user's own words back. Weizenbaum was disturbed by how readily people confided in it. Weizenbaum's original ACM paper (PDF).
PARRY, Jabberwacky, A.L.I.C.E.
A generation of pattern-matching bots refines the same trick. Impressive in demos, brittle in use: everything they can say has to be anticipated by a human author first.
Voice assistants and the messaging-bot wave
Siri, Alexa and Google Assistant put speech interfaces in everyone's pocket, and Facebook opens Messenger to bots in 2016. Intent-based platforms make chatbots mainstream - still classifiers picking from a fixed menu of intents.
Large language models
ChatGPT's release moves the expectation from 'picks the right prewritten answer' to 'writes a new one'. Business chatbots follow, and the frontier is now agentic: bots that call tools and complete tasks.
For the first fifty-odd years, everything a chatbot could say had to be written by a human first. Language models broke that constraint, which is why the hard question changed from "what did we remember to script?" to "how do we stop it saying something we never approved?".
How Do Chatbots Work?
Every chatbot does three things: receive a message, decide what it means, produce a response. As a pipeline, one turn looks like this:
- Ingest. The message arrives from a channel - website widget, WhatsApp Cloud API, Telegram webhook - with metadata about sender and context.
- Preprocess. Text is normalised, speech transcribed, language detected so a multilingual chatbot replies in kind.
- Understand. The bot works out intent and entities - the step that differs most between architectures, unpacked below.
- Decide. Dialogue management picks the next move: answer, clarify, run a flow step, call an API, or escalate.
- Respond and log. The reply is rendered for the channel, sent, and written to the transcript that feeds chatbot analytics.
Rule-based chatbots: a decision tree with a chat interface
A rule-based chatbot follows a tree a human designed: it asks, offers buttons, and branches on the answer. Predictable, free to run per message, and the right choice for structured jobs - collecting a name and email, walking someone through a return, qualifying a lead. The weakness is rigidity: anything outside the tree produces a fallback message. Designing those trees well is a craft; our conversation design masterclass is the deep version.
Intent-based NLU chatbots: classify, then answer
The generation of chatbots that dominated from roughly 2016 works by classification. You define intents - the things users want, such as track_order or cancel_subscription - and give each one example phrasings. The model learns to map new messages onto one of those intents, and separately extracts entities: the order number, the date, the product name. Google's Dialogflow documentation is a clear public reference for both concepts, on intents and entities. In our glossary the same ideas are intent recognition, entity extraction, slot filling and utterance.
These bots handle paraphrase well - "my package never showed up" and "where is my order" land on the same intent - and their answers are exactly what a human wrote, which auditors like. The cost is curation: every new question type means a new intent, new examples, and a check that it does not collide with an existing one. NLP chatbot features and the NLP for chatbots guide go deeper on the mechanics.
Generative and LLM chatbots: write the answer
A generative chatbot passes the conversation to a large language model and lets it compose a reply - no intent list to maintain, no ceiling on phrasing. Two mechanisms make that safe enough for business use:
- Grounding. The model is given the company's actual content - help pages, policies, product docs - and told to answer only from it. The general technique is retrieval-augmented generation: the difference between a bot that quotes your refund policy and one that invents it. On a no-code platform this is pointing the bot at your website and documents through an AI knowledge base.
- Instructions and guardrails. A system prompt sets persona, tone, off-limits topics and the rule that says "if the answer is not in the provided content, say so and offer a human". Our prompt engineering guide covers writing one that holds up.
Skip either and you get hallucination: a confident, fluent, wrong answer - common enough that we wrote guides on preventing chatbot hallucinations and what to do when an AI chatbot gives wrong answers.
Hybrid chatbots: what almost everyone actually ships
In production the strongest pattern is neither pure. Scripted flows own anything with a fixed correct path - bookings, payments, identity checks, lead forms - because you want determinism where money and data are involved. The AI layer catches the rest. A visual AI chatbot builder lets you draw both on one canvas and decide, block by block, what the model is allowed to decide. For the background, our complete guide to conversational AI is the long-form companion to this section.
Types of Chatbots
Chatbots get classified two ways, which is why "how many types are there?" gets a different answer everywhere. The first cut is by mechanism - how the bot decides what to say - and it is the taxonomy that predicts behaviour:
| Type | How it decides | Strength | Weakness | Best fit |
|---|---|---|---|---|
| Menu / button bot | User picks options; no language understanding at all | Zero ambiguity, works anywhere, trivial to audit | Cannot answer anything not on a button | Short fixed processes: hours, order status, routing |
| Rule-based / keyword bot | Matches keywords and follows a decision tree | Predictable, cheap, deterministic, no model costs | Breaks on phrasing the author did not anticipate | Lead forms, qualification, bookings, returns |
| Intent-based NLU bot | Classifies into a trained intent, extracts entities | Handles paraphrase; answers stay exactly as written | Needs training examples and ongoing curation | High-volume support with a stable question set |
| Generative / LLM bot | A language model reads the message and writes a reply | Answers questions nobody scripted, in any language | Invents things unless grounded in your content | Open-ended product, policy and documentation questions |
| Hybrid | Scripted flow where the path is fixed, AI where it is not | Deterministic on money and data, flexible on questions | Two systems to design and keep in sync | Most real business deployments |
| Voice bot | Speech recognition in front, synthesis behind, same brain | Hands-free, works over a phone line | Transcription errors compound; turn-taking is hard | Phone systems, smart speakers, accessibility |
| AI agent | Reasons about a goal, then calls tools to complete it | Finishes tasks instead of describing them | Needs tight scope, guardrails and audit trails | Look up an order, change a booking, file a ticket end to end |
The second cut is by job - what the bot is for, and the one buyers shop by: support chatbots, lead generation chatbots, booking bots covered in the appointment booking use case, ecommerce assistants, internal helpdesk bots described in our guide to internal IT helpdesk chatbots, and survey bots such as an NPS survey chatbot. The two axes are independent: a support chatbot can be rule-based, generative or hybrid, and the mechanism is what decides whether it works.
Chatbot vs Conversational AI vs Live Chat vs AI Agent
These terms describe different layers, not competing products, and confusing them costs buyers real money. Glossary entries for conversational AI and live chat hold the short versions; the clean separation is:
| Term | Who or what replies | Underlying technology | When you want it |
|---|---|---|---|
| Live chat | A human writes every reply | Messaging interface, routing, queueing | Complex, emotional or high-value conversations |
| Chatbot | Software replies automatically | Rules, NLU, or a language model - or all three | Repetitive questions and structured tasks, any hour |
| Conversational AI | Not a product - a technology category | NLU, LLMs, speech recognition, dialogue management | The umbrella term for what makes chatbots smart |
| Virtual assistant | General-purpose, usually branded and voice-first | Conversational AI plus device and app integrations | Siri, Alexa, Google Assistant and enterprise cousins |
| AI agent | Software that acts, not only answers | LLM reasoning plus tool calls, memory and guardrails | When the job is to complete a task, not to explain it |
The practical question is never which one to buy but how they hand off. The pattern that works: the chatbot answers first and resolves the routine majority; live chat takes over when the conversation needs judgement or empathy; and the handoff carries the transcript so the customer never repeats themselves. Our human handoff best practices and the human handoff definition cover the mechanics; the head-to-head sits in chatbot vs live chat and conversational AI vs chatbot. If you are weighing the cost of each, chatbot vs live chat costs puts numbers on it.
Visual flows plus AI answers from your content - on the free plan, no card.
What Are Chatbots Used For?
Every chatbot use case is a variation on one of five jobs. Industry changes the vocabulary, the integrations and the compliance constraints - not the jobs.
Customer support
The largest category by volume. A support bot trained on your help centre answers repetitive questions instantly and escalates the rest. Full playbook: customer support chatbot guide, with operational detail in building a support chatbot and covering after-hours support.
Lead generation and qualification
Instead of a static form, a bot asks qualifying questions in conversation and routes hot leads to sales while the person is still on the page. See the lead generation chatbot guide and lead qualification with chatbots.
Booking and scheduling
Clinics, salons and consultants let customers pick a slot inside the chat, synced to a real calendar, with no phone tag. Calendar booking covers availability; automated reminders cover no-shows.
Ecommerce and conversational commerce
Product finders, order tracking, shipping questions and cart recovery, usually on the channel the customer already uses. See ecommerce chatbots and the conversational commerce guide.
Internal operations
The same technology pointed inward: HR policy bots, IT helpdesk bots and onboarding assistants answering employees from internal docs - internal self-service FAQ bots.
By industry
Constraints change more than the technology does. We publish full guides for healthcare, real estate, education, hospitality, restaurants, insurance, legal, automotive, logistics and HR, plus a browsable library of chatbot use cases. One warning: if the bot touches protected health information in the US, read HIPAA-compliant AI chatbots before building anything.
Where Chatbots Run: Channels and Deployment
A chatbot is not a website feature. It is a conversation engine you attach to whatever surface your customers already use, and the channel decides more than buyers expect: available formatting, who owns the sender identity, what the platform charges and what its policies allow.
- Your own website. A website chatbot is one embed snippet with full control over look and behaviour. Platform-specific installs exist for WordPress, Shopify, Wix, Squarespace, Webflow and WooCommerce.
- Messaging apps. WhatsApp, Facebook Messenger, Instagram, Telegram and LINE, each with its own API, message-window rules and template approvals - WhatsApp's are in Meta's Cloud API docs.
- Workplace tools. Slack, Microsoft Teams and Discord, for internal and community support.
- Mobile apps. Native SDKs put the bot inside an iOS, Android, React Native or Flutter app.
Running one bot across several of these at once is omnichannel deployment, and it differs from simply having many separate bots - omnichannel vs multichannel explains why shared context is the point. Conferbot includes every channel on every plan, free one included, with no per-channel fee; the full list is on the chatbot channels overview.
Benefits of Chatbots
The case rests on five structural properties, not on any clever feature.
- Instant, always. No queue, no opening hours, no time zones. First response time is one of the few metrics customers actually feel.
- Consistent. The documented answer every time, in every supported language - no variance between a new hire and a veteran.
- Parallel. One bot handles a thousand conversations as easily as one, so seasonal spikes stop being a staffing problem; see handling holiday support volume.
- Cheap per conversation. Not free, but far below staffing the same hours. Cost per conversation is the metric to compare, and the free calculators and tools run it on your numbers.
- They capture demand a form loses. A visitor who will not fill in six fields will answer three questions in a chat - chatbots vs forms.
Limitations, Risks and Honest Tradeoffs
Every benefit above has a matching failure mode, and pretending otherwise is how chatbot projects get cancelled in month four.
- A chatbot decays. It is only as good as the content behind it; an unmaintained bot serves last year's prices with total confidence. Budget review time, not just build time - chatbot mistakes to avoid.
- Generative bots are wrong fluently. Ungrounded models invent policies, prices and features. Grounding plus an explicit "I do not know, here is a human" path is the mitigation, not a nice-to-have.
- Some people want a human immediately. Forcing an upset customer through a bot damages the relationship support exists to protect. Visible escalation is a feature, not a leak.
- Transcripts are personal data. Set retention periods, restrict who can read conversations, and keep card and health data out of chat entirely. See chatbot GDPR compliance.
- Prompt injection is a real attack. Text a model reads can contain instructions aimed at the model. It is the top entry in the OWASP Top 10 for LLM applications and it matters the moment your bot can do anything beyond talk: prompt injection and chatbot security.
- Disclosure is becoming a legal requirement. The EU AI Act requires that people are informed they are interacting with an AI system unless it is obvious - Article 50. Practical versions: chatbot disclosure laws and EU AI Act compliance.
The conclusion is unglamorous: automate the repetitive majority, keep humans reachable for the rest, and read transcripts weekly so the bot improves instead of rotting. Broader risk list: chatbot security risks and prevention.
How Much Does a Chatbot Cost?
Chatbot pricing is quoted in at least four incompatible units - per seat, per conversation, per AI resolution, per message - which is why comparison is hard. The market by band, with the trap in each:
| Band | Typical price | What you actually get | What to watch |
|---|---|---|---|
| Free tiers | $0 | Real but capped. Conferbot's free plan covers 600 conversations a month with no card; most competitors cap conversations, seats or AI replies too. | What happens at the cap: hard stop, overage billing, or forced upgrade |
| Small-business plans | ~$19-$99 / month | Where most SMBs land. Conferbot's paid plans are $19 (Starter), $39 (Pro) and $59 (Business) a month, priced on conversation volume rather than per channel. | Per-seat pricing and metered AI replies, which turn a flat plan into a variable bill |
| Enterprise suites | $500-$5,000+ / month | Support suites with AI add-ons, per-seat licences and per-resolution fees. The list price is rarely the invoice. | Per-resolution AI charges: the better your bot gets, the more you pay |
| Custom build | Developer time + model + hosting | Building directly on LLM APIs. Maximum control, and you own retrieval, evaluation, moderation, channel plumbing and uptime forever. | Maintenance, not the build, is where custom chatbots get expensive |
| Channel pass-through costs | Set by the platform, not the vendor | WhatsApp is the big one: Meta bills per message on the Cloud API, separately from your chatbot vendor. | Model WhatsApp volume first - it can exceed the software bill |
The pass-through row is the one people forget. Meta bills WhatsApp conversations independently of your chatbot vendor, under the schedule in the WhatsApp pricing documentation. Run your volume through the WhatsApp API cost calculator before signing anything, and read the October 2026 service-message billing change.
Conferbot's own numbers, for reference: a free plan with 600 conversations a month and no card, then Starter at $19, Pro at $39 and Business at $59 per month on the pricing page, with every channel included on every tier. For the wider market, see our pricing comparison, Intercom pricing breakdown, total implementation cost and the platform comparisons.
Cost is the wrong headline number anyway. Return is what decides whether the project survives: agent hours removed, leads a form would have lost, tickets never opened. The free tools and calculators model it on your volumes; how to calculate chatbot ROI explains the formula.
Metrics: How to Tell Whether a Chatbot Is Working
"Conversations handled" is a vanity number: a bot that answers a thousand people badly scores well on it. Five metrics tell the truth:
- Containment rate - the share of conversations the bot finished without a human. The best single headline number; realistic bands are in our containment benchmarks.
- Deflection rate - tickets that never got created because the bot answered first.
- CSAT measured on bot conversations specifically, not blended with human ones, so a good human team cannot mask a bad bot.
- Escalation rate and where escalations cluster. The clusters are your content backlog.
- First contact resolution - did the customer come back the next day with the same question?
Full treatment in the chatbot KPIs guide and metrics to track; the reporting side is chatbot analytics. To compare two versions of a flow properly rather than by vibes, start with chatbot A/B testing.
How to Build a Chatbot (Without Coding)
A decade ago this needed a development team. On a no-code chatbot builder - the category is defined at no-code platform - it is a setup flow most people finish in ten minutes:
- Start from a template, not a blank canvas. Pick the closest of the 250+ chatbot templates and edit it; the templates guide explains how to choose.
- Design the flow. Drag blocks onto the canvas: greeting, question, condition, form, handoff - the six parts are broken down in chatbot flow templates. Keep the happy path short and the escape hatch obvious - patterns to copy are in conversation flow templates.
- Give it your content. Point it at your website URL, help docs and PDFs so answers come from your material, not the model's general knowledge: training on your business data and on a knowledge base.
- Connect the channels. One embed snippet for the site, then whichever messaging apps your customers use - how to add a chatbot to your website.
- Test with real phrasings, then launch small. Ask it the ten questions your team actually gets, badly-worded ones included, and ship to one page first.
- Read the transcripts weekly. Every fallback is a content gap with a timestamp. This is the whole difference between a bot that improves and one that decays.
Hands-on tutorials: how to build a chatbot without coding and the complete no-code builder guide. Developers who would rather own the stack should weigh Python chatbot vs no-code and the chat API. Still choosing a platform? Our ranked best AI chatbot builders is the honest version; the free chatbot builder is where you stop reading and start clicking.
The Operations Reality: Errors, Limits and Things That Break
Almost every introduction to chatbots stops at "and then you launch". In practice the first month is consumed by operational questions, and they come from the messaging platforms rather than the bot. Meta, Telegram, Slack, Discord and LINE each enforce their own rate limits, message windows, template approvals and error codes.
- Error codes. A failed message returns a number and one line. Our error code directory documents 325 codes across 8 messaging platforms with the literal message, the real causes and an ordered fix - including WhatsApp Cloud API codes, Telegram Bot API codes and Discord API codes.
- Rate limits and quotas. Messaging tiers, per-second caps and payload limits, gathered per platform under platform limits - for example WhatsApp and Telegram, with Telegram's Bot API reference upstream.
- Silent failures. The most common chatbot ticket is "the bot stopped replying", and it is nearly always webhooks: webhook not firing or the widget not showing.
- Template rejections. WhatsApp reviews every proactive template and rejections are routine: why templates get rejected covers the causes, and the template checker catches most before you submit.
None of this argues against deploying a chatbot. It argues for a platform that absorbs it, so tokens, webhook verification and retries are handled for you and failures surface as readable status rather than raw codes.
From Chatbots to AI Agents: What Changed
The word the industry is migrating to is agent, and it is not purely marketing. A chatbot answers; an AI agent decides on a sequence of steps and carries them out - looking up an order, changing a booking, filing a ticket - by calling tools mid-conversation. Same language model; the difference is that it has hands.
In practice this is a spectrum, not a category boundary: one build can answer from your documentation on one turn and complete a task on the next. What matters is scope. An agent should be allowed a short, explicit list of actions, with a human in the loop for anything consequential and guardrails around the rest. Read the side-by-side in AI agent vs chatbot, the conceptual version in agentic AI chatbots explained, and the service-desk application in agentic AI for customer service. The builder side is the AI agent builder.
Frequently Asked Questions About Chatbots
What is a chatbot in simple words?
A chatbot is a computer program you talk to. You type or say a question and it replies the way a human assistant would. Simple chatbots follow a script written in advance; AI chatbots work out what you meant and compose a new answer.
How do chatbots work?
A chatbot receives your message, works out what it means, and produces a response. Rule-based bots match keywords against a decision tree. Intent-based bots classify the message into a trained intent and extract entities such as an order number. Generative bots pass it to a language model alongside your own documentation.
What are the main types of chatbots?
By mechanism: menu and button bots, keyword and rule-based bots, intent-based NLU bots, generative LLM bots, hybrids that combine scripted flows with AI, voice bots, and AI agents that call tools to complete tasks. By job: support, lead generation, booking, ecommerce and internal helpdesk bots. Most production deployments are hybrids.
Is ChatGPT a chatbot?
Yes. ChatGPT is an AI chatbot built on OpenAI's large language models. The difference from a business chatbot is purpose and grounding: ChatGPT is a general-purpose assistant trained on public data, while a business chatbot answers from one company's own content, sits on that company's channels, and escalates to that company's humans.
Do all chatbots use AI?
No. A rule-based chatbot is a decision tree with no machine learning in it, and it is still the right tool for a booking form or a returns workflow where you want the same path every time. AI earns its place when users type things nobody anticipated. Most platforms let you combine both.
What is the difference between a chatbot and conversational AI?
Chatbot is the older, broader word for any program that chats, including a purely scripted one. Conversational AI is the technology category underneath the smart ones: natural language understanding, large language models, speech recognition and dialogue management. Every conversational AI assistant is a chatbot; plenty of chatbots contain no conversational AI at all.
What is the difference between a bot and a chatbot?
A bot is any program that runs automated tasks - a crawler indexing pages, a trading bot, a Discord moderation bot. A chatbot is the subset whose job is conversation with a person. All chatbots are bots; most bots never talk to anyone.
How much does a chatbot cost?
From nothing to five figures a month. Most no-code builders have a free tier, small-business plans commonly run about $19 to $99 a month, and enterprise suites with AI add-ons regularly pass $500. Custom development costs the most over time: build, model usage, hosting and permanent maintenance. Messaging channels such as WhatsApp bill separately, per message.
Are chatbots safe and private?
They are as safe as their configuration. A chatbot sees whatever customers type, so treat transcripts as personal data: control who can read them, set a retention period, keep payment and health details out of chat. In the EU, the AI Act requires that people are told they are talking to an AI system.
Can a chatbot replace human support agents?
Not entirely, and the attempt usually backfires. Chatbots absorb the repetitive, well-documented majority of contacts and free humans for the rest. Emotionally charged situations, edge cases and high-stakes decisions still need a person. The deployments that work pair the bot with a visible route to a human that carries the transcript across.
How long does it take to build a chatbot?
On a no-code builder, roughly ten minutes: start from a template, edit the messages, point it at your website so it answers from your content, and paste one snippet into your site. A production bot with integrations, tested flows and a reviewed escalation path is a few days. Custom-coded bots take weeks to months.
Do I need to know how to code to build a chatbot?
No. No-code builders use a visual canvas, so designing a conversation looks like drawing a flowchart. You only need code when you want behaviour an off-the-shelf platform cannot express, or when the chatbot is itself the product you are selling.
Still hit a term you do not recognise? The chatbot and AI glossary defines every one used here, and the resources hub collects the guides, calculators and tools linked above.
The fastest way to understand chatbots is to build one
Start from a template, point it at your website, go live in ten minutes. Free plan, 600 conversations a month, no credit card.