Chatbots in Artificial Intelligence: How Much AI Is Really Inside? A 6-Level Scale and a Free Behaviour Test

Chatbots in artificial intelligence range from simple keyword scripts to systems that write fresh answers and take actions, yet almost every product page now says “AI-powered”. That label tells you very little. This guide places chatbots on a six-level scale, explains which branch of AI each level borrows from, and gives you a free behaviour test: send the bot nine messages, tick what happens, and see how much AI it actually uses and what risks come with it.

⚡ In short

  • A chatbot is an application of AI, not a branch of it. Different bots borrow very different amounts of AI.
  • Six levels: menu, keyword rules, intent matching, generative, grounded generative, and agents with actions.
  • Higher is not always better. Each level brings new abilities and new ways to go wrong.
  • You can estimate a bot’s level from how it behaves. The free test below takes about ten minutes.
Chatbots in artificial intelligence in LiveChatBot: AI Agent knowledge sources and a sample chat answered from them
How it looks in LiveChatBot: AI Agent knowledge sources and a sample chat answered from them.

The family tree

Chatbots in artificial intelligence: an application, not a field

Artificial intelligence is a broad umbrella. Under it sit several fields that matter for chat: natural language processing (getting computers to work with human language), machine learning (systems that learn patterns from examples instead of following hand-written rules), and more recently large language models (very large machine-learning models trained on huge amounts of text, which can write fluent replies).

A chatbot is none of these. It is a product that may use some, all or none of them. A bot that shows four buttons uses no AI. A bot that matches the word “refund” and sends a fixed reply uses simple text matching, which was standard long before the current AI wave. A bot that writes a new paragraph from your returns policy uses a large language model plus a retrieval step that looks up your content first.

This matters when you buy one. Two products both called “AI chatbot” can sit at opposite ends of this tree. One will never invent a price because it can only send what you wrote. The other can answer questions you never anticipated, and it can also invent things if nothing keeps it to your facts.

So the useful question is not “is it AI?” but “which parts of AI does it use, and what does that mean for my customers?” The scale in the next section answers that. If you’re new to the basics, our explainer on what a chatbot is covers the ground first.

The scale

Six levels of AI in a chatbot

Each level adds one ability. Read across to see what it borrows from AI and the risk it brings.

LevelWhat it doesAI it borrowsMain risk
0. Menu botShows buttons; visitor taps a pathNoneVisitors with other questions get stuck
1. Keyword rulesSpots set words, sends fixed repliesBasic text matchingMisses reworded questions
2. Intent matchingRecognises meaning of trained questionsMachine-learning NLUNeeds many examples; still fixed replies
3. GenerativeWrites new replies with a language modelLarge language modelsInvents facts, drifts off topic
4. Grounded generativeWrites replies from your content onlyLanguage model plus retrievalOnly as good as your content
5. AgentReads live data, takes actionsLanguage model plus tool useWrong actions, permission mistakes

Most website bots today are a mix. A common pattern is a Level 4 assistant for open questions, a Level 1 keyword fallback for when the AI is unavailable, and a Level 0 menu of buttons for the most common paths. That mix is often the most reliable setup, because each part covers another part’s weakness.

Why labels blur

What “AI-powered” can mean on a product page

Marketing pages have every reason to stretch the word. Here’s what the label can quietly mean:

  • “AI-powered” on a Level 1 bot. The product matches keywords, perhaps with typo tolerance, and calls that AI. It’s honest in a narrow sense, since fuzzy matching is a technique from the field, but you won’t get answers to questions you didn’t set up.
  • “Trained on your data” on a Level 3 bot. Sometimes the product feeds a few of your pages into a general model without any rule that it must stay within them. It sounds grounded, but ask about something your pages don’t cover and it may make up a confident answer.
  • “Autonomous agent” on a Level 4 bot. It answers well from your content but can’t actually do anything, such as look up an order or change a booking. That’s fine, as long as you know.

None of this is necessarily dishonest. Discussions of chatbots in artificial intelligence often mix three meanings: the technique used, the experience the visitor has, and the vendor’s roadmap. Your job is to separate them. Behaviour is the most reliable evidence, because you can test it yourself in ten minutes without reading any technical paper.

That’s what the tool below does. It doesn’t look inside anyone’s code. It asks you to send specific messages and record what came back, then estimates the level from the pattern.

Free tool

Test how much AI a chatbot really uses

Open the bot you’re judging in another tab. Send the test messages described in the next section, tick each behaviour you saw, and press the button.

Run each test on the bot you are judging, then tick what it actually did.

Runs only in your browser; nothing is sent or saved. This is a rough guide from behaviour, not a look inside the vendor’s code.

Running the tests

Nine messages to send, and what to watch for

Use a fresh browser window so the bot has no memory of earlier chats. Send these one at a time and note what happens. Replace the examples with words that fit the business you're testing.

  1. Typed question. Type “What are your delivery charges?” If the bot only offers buttons and ignores typing, stop here: it's Level 0.
  2. Typo. Type “delivary chargs?” A keyword bot with exact matching will fail; one with fuzzy matching will cope.
  3. Rewording. Type “how long till my parcel turns up” without using the words delivery or time. Understanding this means at least intent matching.
  4. Repeat. Ask question 3 again in a new chat. Identical wording both times suggests fixed replies; different wording suggests a language model writing fresh answers.
  5. Memory. Ask about delivery to Mumbai, then “and Chennai?” A bot that understands the second message is carrying context.
  6. Own detail. Ask about something that appears only on that business's site, such as a specific product size or a branch address.
  7. Gap. Ask something the site clearly doesn't cover, such as “Do you deliver to Antarctica on Sundays?” Watch for invention.
  8. Off-topic. Ask for a cricket poem. A business bot that writes one has no topic boundaries.
  9. Action. Ask it to check an order or book a slot, if the business offers either. Watch whether it really does, or just tells you how.

Ten minutes, and you know more than most product pages will tell you.

Reading the result

Pick the level that fits the job, not the highest one

It's tempting to treat Level 5 as the winner. For most small businesses, it isn't.

Levels 0 and 1 are ideal when you have a handful of common questions and want complete control over every word. A tuition centre with five questions about batches, fees and timings doesn't need a language model. Our rule-based chatbot guide shows how to make keyword rules that don't misfire.

Level 2 made sense when building a generative bot was expensive. Today it's mostly found in older enterprise tools. It needs lots of training examples per question, which small teams rarely have time to write.

Level 3 is the level to avoid for customer-facing chat. Fluency without grounding is how bots end up quoting the wrong return window or promising discounts that don't exist.

Level 4 suits most businesses that get varied questions: shops, clinics, coaching institutes, travel agents. It answers questions you didn't anticipate, in your own facts, and hands over when it can't help.

Level 5 is worth it only when the action saves real work and you've tested every way it can go wrong. A bot that can cancel orders needs strict limits.

The honest summary: match the level to the variety of questions you get and the cost of a wrong answer.

Where LiveChatBot sits

A grounded assistant with a rules safety net

For transparency, here's how our own product maps onto the scale.

LiveChatBot's AI Agent is designed as a Level 4 assistant: it answers from the knowledge you give it, and when it can't help, the chat goes to your team in the inbox or the Android app. You decide what it knows by writing that knowledge, and our guide to training your AI chatbot covers how.

Underneath sits a Level 1 safety net. When AI replies are unavailable or the allowance is used up, the free keyword bot answers from your saved replies. That means visitors still get your approved answers to common questions instead of silence.

Flows add Level 0 and rule-based behaviour on top: proactive messages triggered by conditions such as page, source, country, device, number of visits or lead score, plus a chat rescuer for chats nobody has picked up.

It doesn't take actions in your other systems by itself. If you want chat to create records elsewhere, the REST API, webhooks and Zapier connect it to other tools, which keeps the risky part, changing data, under rules you set.

The free plan covers two websites with 100 visitors per website each month, with no per-seat charges. See pricing for visitor packs.

Questions for vendors

Seven questions that reveal the level

When a demo looks impressive, these questions get past the label:

  • “What happens when a visitor asks something I never set up?” Level 1 says “it shows a fallback”. Level 4 says “it answers from your content or hands over”. Level 3 often goes quiet on this one.
  • “Where do its facts come from?” Look for a clear answer: your pages, your documents, your written answers.
  • “Can I stop it answering outside my business?” If not, expect cricket poems.
  • “What does it do when the AI is down or my allowance runs out?” A good answer includes a fallback that still helps.
  • “Can I see and edit what it knows?” You should be able to fix a wrong fact in minutes.
  • “Which actions can it take, and who approves them?” Only relevant at Level 5, but critical there.
  • “How do I hand a chat to a person?” Every level needs a human path.

Write the answers down and run the behaviour test yourself. Claims about chatbots in artificial intelligence are easy to make; behaviour is harder to fake. For a wider comparison of bots and human chat, see chatbot vs live chat.

Myths

Common beliefs, corrected

🚩 MythAny bot that understands typos must be AI.
✅ RealityFuzzy matching handles typos with simple maths. It's useful, but it doesn't understand meaning.
🚩 MythA bot trained on my website will never invent things.
✅ RealityOnly if it is told to stay within that content and admit gaps. Test it with questions your site can't answer.
🚩 MythMore AI means fewer complaints.
✅ RealityMore AI means more flexibility. Complaints drop only when the bot's knowledge is accurate and handover works.
🚩 MythButton bots are outdated.
✅ RealityButtons are still the fastest path for common choices, and many good bots combine them with AI.
🚩 MythAn AI bot replaces the team.
✅ RealityIt handles routine questions. People still handle exceptions, complaints and sales conversations.

Getting started

Test before you trust

Before you choose a bot for your website, run the nine-message test on two or three products and record the level each one reaches. Then decide which level suits your questions. If most of your enquiries are the same ten questions, Level 1 with good saved replies may be all you need. If they vary widely, a grounded Level 4 assistant with a keyword fallback is usually the safest choice.

You can try LiveChatBot on your own site for free at livechatbot.in: set up the widget, write your saved replies and your AI knowledge, then run the same test on your own bot. It's the quickest way to see what your visitors will experience.

Frequently asked questions

What role do chatbots in artificial intelligence play?

They're an application of AI fields such as natural language processing, machine learning and language models. A given chatbot may use several of these or none.

Is a keyword bot real AI?

Not in the modern sense. It uses text matching, which is predictable and cheap. Some vendors still call it AI.

What does grounded mean?

A grounded bot writes answers from content you provide, such as your pages or written answers, and says it doesn't know when the content doesn't cover the question.

Which level do most small businesses need?

Usually a grounded assistant for varied questions, with a keyword fallback and a clear path to a person.

Can I test a bot without technical knowledge?

Yes. Send the nine test messages in this guide and tick what happens in the free tool. It takes about ten minutes.

Keep reading

Related guides

📘

Companies With Chatbots: 9 Patterns Big Brands Use That a Small Business Can Copy This Week

Read the guide →

Run the test on your own bot

Set up a grounded AI assistant with a keyword fallback and a human handover. Start free on livechatbot.in.