The difference is that a regular AI chatbot follows rules someone wrote by hand, while an LLM (large language model) generates its replies word by word from patterns it learned during training, so it can answer questions nobody programmed it for.
A scripted chatbot matches your message to a fixed list of intents and returns the response attached to that intent. An LLM predicts the next chunk of text based on everything in front of it. Same chat window, completely different engine underneath.
What a scripted chatbot can and cannot do
Think about the automated helper on a bank website. You type "what's my balance" and it replies with a balance link. You type "card lost" and it starts the replacement flow.
Behind that is a decision tree: the system looks for keywords or patterns, picks the closest matching intent, and fires the response a human wrote for it. The bot never writes a sentence. It selects one.
That is why these systems are cheap, predictable, and fast, and also why they fall apart the moment you phrase something in a way nobody anticipated. Ask the same bot "my card went missing while I was abroad and I'm worried about fraud" and it may return the generic help menu, because no single intent covers that sentence.
The failure isn't stupidity. It's the architecture. There is no response in the list that fits, so the bot guesses or gives up.
What makes an LLM different
An LLM is trained on enormous amounts of text and learns statistical patterns about how language fits together. When you send a message, it doesn't search a list of canned replies. It generates a response token by token, where a token is a small piece of a word, choosing each one based on probability given everything that came before. That is why the same question can produce a slightly different answer twice. Nothing is hard-coded. The model has no stored reply for "my card went missing abroad." It composes one on the spot.
This is also why ChatGPT and Claude count as LLM-based chatbots rather than regular ones. According to our AI tool database, ChatGPT is OpenAI's flagship assistant and Claude is Anthropic's safety-first assistant, and both are built around large language models rather than intent lists. When people say "AI chatbot" today, they usually mean this kind, which is exactly why the old distinction has become confusing.
A worked example
Say you run a small bike shop and you want a helper on your site. With a scripted bot, you'd sit down and write intents: "store hours," "return policy," "shipping cost," "book a repair." Maybe forty intents total.
A customer asks "can I bring my bike in Saturday if the rear wheel is wobbling?" The scripted bot has no intent for that combination. It might match "Saturday" to store hours and send the wrong answer.
An LLM-based assistant handles it because it can read the whole sentence, notice the repair request and the day, and reply with something like "yes, bring it in Saturday, and mention the wobbly wheel so we can check the spokes." No one wrote that sentence. The model produced it from patterns.
The trade-off is real. The scripted bot will never invent a discount that doesn't exist. The LLM might, which is why the same flexibility that makes it useful also makes it worth checking. That problem has a name, and it's covered in our explainer on what an AI hallucination is and why AI tools make things up.
Where the line blurs, and what it costs
Here's the honest part: most things marketed as "AI chatbots" in 2026 are LLM-backed, so the two categories have merged in everyday speech. When a company says it uses AI for support, you often can't tell which kind without asking. The distinction still matters because it predicts behavior. Scripted bots are consistent and narrow. LLMs are flexible and occasionally wrong. If you're choosing a tool, ask one question: can it answer something nobody wrote a script for? If yes, you're dealing with a language model.
The limits run both ways. LLMs cost more to run because every reply is generated, and pricing varies by vendor, so check the vendor's own page rather than trusting a number you saw months ago. They can also be slower and harder to audit than a decision tree, which matters in regulated industries where you need to prove exactly why a system said what it said. Scripted bots win there. Neither approach is universally better. It depends on whether your problem is "handle these twelve known requests" or "handle whatever a human types."
One practical tip: when you're evaluating any chatbot, throw it a sentence that mixes two topics and uses unusual phrasing. Scripted bots stumble. LLMs usually cope. That single test tells you more than any feature list.