What Is a Chatbot? And How Does It Decide What to Say?

Chapter 1: What a chatbot actually is, in one plain paragraph

A chatbot is a program that turns a text message into a useful text reply without a human typing it. That is the whole definition. Everything else — the terms like “AI chatbot”, “conversational AI”, “virtual assistant” or “auto-responder” — describes the same machine with different levels of brainpower.

At the low end sit fixed scripts: if the customer types “timing”, the bot sends the exact message you wrote about timings. Predictable, fast and completely incapable of going off-script. At the high end sit AI replies, where a language model writes a fresh sentence every time, using your own approved facts as the only source of truth. Between the two sits the hybrid most small businesses end up with: scripts for the ten predictable questions, AI for everything else, a human for anything sensitive.

One useful correction before we go further: a chatbot is not a salesperson and it is not a decision-maker. It is a triage layer. It absorbs the repetitive, predictable, information-seeking traffic so that your attention — the scarce, expensive thing you actually sell — lands on the conversations that need a human. If a vendor ever promises you a bot that “closes sales for you”, run.

Chapter 2: Step one — reading the message and working out intent

The very first thing your chatbot does is decide what kind of request it is looking at. This is called intent detection, and it is the least glamorous and most important part of the whole system. Get this wrong and everything downstream fails: a pricing question treated as a complaint produces a cold, unhelpful answer.

Modern systems do not just match keywords. The message is converted into a mathematical fingerprint called an embedding, and that fingerprint is compared against the fingerprints of every intent you defined. “Do you have any slots for Saturday morning?” and “Need appointment this weekend, any time?” share almost no words but land in the same place: book an appointment.

Notice what is not on that list. Sentiment analysis is the other half of this step: a bot that detects mild irritation and quietly routes the chat to a person prevents your worst reviews. Research is consistent on this — a 2025 peer-reviewed evaluation of AI-chatbot adoption found significant measurable effects on response time and on how customers perceive the interaction, which is precisely the trade-off you are managing.

Chapter 3: Step two — the knowledge base, where the answer comes from

Here is where most “AI” explanations go wrong. A language model on its own knows a spectacular amount about the world and knows nothing whatsoever about your business. It does not know your timings, your fees, your stock, your refund policy or which doctor is on duty. Left to answer freely, it will confidently invent all five. That behaviour has a name: hallucination.

So the second step is retrieval. Your approved information — price sheets, FAQs, service pages, policy documents, a live spreadsheet of stock or slots — is chopped into small chunks, given a fingerprint, and stored in a searchable library called a vector database. When the bot has worked out the intent, it searches that library for the chunks closest in meaning to the customer’s question, pulls the best few, and hands them to the writing step along with strict instructions: answer only from these, and if they do not contain the answer, say so and pass it to a human.

The industry name for this pattern is Retrieval-Augmented Generation, usually shortened to RAG. It matters for you commercially, not just academically. Analysts benchmarking enterprise RAG systems in 2025 found they still hallucinate at rates above 10% on real-world queries — which is a reality check worth internalising. Grounding dramatically reduces the risk; it does not eliminate it. That is exactly why every serious implementation also includes a refusal path and a human handover.

Chapter 4: Step three — turning retrieved facts into a worded reply

Now the bot writes. It receives three things: the conversation so far, the retrieved facts, and your instructions — usually written as a short “operating brief” that sounds remarkably like a job description. For example: You are the front desk for a dental clinic in Lucknow. Answer in simple friendly English or Hindi. Never give medical advice. If asked about discounts, say a manager will call back. Keep replies under four sentences. Always end by offering a next step.

The model then predicts the reply one token at a time, choosing each next word based on everything that came before. This is why the same question can get two different sentences — and also why the same question can get a wrong sentence. The model is optimised to sound plausible, not to be correct. Correctness comes entirely from what you allowed it to read in step two.

Two settings matter here. Temperature controls how adventurous the wording gets: low keeps it consistent and safe, high makes it varied but riskier. For business replies, low-to-medium is the sensible default. Guardrails are the hard rules layered around the output — banned topics, banned promises, a required format, and an automatic trigger to hand over when a complaint, a medical question or a high-value negotiation is detected.

• Intent → Retrieval → Generation. Three steps, every time, for every message. There is no fourth step, no secret sauce.
• Garbage in, garbage out. A chatbot is only ever as good as the information you give it. Stale price sheets are the number one cause of embarrassing answers.
• Knowing when to quit. A good bot says “I’ll have the team call you” more often than a nervous one. Refusing gracefully is a feature, not a failure.

Chapter 5: Fixed scripts versus AI replies — and why the difference decides quality

This is the single most important decision you will make, and it is often made for the wrong reason. Scripted bots are sold as “safe but limited”; AI bots are sold as “smart”. Both descriptions are half wrong. A scripted bot is not limited because of its architecture — it is limited because you decide how many variations to write. Twenty well-planned branches handle more than most businesses ever need.

The honest comparison looks like this:

The pattern that works for almost every local business is a layered setup. Layer one: scripts for the questions you know cold — timings, address, what you do, how to book. Layer two: AI, grounded on your real documents, for everything else. Layer three: a human, instantly, for complaints, medical or legal advice, negotiations and anything the bot is not confident about.

What this layered structure really gives you is speed. When the bot handles the repeatable half of a conversation, your team only ever sees leads that have already been qualified — and that is the argument for pairing a chatbot with automatic follow-up, so the conversation does not simply die the moment someone walks away from the chat window.

Chapter 6: Five tests any business chatbot should pass before you trust it

Do not take a demo at face value. Demo conversations are rehearsed. Run these five tests yourself, with your own real questions, on your own phone. Ten minutes, and you will know more than any sales deck can tell you.