A customer lands on your site at 11pm and asks whether the blue jacket comes in a bigger size. Nine seconds later they have an answer, a size chart, and a link to the returns policy. That is a chatbot working well. It is also why so many people groan when one pops up in the corner of the screen. The gap between the two comes down to which type of system you build and how carefully you prepare it — both far less technical than the jargon suggests.
Rule-based bots and learning bots
Almost every chatbot belongs to one of two families.
Rule-based bots follow a script. If someone types "track my order", show the tracking form; if they type "returns", show the policy. You build them with a flow builder — boxes, arrows, fixed answers. They are cheap to run, predictable, and easy to audit. Their weakness is literal-mindedness: a bot that knows "track order" may have no idea what to do with "any news on my parcel?"
Learning bots are trained on large amounts of language instead of being written line by line. They handle typos, slang, and messy grammar, and can answer questions nobody wrote a rule for. The trade-off is control — they can sound confident while being wrong, so they need guardrails, a defined scope, and someone reviewing what they say.
Most real setups are hybrids. The scripted layer handles actions where accuracy is non-negotiable: resetting a password, checking a balance, booking a slot. A learning layer handles the open-ended questions in between.
How a bot reads a message
Intents: what the person wants
An intent is the goal behind a message, stripped of wording. "I want to cancel", "how do I stop my subscription", and "please close my account" are three sentences with one intent. Bots sort messages by intent, then decide what to do about it.
Entities: the details inside it
Entities are the specifics: a date, an order number, a city, a product name, a size. If intent is the verb, entity is the noun. "Move my Tuesday appointment to Friday" carries a rescheduling intent and two date entities, which the bot can reuse later in the conversation.
NLP: the machinery in the middle
Natural language processing is the umbrella term for everything that turns raw text into something a machine can act on: splitting sentences, spotting names and numbers, detecting language, reading tone. You rarely touch it directly, but it explains why one bot shrugs off a typo and another collapses at a single misspelling.
Training phrases and confidence
To teach a bot an intent, you give it examples, usually called training phrases. Variety beats volume — ten differently worded examples teach more than forty near-copies of the same sentence. When a real message arrives, the bot scores how closely it matches each intent. That score is its confidence. Anything below your threshold goes to a fallback or a person rather than a guess.
The vocabulary you will keep hearing
- Utterance: one message from a user, start to finish.
- Context: what has already been said. It is what lets a bot understand "the second one" three messages later.
- Fallback: the default reply when nothing matches. A good one offers a next step; a bad one apologises in a loop.
- Prompt: the instructions and background you give a generative model before it answers.
- Hallucination: a fluent, plausible answer that happens to be false — the main reason to keep a bot's replies grounded in your own documents.
- Guardrails: rules that limit what a bot may say or do, from forbidden topics to actions it must confirm first.
- Handoff: passing the conversation to a person, ideally with the transcript attached.
- Knowledge base: the help articles, policies, or PDFs a bot searches before answering.
Why bots sound robotic, and how to fix it
Most complaints trace back to a few fixable habits. Replies run too long, so people skim and miss the point. The bot asks for information the user already gave. It answers every unmatched message with the same flat apology, or ignores the question and pushes a link instead.
Three changes do most of the work. Keep replies to two or three short sentences, because people read chat in glances. Acknowledge before answering: "Got it — that order shipped Tuesday." And write fallbacks that move things forward, offering two or three concrete options rather than "Sorry, I didn't understand."
Then read the logs. Once a week, look at the messages that ended in a fallback. Each one is either a new intent to add or a document your knowledge base is missing. That habit improves a bot faster than any rebuild.
Channels, handoffs, and the limits
Where a bot lives changes how it should talk. A website widget can show a size chart and a button. Messaging app users expect quick, conversational replies. Text messages punish long answers. Voice needs short turns and no menu with seven options, because nobody remembers seven options out loud.
Decide early when a human takes over: when someone asks for one, after two failed attempts to understand, when the tone turns frustrated, or whenever the topic is on a sensitive list such as billing disputes or complaints. Pass the whole conversation across so the person does not have to start over — that detail shapes how the entire experience feels.
If your bot handles health, money, or legal questions, keep a qualified person in the loop and route anything specific to them. A chatbot can explain a general process; it should not be the final word on someone's medication, debt, or contract.
Your first bot, in five steps
- Pick one narrow job. Answering "where is my order?" well beats answering everything badly.
- Collect twenty real questions from your inbox, support tickets, or site search.
- Write short answers in your own words, then check them against your published policies.
- Build the escape hatch first: a visible "talk to a person" path and a fallback that offers options.
