Answer in brief
A chatbot is software that talks to customers by text in a website widget or a messenger. A scripted bot leads people through buttons; a bot on a language model understands free text but can state something false. For a small company it closes frequent questions, requests, bookings and order status at any hour, provided it has a written knowledge base, a handoff to a person and an owner who reads the dialogues.
A chatbot is a program that talks to customers in a chat window
Type chatbot for business into a search engine and you get a wall of promises, so start with a plain definition. IBM gives one: a chatbot is a software application that communicates with people through text or voice. It lists where such programs sit: websites, messaging apps, SMS, WhatsApp, customer service portals. Picture a dental clinic. A patient asks at eleven in the evening about a free slot on Thursday, and within seconds something offers two times. That something is the bot.
An older, narrower definition is worth keeping beside it. Nielsen Norman Group, in a usability study, called a chatbot a domain-specific text-based conversational interface that supports users with a limited set of tasks. The last five words carry the weight. A bot is not an employee and not a digital copy of the owner. It is a narrow tool, and the companies content with theirs chose its few tasks on purpose.
Two kinds: a scripted bot with buttons and a bot on a language model
IBM separates two families. Traditional chatbots follow predefined rules, decision trees and scripted conversation flows. To the customer this looks like a menu inside a chat: press “Book a visit”, pick a service, pick a day, leave a phone number. By IBM’s account such bots work well for predictable requests and struggle with questions outside their programmed responses. Nielsen Norman Group noted their useful side: predetermined links and buttons saved users from typing.
The second family is the AI chatbot, built, in IBM’s words, on natural language processing, generative AI and large language models. It reads an ordinary sentence, keeps the thread and composes a reply instead of picking one from a list. IBM clears up a common confusion in one line: every AI chatbot is a chatbot, but not every chatbot is an AI chatbot. Nothing forces a choice: one bot can offer buttons for the usual routes and a model for whatever is typed by hand.
Where the bot lives: a website widget, Telegram, WhatsApp
The most familiar home is the widget, a small chat window in the corner of a site. It catches a visitor at the moment a question appears, before they leave to compare someone else’s page. The weak side: a visitor who closes the tab is often gone for good, so the widget should ask early how to reach them. A messenger keeps the thread, and the customer can return to it a week later.
Telegram describes its bots as small applications that run entirely within the Telegram app and says the platform is free for both users and developers. A bot is registered by writing to @BotFather, and its chat carries a “bot” label, so nobody takes it for a person. One line matters for planning: bots can’t start conversations with users, so the customer writes first. The documentation also covers keyboards with predefined reply options and bots that Telegram Business users connect to answer messages on their behalf.
WhatsApp is stricter. Meta’s developer documentation describes a customer service window: a user’s message starts a 24-hour timer that resets with each new message, and once it closes only pre-approved template messages can be sent. The WhatsApp Business Messaging Policy, in its version of 23 September 2026, requires opt-in permission before a business contacts people. It allows automation only with prompt, clear and direct escalation paths, such as a transfer to a human agent, a phone number or a support form. Such rules change, so read the current text first.
The tasks a bot really closes for a small company
The list is shorter than sellers suggest, and that is good news. IBM names the common uses: answering product and policy questions, helping customers track orders, booking appointments, qualifying leads and collecting customer information. For a small company that means five jobs: answer the questions that repeat daily, take a request with contact details, book a visit, say where an order is, and do it all at night. On the last point IBM is brief: unlike teams with limited business hours, chatbots can respond at any time of day.
What these jobs share is their shape: each is short, predictable and has an obvious end. Nielsen Norman Group saw the same from the user’s side: the tasks bots support are best conceptualised as linear flows with a limited number of branches. Its advice was not to be overly ambitious and to create bots for simple tasks. A complaint, a dispute about money or an unusual order has a different shape. Those belong to a person from the first message, and a chatbot for business earns its keep by leaving them alone.
Where a bot fails and why the handoff to a person is designed in
Scripted bots fail at the edge of the script. In the Nielsen Norman Group study, problems occurred as soon as users deviated from the prescribed flow, and people were annoyed when the same answer came back over and over. Customer service bots were perceived as generally less helpful than human representatives. The study dates from 2018 and had eight participants, so it describes the scripted generation.
Bots on a language model fail differently: they answer smoothly and sometimes wrongly. IBM states that AI chatbots can occasionally generate responses that are inaccurate, incomplete or misleading. A separate Nielsen Norman Group article defines such a hallucination as output that seems plausible but is incorrect or nonsensical, often presented confidently. Supplying the model with retrieved documents, it says, has reduced the problem without entirely fixing it. In business terms, a bot may quote a price or a return term that does not exist.
That is why the handoff is a part of the product, not an admission of defeat. IBM writes that a well-designed chatbot should recognise its limitations and provide a clear path to a human representative. Nielsen Norman Group says it more plainly: be honest about not understanding and offer an escape hatch in the form of a real human or a phone number. Its participants preferred an honest “I don’t understand” to a blatantly wrong answer. Both sources also want people told upfront that they are talking to a bot.
What a bot needs from the business: knowledge base, scenarios, CRM and an owner
A bot knows only what it was given. First it needs a knowledge base: the real answers of this company, written down. Opening hours, what a price depends on, delivery and return terms, what the company does not do. IBM counts consistent information among a chatbot’s benefits because it replies from approved content and business rules. If those answers live only in the owner’s head, the project starts with a notebook, not with software.
Next come scenarios: the few routes a conversation takes, each drawn from the greeting to the last message, including the branch that calls a person. Then a link to wherever requests are kept, a CRM, a calendar or a shared table. Last comes an owner, one named person who reads the dialogues and corrects the answers; IBM notes that chatbots require regular updates as business information, policies and customer needs change. Chats also hold personal data, and which privacy law applies to you is a question for a specialist.
How to tell that the bot works: the outcomes worth counting
Counting messages says little; count how conversations end. Microsoft’s documentation for Copilot Studio, its tool for building conversational agents, sorts every engaged session into three outcomes. Resolved means it ended successfully, confirmed by the user or implied. Escalated means a handoff to a person was triggered, with the reason recorded: a rule the builder set, the user’s own request, or a threshold showing that the user was stuck. Abandoned means the session timed out, after 30 minutes in that product, without reaching either result. Other platforms count in their own way, so read the definitions in the one you use.
The same documentation describes two more figures: a satisfaction score out of 5, averaged from the surveys users fill in after a session, and an answer rate, meaning how many questions the bot answered and how many it could not. The unanswered list is the most useful page of any report, a ready plan for the knowledge base. Then look at what the business wanted in the first place: requests received, visits booked. A bot with many dialogues and no requests is entertainment.
How to start small and what to prepare before ordering a bot
Start with one channel and one task. Choose the channel where customers already write and the task that repeats most, then collect the real questions from your chat history and write the answers you would give yourself. Decide who takes over when the bot hands a conversation on, and during which hours. Read every dialogue for the first weeks and extend the bot only where the transcripts ask for it.
Whether to build it yourself depends on the kind. A scripted bot for one task in one channel can be assembled by a patient owner with a visual builder. A bot on a language model that reads your documents, writes to the CRM, hands over to an operator and reports its outcomes is a project with testing and upkeep, and ordering it is reasonable. Bring the question list, the answers, the routes and the name of the person responsible. VITON13 Studio is one of the places that builds such bots; the linked guide and service page give the scope and the prices.
Practical checklist
- Pick one channel where customers already write to you and one task that repeats most often.
- Collect the real questions from your chat history and write a short approved answer to each.
- Draw each route from the greeting to the last message, including the branch that calls a person.
- Name the person who receives handed-over chats and set the hours when they answer.
- Read the dialogues every week and add every unanswered question to the knowledge base.
Questions and answers
Does a small company need a chatbot with AI, or are buttons enough?
For one or two predictable tasks, such as booking a visit or taking a request, a scripted bot with buttons is often enough, and it never invents an answer. A language model pays off when customers ask in their own words about many products or terms. The two can be combined in one bot.
Can a chatbot replace a manager or a support employee?
No. It takes over the short, repeating conversations and collects requests outside working hours. Complaints, disputes, unusual orders and anything that needs judgement still go to a person, which is why a handoff is built into every sound design.
Can a Telegram bot write to a customer first?
According to Telegram’s documentation, bots can’t start conversations with users: the person has to open the bot and send the first message. After that the bot can reply. Rules of this kind change, so check the current documentation before building a scenario around them.
Why does a bot on a language model sometimes give a wrong answer?
A language model works out the most likely next word, so it can produce a reply that sounds right and is false. IBM says this is more likely when the bot lacks reliable information. A knowledge base, a narrow subject and a handoff lower the risk; Nielsen Norman Group writes that it is not yet known whether it can be eliminated.
Which numbers show that a chatbot for a small company is working?
Look at how conversations end rather than at how many there were: the share resolved, the share handed to a person and why, the share abandoned. Add the satisfaction score, the list of unanswered questions and the business result, meaning requests and bookings received through the bot.