A couple of years ago, a chatbot on a website meant a set of buttons and canned answers: press “1” to get the price, press “2” to reach a manager. Today a bot increasingly means a system built on a large language model that understands a question asked in plain words and answers on the merits rather than following a script. For a business this sounds like a dream: the customer writes at any hour, gets an answer in seconds, and managers handle only the conversations that truly need a human.
But there is a long way between a conference demo and a bot working in a real company. A bot can confidently tell customers things that are not in your price list, miss the context, or become one more point of data leakage. In this article we look at how such bots work, which tasks they really cover, how to roll them out without big losses and which risks to watch for in advance.
How an AI bot differs from a regular chatbot
A classic bot follows a pre-written tree of scenarios: every possible question and answer was invented by a person. This is predictable and safe, but fragile: as soon as a customer phrases a question in an unforeseen way, the bot replies “I didn't understand you.” Such a bot is also hard to maintain, because every new exception has to be added to the script by hand.
An AI bot understands the meaning of a question, not just matching words, and composes the answer itself based on company materials. Flexibility comes at a price: an answer can no longer be guaranteed a hundred percent, because the model generates text rather than picking from a ready list. So during implementation the key question is not “how smart is the model” but “how will we limit and verify what it tells customers.”
Many companies end up choosing a hybrid: the AI understands and answers typical questions, while sensitive actions — payment, refunds, contract changes — go through rigid scripts or a human.
Which tasks a bot handles well and which it doesn't
AI bots handle repetitive informational questions best: working hours, addresses, delivery and payment terms, product composition and specifications, order status, preparation for a service, answers to typical “how does it work” questions. Such questions take up a significant share of the time of managers and support staff, and this is exactly where a bot pays for itself quickly, because it answers instantly and at any time.
Initial qualification of inquiries works well too: the bot clarifies what the customer needs, collects a contact and budget, books a consultation and passes the ready lead to the CRM so the manager does not start the conversation from scratch. This continues the sales automation logic we covered in our articles on CRM and round-the-clock sales.
A bot copes worst with non-standard, emotionally charged and legally significant situations: a complaint, a disputed refund, price negotiation, a complex individual case. Here a mistake is expensive, and the customer needs to feel heard by a human.
How an AI bot works: knowledge base, model, integrations
A modern business bot usually rests on three layers.
Layer one — the knowledge base
Your documents, price lists, regulations, product descriptions, answers to frequent questions. This is the most important part, because the quality of the bot's answers depends directly on the quality and freshness of the materials it relies on. If the base has contradictions or outdated prices, the bot will confidently repeat those errors.
Layer two — the language model
It understands the question, finds the relevant fragments in the knowledge base and formulates an answer in the set tone. Proper setup includes instructions on what the bot can and cannot discuss, how it addresses the customer and what it does when it is not sure of an answer.
Layer three — integrations
Connections to the CRM, catalog, booking system and order statuses. Thanks to them the bot not only talks but acts: checks availability, creates a lead, books a time slot. External systems must be connected securely and with the minimum necessary permissions.
Channels: website, WhatsApp, Telegram, Instagram
A bot is useful where customers already talk to you. For Kazakhstani businesses this is most often WhatsApp and Instagram, then Telegram and a chat widget on the website. Choose the channel by the behavior of your audience, not by fashion: if most leads come from a messenger, the bot should live there, not only on the site.
Technically, connecting to messengers goes through the platforms' official interfaces, which have their own rules, limits and terms of use, so it is worth studying them before development starts.
A good practice is a single point where all inquiries from different channels land in the CRM and in one history of communication with the customer; we described how this works in our article on integrating WhatsApp and Instagram with a CRM. Then a manager joining a conversation sees everything the customer has already discussed with the bot.
Handoff to a human: the most important feature
If you pick one feature that trust in the bot is built on, it is a quick and clear handoff of the conversation to a live employee. A customer must be able to say “get me a human” at any moment and receive it without arguing. A bot that stubbornly “holds the line” and won't release an irritated customer does worse than no bot at all.
A well-designed handoff includes several conditions:
- The bot itself recognizes situations that need a human: a complaint, a refund request, a complex question, repeated misunderstanding
- On handoff the manager receives a short summary of the conversation and the customer's data, so the customer isn't forced to repeat everything
- When no managers are online, the bot honestly says when the customer will be contacted and records the request
This approach turns the bot from an obstacle into the first effective stage of service.
Ask a question or leave a request
We'll start with a narrow task and a pilot: prepare the knowledge base, set up the bot in the right channel, handoff to a manager and leads into the CRM. We'll expand features only after confirmed results.
By submitting the form, you agree to the processing of personal data.
Risks: model errors, data, security
Model errors (“hallucinations”)
The main risk is so-called “hallucinations”: a model can confidently state something that is not in your materials, such as a non-existent promotion or a wrong price. This risk is reduced in several ways:
- limit the bot to answers based on the knowledge base
- set clear bans on promises outside its authority
- add checks for sensitive answers
- regularly review real conversations to find and fix weak spots
Errors cannot be eliminated completely, so it matters that the consequences of any single error are limited: the bot must not change prices on its own, confirm discounts outside the rules or make legally significant decisions.
Personal data
Customers write names, phone numbers, addresses and sometimes more sensitive information to the bot, which means personal data protection requirements, covered in a separate article of ours, apply to such a system. You need to understand where conversations are stored, who has access to them, whether data is passed to an external model provider and on what terms.
The bot's own permissions
The more actions the bot is allowed to perform in your systems, the more carefully access must be limited. The recent PixelLeak case, which we also wrote about, showed how an autonomous AI agent without proper limits can cause a leak without anyone hacking anything. The principle is simple: give the bot the minimum necessary permissions and nothing beyond.
How to implement: from pilot to scale
A sensible path starts with choosing a narrow task rather than trying to replace all support. For example, the bot answers the ten most frequent questions and collects leads in one channel. During preparation the knowledge base is gathered and put in order: this is usually where you discover documents are outdated or contradict each other, which is worth fixing regardless of the bot.
Then comes a pilot with a limited audience or at certain hours, with all conversations reviewed by people. This shows which questions the bot doesn't understand, where it errs and which scenarios should be fixed rigidly. Only after stable results is the bot's area of responsibility expanded: channels, integrations and new topics are added.
It is important to appoint an owner right away who regularly updates the knowledge base and reviews conversations, because an unattended bot becomes outdated over time and starts making mistakes.
How to measure results
Without metrics it is impossible to tell whether the bot is useful. Basic indicators:
- the share of inquiries the bot closed without human involvement
- first response time
- the share of handoffs to a manager and the reasons for them
- the number of leads collected and their quality
- customer ratings of the conversation, if you collect them
It is important to look not only at “how much the bot answered” but also at whether customers were satisfied and reached a purchase.
The economic effect is estimated through saved employee time, the number of inquiries handled outside working hours and leads that used to be lost to slow replies. Baseline figures must be recorded before launch, otherwise there is nothing to compare with. We described the approach to calculating payback in our article on ROI metrics.
How much it costs and what the price depends on
An exact cost cannot be named, because it depends on several factors: the volume and condition of the knowledge base, the number of channels, the depth of integrations with the CRM and other systems, security and data hosting requirements, and whether you need custom development or a ready-made platform will do.
A typical cost structure includes development and setup, knowledge base preparation, payment for model and channel usage (often dependent on conversation volume) and support.
A practical tip is to start with a ready-made solution or a small pilot and decide on deeper custom development only after real results are visible. That way you pay for proven value rather than a hypothesis. When comparing offers, look not only at the development price but at the cost of ownership: how much support and growth in conversation volume will cost.
Common mistakes
- Launching a bot without a prepared knowledge base: then it answers with generic phrases or invents details
- No handoff to a human: trying to “retain” the customer in a conversation with the bot at any cost causes irritation and lost customers
- Giving the bot too broad authority, such as changing prices or promising discounts, without limits and checks
- No owner for the bot and no review of conversations — quality gradually declines
- Forgetting the legal side: the customer isn't told they're talking to a bot and personal data handling isn't thought through
- Expecting the bot to replace staff entirely: in reality it takes over the routine, while the most important and complex conversations stay with people
Pre-launch checklist
- A narrow task is chosen for the first stage, not “everything at once”
- The knowledge base is assembled, up to date and free of contradictions
- Channels where customers actually talk have been identified
- Fast handoff to a human with context preserved is set up
- The bot's authority is limited: it does not change prices or make legally significant promises
- Personal data protection and the storage location of conversations are thought through
- It is clear to the customer that they are talking to a bot
- The bot is integrated with the CRM, leads reach managers
- A pilot with human review of conversations is planned
- An owner is appointed for updating the knowledge base and reviewing conversations
- Baseline metrics are recorded for evaluating the result