You can build an AI chatbot in an evening: connect a language model API and write an instruction like "you are a polite company assistant." The trouble starts with the first few dozen real customers. The bot promises a discount that does not exist, mixes up pricing plans, invents delivery times and happily discusses a pasta recipe instead of your product.
The gap between a demo and a production assistant is not the choice of model but what is built around it: the knowledge base, search, guardrails, handoff to a human, integrations and quality control. Here is what an AI assistant for business is made of and how to get it to the point where you can trust it with customers.
How an AI assistant differs from a regular chatbot
A classic bot follows a script: buttons, branches, prewritten answers. It is predictable, but it breaks as soon as a customer writes in their own words or asks something that is not in the tree.
An LLM-based assistant understands free text, rephrasing and typos, can ask a clarifying question and assemble an answer from several sources. The flip side is less predictability: the model can answer confidently and wrongly.
| Scripted bot | LLM assistant | |
|---|---|---|
| How it works | Buttons, branches, prewritten answers | Language model, knowledge base, search |
| Free text | Breaks on questions outside the tree | Understands rephrasing and typos |
| Predictability | High | Lower: can answer confidently and wrongly |
That is why production projects almost always end up hybrid. A language model handles open-ended questions, while actions with consequences (placing an order, booking, a refund) follow a strict script with buttons and checks. A business has no use for a customer support bot that "just chats." It needs one that solves the customer's problem or quickly hands it to a person.
A business does not need a bot that "just chats." It needs one that solves the customer's problem or quickly hands it to a person.
How an LLM-based assistant is built
Take apart a typical AI bot for a website or Telegram, and you will find these components inside:
- Channel. A website widget, a Telegram bot, WhatsApp or a chat in a mobile app. There can be several channels with one shared logic.
- Backend orchestrator. Receives the message, stores the conversation history and decides what to do next: search the knowledge base, call the CRM or hand off to an agent.
- Language model. A cloud model via API or one deployed in your own infrastructure. It handles understanding the question and phrasing the answer.
- Knowledge base and search. The documents the model takes facts from.
- Tools. Functions the model can call: check an order status, find an open booking slot, create a ticket.
- Logs and admin panel. Where you see all conversations, ratings and handoffs, and where the knowledge base is edited.
A single message travels like this: the customer asks a question → the orchestrator finds relevant passages in the knowledge base and, if needed, pulls data from the CRM → the model receives the instruction, the conversation history and the facts found → it drafts an answer → the system checks it against the rules and sends it to the customer or calls an agent.
The knowledge base and RAG
RAG (retrieval-augmented generation, generating answers grounded in retrieved documents) is the standard way to make a model answer from your data without fine-tuning. Documents are split into meaningful chunks, and a vector representation is built for each. When a question arrives, the system finds the chunks closest in meaning and passes them to the model with the instruction "answer only based on this."
In practice, an assistant's quality is defined first by the knowledge base, not the model. What matters in it:
- A single source of truth. If one document says delivery takes 3 days and another says 5, the model will pick either one.
- Freshness. Every section needs an owner and a clear update process.
- Structure. Short sections with clear headings and question-and-answer phrasing are found more easily than a hundred-page policy manual.
- Live data, not text. Prices, stock and order statuses are better fetched from the back-office system at the moment of the question than copied into documents that will be outdated tomorrow.
An assistant's quality is defined less by the model than by the knowledge base.
Why the bot "lies" and how to fix it
A language model does not look for the truth; it generates plausible text. When facts are missing, it fills them in itself. Typical causes of made-up answers:
- the knowledge base has no answer, and the model fills the gap;
- search found the wrong passage, similar in wording but about something else;
- the knowledge base contains contradictory documents;
- the instruction is too loose and does not forbid guessing;
- the data is outdated.
What to do about it:
- Allow "I don't know." An explicit instruction: if the retrieved passages do not contain the answer, the bot says so honestly and offers to connect the customer with a team member.
- A relevance threshold. If search does not find close enough passages, the model does not generate a substantive answer at all.
- Source links. The answer comes with a link to the knowledge base section or a page on the website. This keeps the model disciplined and helps when reviewing errors.
- Topic limits. Questions unrelated to your product are politely declined.
- No commitments. The bot does not promise discounts, compensation or custom timelines; those questions go to an agent.
- A test set. 100 to 200 real questions with reference answers, including provocative ones. It is rerun after every change to the knowledge base, the instructions or the model version.
Errors cannot be eliminated entirely, but they can be made rare, visible and safe.
Handoff to a human agent
A good assistant knows when to call a person. Typical triggers:
- the customer explicitly asks for an agent;
- negativity, a complaint, a threat to leave;
- money questions: refunds, disputed charges, custom terms;
- low search confidence, or the same question repeated two or three times in a row;
- complex cases the business has flagged in advance as manual.
It is not just the handoff that matters but its quality. The agent receives the full conversation history and a short summary from the model: who the customer is, what they want, what they have already been told. The customer does not have to repeat everything.
Integrations with the CRM and other systems
Without integrations, the assistant is a reference desk. With them, it starts getting things done:
- creates a lead or a deal in the CRM with the conversation text and source;
- recognizes the customer by phone number or account and takes their order history into account;
- reports the status of an order or request;
- books a service into an open slot;
- adds tags for analytics: topic, channel, outcome.
A Telegram bot with GPT has its own specifics: the user is already identified by their account and status notifications are easy to send, but you need to plan how to link the Telegram account to the customer in the CRM, for example through phone number verification. If personal data passes through the bot, account for GDPR and other data protection requirements on storing and transferring it.
Technically, an assistant with integrations is a full web service with its own architecture, access rights and monitoring, and it should be designed as one, with the same end-to-end development approach as any other product.
Quality metrics
"The bot seems to work" is not a metric. What to track from day one:
- Share of conversations resolved without an agent. The main savings indicator, but only paired with quality.
- Answer accuracy. Manual spot checks: for example, every week 50 random conversations are labeled "correct," "partly correct" or "wrong."
- Share of "I don't know" answers. Too high means gaps in the knowledge base. Too low may mean the bot is making things up.
- Handoffs to agents and their reasons. They show which topics to add to the knowledge base.
- Customer rating after the conversation.
- Conversion to an inquiry or purchase for website assistants. Tie it to your web analytics: goals and events in Google Analytics.
The most useful regular habit is reviewing logs: where the bot got it wrong, where the customer left, which questions are missing from the knowledge base. That is a source of improvements for months ahead.
Stages of implementing an AI assistant
- Analysis. We go through request history, identify frequent topics and decide what the bot handles itself and what it passes to people.
- Knowledge base. We collect it, clean out contradictions and assign owners for updates.
- Prototype and test set. We check answer quality on real questions before customers are connected.
- Integrations. CRM, back-office system, the agent channel.
- Closed test. Employees and a small share of customers use the bot, and every conversation is reviewed.
- Launch and growth. We gradually expand the audience, review metrics weekly and grow the knowledge base.
This sequence mirrors how we at Sphere, a digital agency, run every product: research, design, development and testing, launch and growth. Learn more about our approach on the agency page.
FAQ
Can I build a Telegram bot with GPT?
Yes, technically it is a standard task: the bot receives messages through the Telegram Bot API and sends them to a language model via the backend. The main work is not connecting the model but the knowledge base, guardrails, CRM integration and handoff to an agent. Choose the model with an eye on what data passes through the bot.
Which language model should I choose for an assistant?
The one that performs best on your test set of questions at an acceptable cost and with suitable data processing terms. Mid-sized models are enough for many support tasks, and accuracy depends more on the knowledge base and search than on the model.
Can an AI bot fully replace customer support?
Usually not. It handles frequent routine questions well and takes load off the first line, but complex, contentious and money-related questions are better left to people. The right goal is for agents to spend their time where they are really needed.
How do I update what the bot knows?
Through the knowledge base: change a document or section in the admin panel, and after reindexing the bot answers from the new version. Prices, stock and statuses are better not stored in documents but fetched from the back-office system at the moment of the question.
How much does an AI assistant for business cost?
The cost is made up of preparing the knowledge base, developing the backend and integrations, paying for the model per request or running your own servers, and ongoing support. A simple FAQ assistant and a bot that works with the CRM and orders differ in scope several times over, so the estimate is best made after reviewing your customer requests.




