AI agents for business: what they do, cost and get wrong
How an AI assistant on your site and Telegram works: knowledge base, tools, handover to a human. Real development cost and monthly model expenses.
Published — 7 min read
An AI agent is not a button bot. It understands the question, answers in its own words from your knowledge base and can act on its own: check an order status, book a slot, create a CRM lead. We deploy such agents on websites, Telegram and WhatsApp.
What an agent does well
- Answers routine questions 24/7: prices, timelines, addresses, delivery and returns.
- Qualifies leads: clarifies the task, budget and deadline, then hands a ready card to a manager.
- Books free slots and sends reminders.
- Searches the catalog: “a gift for a 5-year-old girl under $40”.
- Reads incoming documents and fills in spreadsheets.
How it is built
- Knowledge base: your prices, services, rules and FAQs, collected from the site, documents and chats.
- Retrieval (RAG): for each question the agent finds the relevant pieces instead of memorizing everything.
- Tools: functions the agent can call — “find order”, “book”, “create lead”.
- Guardrails: what it must not discuss, when to call a human, which prices it may quote.
- Human handover: tricky chats go to an operator in Telegram and the reply returns to the same chat.
What it costs
| Item | Price |
|---|---|
| Agent for a site or Telegram | from $300, 3–5 days |
| CRM or booking integration | from $100 |
| Model usage for a small business | $5–30 a month |
| Support and knowledge updates | from $30 a month |
We cut model costs with caching and by matching the model to the task: system instructions are not paid for again on every question, and simple answers come from a cheaper model. In one of our projects switching the default model halved the bill without hurting answer quality.
Where agents fail
Agents fail where they lack data: if a price is not in the base, they may invent one. So prices and stock come only from tools, never from “memory”, and questions outside the base get an honest “let me check with a manager”.
Questions and answers
Which model do you use?
Claude, GPT or YandexGPT, depending on budget, customer language and data residency needs.
Does it work in Georgian?
Yes. Modern models understand Georgian well, but the knowledge base should be in the language your customers write in.