An AI sales assistant that never invents a price
An AI agent answers customers in the brand's own Telegram Business account, in Uzbek and Russian. It quotes real prices and stock, takes orders, and hands the chat to a person when it isn't sure.
- Client
- A fashion brand
- Year
- 2026
- Type
- AI & bots, ERP & CRM
- Role
- Sole developer: design, engineering, deployment
Overview
A clothing brand sells mostly through Telegram, and its managers spent the day answering the same questions about sizes, prices and delivery. I built an AI agent that works inside the brand's own Telegram Business account, and the admin behind it: catalogue, orders, customers, chats, a knowledge base and AI cost tracking.
The problem
Messages arrive day and night, most of them repeat, and slow answers lose sales. But an AI that guesses a price or promises stock that isn't there costs more than it saves.
The challenge
Letting a language model talk to customers without letting it make things up, in two languages, and knowing exactly when to step back.
My role
- Agent design: tools, instructions and hand-over rules
- Admin: catalogue, bundles, promotions, orders and customers
- Message queue, voice and photo handling
- Analytics, AI cost tracking and a chat simulator for testing
- Deployment and logging
Approach
Facts come from tools, not from the model
The agent can only quote what its tools return: product, price, size, stock, delivery rules. When a tool has no answer, the agent says so and calls an operator.
Gaps become knowledge
Questions the agent couldn't answer are collected as gaps. The owner answers each one once in the knowledge base, and the agent uses that answer from then on.
The owner sees everything
Every chat, order and AI cost is in the admin, with a simulator for testing new rules before customers see them.
Key features
- Replies from the brand's own Telegram Business account
- Real prices, sizes and stock from the database
- Orders created in the chat, with printable order slips
- Hand-over to an operator with the full conversation
- Knowledge base that tracks unanswered questions
- Voice messages and photos
- AI usage and cost per conversation
Architecture & technologies
- Next.js
- TypeScript
- Prisma
- PostgreSQL
- BullMQ
- Redis
- OpenAI
- Telegram Business
Design & UX
- Customers write in a normal chat, not a bot menu
- Operators pick up exactly where the AI stopped
- The dashboard answers one question first: how is the shop doing
Outcomes
- About 40% of conversations are handled by the AI alone (30-day dashboard)
- The same system was adapted for a second shop, selling wellness products