The challenge
A B2B supplier and authorized service center for trade-automation equipment runs on an inbound-sales model. Its most time-consuming daily routine is invoicing: creating a customer order in the ERP. Right behind it sits picking the right equipment for each customer, which relies on the tacit knowledge of experienced staff.
The pain points were concrete:
- Invoicing is slow and repetitive. Each invoice takes 10 to 15 minutes, and every manager does it every day: find the company card, copy the tax ID, locate items in the catalogue, set prices, generate the printable form
- Hiring and training is hard. Selection knowledge lives with senior staff. New managers ask questions in the wrong order or skip steps, and clients notice
- Selection logic is explainable but fuzzy. It rests on heuristics and experience, and it isn’t captured anywhere reusable
- Partners wait on a human for a simple quote. Resellers who already know the model just need price and availability, but still queue for a manager
- Regulatory changes trigger waves of near-identical requests, and the selection and invoicing work for each one is done by hand
The approach
The engagement is staged so value lands early.
Stage 1: invoice automation. The agent takes two clean inputs, a manager’s free-text description in Telegram or a website order with ready item codes and prices. It identifies the items in the ERP catalogue via fuzzy search, sets correct prices, pulls or requests the customer’s details, creates the customer order, and produces the printable invoice as a PDF. A manager reviews and sends. A human stays in the loop on every order.
Stage 2: equipment recommendations. A knowledge base of per-category selection guidelines drives an adaptive dialog. The assistant decides which clarifying questions to ask, checks recommendations against the live catalogue, and applies business priorities such as stock availability and margin before handing a finished specification to stage 1.
Stage 3: self-service partner bot. A Telegram and website bot that returns price and availability for a known model and sends a ready quote on its own. No manager in the loop: partners get an answer in minutes, at any hour.
Architecture
- AI agent: built on a self-hosted open-source agent framework that works with any LLM. Behavior is constrained: it never commits anything to the ERP without manager confirmation
- ERP integration: the company runs its sales on a trade-management ERP. The agent creates customer-order documents and pulls the printable invoice form through the ERP’s HTTP API
- Messaging: Telegram Bot API as the primary interface for both managers (invoice and recommendation modes) and partners, with email and website-widget channels planned
- Knowledge base: formalized per-category selection heuristics, built as a fuzzy-matching aid rather than a rigid decision tree
- Data: ERP catalogue, stock, and counterparties; daily partner stock and price feeds; the website as the price source of truth
Results
The design removes the most repetitive daily work first (invoicing), then layers on equipment recommendations and partner self-service. Order documents in the ERP keep a manager in the loop; partner quotes need no human at all.