How an AI sales assistant actually works

Online stores receive a steady stream of messages: "Is it in stock?", "How much?", "When will it arrive?" Many are repetitive, yet leaving them unanswered loses the customer. An AI sales assistant is built for exactly that point.

This article walks through the parts of such an assistant, based on building Pasokhinoo.

Assistant versus old-style bot

Older bots were rule-based: if the user types "price", send this message. They stop working at the first question outside the script.

An assistant built on a language model understands meaning. "Would this handle graphic design work?" contains no keyword, but the intent is clear. I compare the approaches in AI agent versus chatbot.

Part one: the store's real knowledge

A language model knows nothing about your products, and when pushed it guesses. Guessing about price and stock is a liability.

So before answering, the assistant retrieves the relevant facts from that store's own knowledge base: products, prices, stock, shipping terms and FAQ. The answer is composed from those facts, not from the model's general memory.

Part two: intent

Not every message has the same value. Someone asking about opening hours and someone asking "if I order today, when does it arrive?" are at different points in the buying journey.

A good assistant detects the intent and acts on it: it answers informational questions, records a buying signal, and guides a ready customer towards completing the order.

Part three: every channel

Customers aren't in one messenger. Pasokhinoo answers on Telegram, Bale, Eitaa, Rubika, Instagram, WhatsApp and a website widget at the same time.

The architectural point is that the assistant's logic must not depend on any channel. Each channel is an adapter that delivers messages to the core in one common shape. Adding a channel means writing an adapter, not changing the brain.

Part four: checkout inside the chat

Every time you send a customer from the conversation to another page, you lose some of them. The purchase should finish where it started: choosing the product, collecting shipping details and recording the payment receipt.

Part five: human handoff

No assistant should pretend to know everything. When confidence is low, the customer is unhappy or the topic is sensitive, the conversation goes to a human operator — with a summary, so the customer doesn't have to repeat themselves. See human handoff.

Part six: brand tone

The assistant represents the store. A sports shop and a law office shouldn't sound the same. Tone has to be configurable and consistent across channels. More in persona and brand tone.

The engineering you don't see

• Isolating each store's data. In a SaaS product, no customer's data may leak to another. That is multi-tenant architecture.
• Cost control. Every message costs money. Without managing model cost, growth means growing losses.
• Measuring quality. You need to know whether the assistant answers well, which takes systematic evaluation.

The takeaway

An AI sales assistant is not a language model wired to a messenger. It is accurate knowledge, intent detection, channel adapters, a purchase path and a clear rule for handing work to a person.

If you want AI doing real work at the core of your product, see AI products.