A large Ukrainian women's-fashion online store runs an AI sales assistant built by Grow2.ai, the AI practice that grew out of Auspex. Since late March 2026 the agent has handled 6,400+ customer dialogues in Instagram, Viber and Telegram — two-thirds of them in Instagram Direct, a third outside business hours — at roughly $0.11 in LLM costs per conversation, with a median reply time of 13 seconds.
Most "AI for e-commerce" articles quote hypothetical percentages. This case study shows production telemetry: how many conversations an AI agent actually handles for a real store, when customers write, what it costs per dialogue — and what the agent deliberately does not do yet.
The Challenge: Consultants Don't Scale to Traffic Peaks
Client: a large Ukrainian women’s-fashion e-commerce store (name withheld by agreement; industry and numbers are real).
The store’s product consultants were the bottleneck. On promo days incoming questions spike into the hundreds per hour — "does this dress come in M?", "where is my order?", "what goes with this skirt?" — and no support team scales that way. Off-peak the opposite problem: a third of all customer messages arrive outside the 9-to-18 working day, when nobody answers at all. A customer ready to buy at 10 p.m. cools down by morning.
The answers existed — in the product catalog and the ERP. What was missing was someone to deliver them at any hour, instantly, without queueing.
The Solution: A Sales Assistant Inside Instagram, Viber and Telegram
Grow2.ai built a conversational AI agent that lives where the store’s customers already are — Instagram Direct, Viber, Telegram and the site chat — and works the way a good shop consultant does:
- Finds products by description, using retrieval over the live catalog — synchronized with inventory, so the agent never recommends what is out of stock.
- Searches by photo. A customer sends a picture — the agent recognizes the garment and finds the same or similar items in the catalog.
- Checks sizes and availability in real time, per store and per warehouse.
- Tracks orders and loyalty status — "where is my parcel?" and "how many bonus points do I have?" are answered from live data.
- Handles Instagram-native activity. A story mention gets a warm thank-you, and blogger or partner inquiries are routed to a separate cooperation channel — kept out of the customer queue.
- Writes everything to the CRM. Dialogues become contacts and deals in Bitrix24, so the sales team sees the full history.
The agent answers in Ukrainian, keeps the brand’s tone of voice, and holds context across the conversation — a customer can go from "show me linen dresses" to "which of them is available in Kharkiv in size S?" without repeating themselves.
The Numbers: Production Telemetry, Not Estimates
Metric | Value |
|---|---|
In production since | late March 2026 |
Customer dialogues (April–July) | 6,400+ |
Instagram share of dialogues (June) | 67% |
Unique customers, last two weeks | 1,256 |
Messages outside 9–18 business hours | 33% |
Messages at night (22:00–08:00) | 10% |
Median reply time | 13 seconds |
LLM cost per dialogue (June) | ~$0.11 |
Two numbers carry the business case. A third of demand arrives when no human is on shift — that demand used to be silently lost. And at ~$0.11 per conversation, the marginal cost of serving one more customer is effectively zero: the agent absorbs promo-day peaks without hiring, training or burnout. Every dialogue runs through a self-hosted observability platform (Langfuse), so these economics are measured, not modeled.
The channel split matters too: two-thirds of June dialogues (67%) came through Instagram Direct — for a fashion store, that is where the audience lives. Viber and Telegram add roughly 30% between them; the site chat covers the rest. An e-commerce agent that skips Instagram misses most of the conversation.
Quality Control: A Reviewer That Can Only Delete
The store’s brand voice is an asset, and the first question every owner asks is "what if the bot says something stupid?" The architecture answers it in three layers:
- Grounded answers. The agent retrieves from the catalog and knowledge base — it does not improvise product facts.
- A compliance reviewer with a one-way permission. A second model reviews every outgoing answer against brand rules and can only remove problematic fragments. It physically cannot add new claims, promises or discounts — a reviewer that can only delete cannot hallucinate.
- Human handoff. When the agent is not confident — or the customer asks for a person — it transfers the dialogue to a manager, with full context attached.
Every conversation is traceable end-to-end: what the customer asked, what the agent retrieved, what the reviewer cut, how long each step took and what it cost.
What the Agent Deliberately Does Not Do (Yet)
As of July 2026 the agent consults, checks and tracks — but the final order is still placed in the ERP by a human manager. The next phase is already in development: the agent will place the order itself and send a payment link in the same chat, closing the loop from "show me dresses" to a paid order. That phase was prioritized by the client after three months of production — which says more than any satisfaction survey.
The agent also does not invent discounts, does not answer "will this fit me?" with fake confidence, and does not pretend to be human.
The pattern — a conversational agent over a live catalog, with photo search, CRM integration and a delete-only reviewer — fits any retail business where customers ask the same questions about products, availability and orders. For the full cost mathematics of a comparable project, see the dealer-network ROI case study and the breakdown of an AI agent’s three protection levels.
Want the same math for your store? Take the 2-minute AI audit or talk to us — we will show where an agent pays back in your process.
Case documented by Andrew Maryasov, founder of Grow2.ai — AI agents for business. Client name withheld by agreement; all numbers are production telemetry (Langfuse), July 2026.