An AI salesperson in Direct turns a comment under a reel into a shipping label
The agent takes each shopper from her first question in Direct to an order in the CRM and a shipping label. Payment, exchanges and tricky cases stay with the owner.
- Client
- Clothing store that sells on Instagram17 products in ≈350 variants; the owner answered every Direct message and comment herself
- Service
- emilus. Agentinbound inquiries
- Status
- In production at the clientthe agent serves the store's shoppers in Direct and in comments
- Systems
- Instagram, CRM, Nova Poshtafrom a comment under a reel to a shipping label
A human in the loop. Only the owner confirms payments; exchanges and tricky cases stay with her too.
The client's situation
The store sells clothing on Instagram, and every sale happens in Direct messages and comments. The owner handled each inquiry herself: price, size, stock, photos, the order, the shipping label with Nova Poshta (Ukraine's largest delivery service), the prepayment check. The first reply took 40–50 minutes, one inquiry in five came in the evening or at night and waited until morning, and there was simply no time left to upsell.
until the first reply in Direct
per client datainquiries arrive after hours and wait until morning
per client dataof people left when the reply took longer than an hour
per client dataupsells: there were none at all
per client dataThe owner's requirementThe agent speaks in her voice and sells by her rules. What the agent says always matches the stock.
What the system does
- Replies in Direct and under posts within a minute, so the shopper gets an answer while she is still scrolling her feed.
- Works nights, weekends and holidays: a message at 23:00 turns into an order right away.
- Recognizes the product from a photo, a reel, a story or a forwarded post and transcribes voice messages, so the shopper never has to explain what exactly she means.
- Takes prices, sizes, colors and stock only from the store's CRM at the moment of the question, picks the size from the size chart and sends photos of every color, so what it says always matches the stock.
- Suggests a fitting add-on before checkout: only from items the owner picked, and once per order, so the order value grows without pushing.
- Places the order and reserves the item after the shopper's clear “yes” and creates the Nova Poshta shipping label itself, so the owner types nothing by hand.
- Checks the prepayment receipt: recipient, date, amount and duplicates, then brings the owner a “Payment received / Not received” card, because only a person confirms a payment.
- Follows up with the shopper the next day if a sale has stalled, and answers “where is my order?” with the tracking number.
- Takes orders from abroad: collects all the details and passes them to the owner to price the delivery, so a buyer from another country is not lost.
- Gives the owner a dashboard: every conversation with each step the agent took, orders by status, tone, terms, answers to common questions, AI costs and a daily evening review. She can take over any conversation and hand it back to the agent.
The architecture in plain terms
The agent takes prices and stock from the store's CRM, and no message gets lost, even during updates. Payment, exchanges and tricky cases are the owner's decision.
- Channels
- Instagram Directcomments
- Integrations
- Instagram, official Meta APIthe store's CRMNova PoshtaTelegram for the ownerspeech recognition
How it works
- shopperQuestionWrites in Direct, comments on a reel or sends a photo of an item
- agentRecognitionIdentifies the product from a photo, post, reel or caption
- CRMStockPrice, sizes and availability at the moment of the question
- agentAdviceSize, colors, photos, delivery and payment, a fitting add-on
- shopperSummary and “yes”The agent shows the summary and waits for a clear yes
- CRMOrderCreated with the item reserved
- Nova PoshtaShipping labelCreated automatically, number added to the order
- ownerA human confirmsA card in Telegram; only the owner confirms payment
How we test the agent
- Replay on real messages from a year of chat archives: the agent sees the same history the owner saw back then, and we compare the replies.
- An exam on sales scenarios and a stress test where the “shopper” deliberately tries to throw the agent off.
- Automated tests inside the agent's system.
- End-to-end check: an outside person went from a comment to an order, and the shipping label was created automatically.
Result and payback
revenue in the first 2 months with the agent
per client dataaverage order value: there was no upselling before at all
per client datato the first reply, around the clock
per client dataProject payback: ≈2 months per client dataFast replies brought back the ≈15% of shoppers who used to leave, and upselling raised the order value. The pilot paid for itself in 2 months.
What is running
- running
A reply within 1 minute, day and night: Direct and comments, the product from a photo, stock and prices from the CRM
- running
Orders with reserved items, Nova Poshta shipping labels, the owner's card in Telegram
- running
Upselling that did not exist before: average order value +20%
A similar task in your business?Solomiia, Pavlo's AI assistant, will ask about the details and prepare your conversation with him.
emilus. Agent
AI agent for customer inquiries. Diagnostic $300 → pilot $2,500–3,500 in 3 weeks → system $6,000–12,000; support $300 or $500 a month; AI usage paid by you directly.