Skip to content
Sadia Khanam PrityAutomation Developer
E-commerce / RetailRetail Facebook Page

Rolling out an AI Messenger agent for a retail page

Replacing round-the-clock manual replies with an automated agent that answers, orders, books and escalates on its own.

n8nMeta Graph APIGoogle GeminiGoogle SheetsGoogle CalendarOCR.space

Seconds instead of hours outside business hours

Response time

Reserved for conversations that need judgment

Human escalations

Matched automatically via code lookup or vision model

Photo enquiries

Challenge

Customer messages came in at all hours through Messenger — stock questions, order requests, booking requests, and photos of product tags. Outside business hours, everything queued for the next available reply, and even during business hours a large share of the volume was questions the catalogue already answered.

Outcome

An orchestrated automation now handles the majority of the conversation on its own: answering from live catalogue data, writing orders and bookings straight into the sheet the business already runs on, and booking calendar slots directly. It escalates to a human only when a conversation genuinely needs one, logging the handoff so nothing is dropped silently.

The situation

The page ran on manual replies. A member of staff was, in effect, a full-time router: deciding whether an incoming message was a stock question, a price question, an order, a booking request, or a photo that needed matching against the catalogue — and then doing the matching by hand.

What changed

Two workflows now sit in front of that inbox. Text and photo messages are separated immediately. Text goes to an AI agent wired up with tools for the catalogue, order-writing, booking-writing, calendar creation, and escalation logging — plus per-customer session memory so a conversation doesn't lose context between messages. Photos run through a cheap OCR pass first, and only escalate to a vision model call when there's no printed code to read directly.

Why it holds up

Every AI decision point has a defined fallback to a human, logged rather than silent. The system is designed so that when it doesn't know, it says so — to the customer and to the business — instead of guessing.