Free playbook
← All playbooks
SalesWe run this

Restaurant sales agent on Instagram, Messenger and web chat

Menu, hours, location and price questions answered in seconds. Orders, complaints and payments go to a person.

For
Restaurants and food businesses that get the same questions all day
Setup time
1–2 days
Runs on
Custom codeLocal model

Why we run it

We run this for Ricki Pollo, a wood-fired rotisserie chicken restaurant in Puebla, Mexico. The agent is called Fernanda and answers Instagram DMs, Facebook Messenger and the chat on the restaurant's website.

Steps

  1. Write down the official facts once: address, hours, menu with exact prices, order phone, delivery notes. The agent may only quote these.
  2. Give the agent a name and a voice that fits the brand. Ours is warm, short and in Mexican Spanish, six lines max.
  3. Route anything risky to a human: orders in progress, complaints, refunds, invoices, payments, or anyone asking for a person.
  4. Connect Instagram and Messenger through the Meta Graph API and the website chat through a small endpoint.
  5. Run the model on your own hardware or a private endpoint so customer messages stay with you.
  6. Send the owner an alert for every handoff, with the customer's last message.
The promptIn the kit
You are Fernanda, the customer assistant for [RESTAURANT] on Instagram, Messenger and web chat.
Reply warmly, like a person from the restaurant, max 6 lines. Introduce yourself only once.
Copy-paste kit · $29Full prompt and config, setup checklist, standing guardrails, updates.

Also in The Revenue Desk, The Full Bench.

Guardrails

  • Facts are a fixed list in the prompt. The agent can't make up a promotion.
  • Money and complaints always go to a human.
  • Customer messages are answered by a model running on our own hardware, not a public API.

What broke

  • Pasting the whole menu into every reply buried the answer. The agent now answers the exact question first and keeps replies to six lines.
  • On WhatsApp we kept the agent in observe-only mode. WhatsApp orders are where mistakes cost real money, so the staff still answer there while the agent logs conversations.

Our stack

Python service on a small VPS, Meta Graph API, local Qwen model on an NVIDIA GX10, owner alerts by email.

Want it running on your accounts?We install and adapt it for you, from $500.
Have us install it
Next playbookA team of ten agents with one orchestrator →