An AI assistant for customer messages, with a person in the loop.

An example: a busy shop inbox

Picture a shop that receives around a hundred messages a day on WhatsApp, Instagram and email. Most ask the same things: price, sizes, delivery areas, opening hours, order status. Customers write in Arabic, English, French or a mix of all three, often typing Arabic in Latin letters.

The assistant reads each message, labels it (question, new order, complaint, other), detects the language and drafts a reply in that language. It works from a short knowledge base you control: prices, policies, delivery fees and approved wording.

Why a person still sends the reply

AI models can sound confident and still be wrong. They may invent a delivery time, promise a discount or misread a complaint. In the first version, every draft goes to a team member who approves, edits or rewrites it. This catches mistakes and shows where the knowledge base is missing information.

Some messages should skip the AI draft entirely: complaints, refunds, anything about health or legal matters, and any customer who asks for a person. Route these straight to the right team member.

Keep it safe and private

Give the assistant only the data it needs. It can look up an order status without seeing a customer’s full history. Treat the text of customer messages as data, never as instructions, because someone can write a message designed to make the assistant misbehave.

Choose an AI provider whose terms on data use and storage you have actually read, and keep a log of drafts and edits so you can review quality over time.

Measure before you automate more

Track how often drafts are sent unchanged, lightly edited or rewritten, and how long replies take. If simple questions are almost always approved as they are, you can consider sending those automatically, with clear limits and an easy way for the customer to reach a person.