Karina is a Tunisian cosmetics and perfume brand whose customers live in their direct messages. Questions about orders, deliveries, products, and complaints arrive all day across Messenger and Instagram, in a mix of Tunisian Derja, French, and Arabic, often with a photo attached. We built an AI assistant that meets them there, sorts what comes in, answers what it can, and pulls in a human the moment a conversation needs one.

The challenge

The volume was real and the messages were messy. A customer might switch languages mid-sentence, send a picture instead of a description, or arrive with a complaint that no bot should try to resolve on its own. A generic chatbot that answered everything confidently would have done more damage than good. The brand needed something that understood the message first, and knew its own limits.

What we built

An assistant designed around triage rather than canned answers, deployed across both channels.

• Intent classification on every incoming message, so an order question, a delivery question, a complaint, and small talk each get the right treatment.
• Multilingual understanding and replies in Derja, French, and Arabic, matching how customers actually write.
• Image understanding, so a photo of a product or a delivery becomes something the assistant can engage with rather than ignore.
• Order and complaint capture, collecting the details the team needs and routing them to the right place.
• A robust human handoff: the assistant pauses, notifies the team with the full conversation and the reason, and hands control back with a single click when the person is done.

The most valuable thing the assistant does is recognize the moment it should stop and bring in a person.

How we worked

The system runs on OpenAI models, with a fast model handling intent on every message and a stronger one reading images. It is aware of business hours and sets honest expectations outside them, and it waits a beat before replying so a burst of quick messages becomes one coherent answer instead of three fragments. Anything sensitive, a complaint, a voice note, a frustrated customer, is an automatic signal to escalate rather than improvise.

We were deliberate about what the assistant does not claim to do. It triages, answers the common questions, captures what the team needs, and escalates the rest with context intact, rather than inventing answers it cannot stand behind.

The outcome

Karina's customers get fast, multilingual replies in the channels they already use, and the team gets clean, context-rich handoffs instead of a wall of unread messages. The routine load is handled automatically, and the conversations that need a human reach one without the customer ever feeling the seam.