The fastest way to embarrass a brand is to ship a support bot that confidently answers questions it does not understand. We have all met one. You ask something specific, it returns a wall of generic text, and you spend the next two minutes hunting for the word "agent."

When we built a customer assistant for a cosmetics retailer, we started from a different premise. The bot's first job is not to answer. It is to understand what kind of message just arrived, and route it correctly. Get the triage right and everything downstream gets easier. Get it wrong and no amount of clever answering will save you.

Sort before you solve

Every incoming message gets classified first. Is this an order question, a delivery question, a complaint, a product query, or just small talk. We use a small, fast model for this single decision, because classification is cheap and you run it on every message.

Once you know the category, the response strategy is obvious. A delivery question has a real answer. A complaint does not get an answer at all, it gets captured and handed to a human. Small talk gets a short, warm reply and nothing more. The model is not trying to be everything. It is trying to be a good receptionist.

A support bot's first job is to know what kind of conversation it is in.

Speak the customer's language, literally

Our users write in Tunisian Derja, French, and Arabic, often switching mid-sentence. A bot that only handles formal French would be useless to most of them.

So the assistant understands customers writing in any of those languages, even when they switch mid-sentence. This is not a nice-to-have in a market like this, it is the difference between a tool people use and one they abandon. If your audience code-switches, your bot has to keep up.

Understand more than text

People do not just type. They send photos of a product they want, a screenshot of an order, a picture of something that arrived damaged. A text-only bot is blind to half of what customers actually send.

Ours reads images. An incoming photo gets described and folded into the conversation, so "is this the right shade" becomes a question the assistant can actually engage with. Voice messages are different. Rather than risk misinterpreting one, the bot treats a voice note as an automatic signal to bring in a human. Knowing what you cannot reliably handle is as important as knowing what you can.

Respect the clock and the queue

The assistant is aware of business hours. Outside them, it does not pretend a person is around. It sets expectations honestly instead of promising a follow-up nobody will send until morning. It also waits a few seconds before responding, so a customer firing off three quick messages gets one coherent reply instead of three disjointed ones. Small touches, but they are the difference between a bot that feels considered and one that feels like a script.

Be honest about the boundary

The temptation with a support bot is to claim it handles everything. It should not, and the good ones do not. Ours is excellent at sorting, answering the common questions, capturing the details a human will need, and escalating the rest with full context. It does not invent prices it cannot verify or promise outcomes it cannot guarantee.

That honesty is the whole design. A bot that triages well and escalates gracefully earns trust. A bot that pretends to be a full support team loses it the first time someone asks a real question. Start with triage. The rest follows.