An AI agent for customer service: what it does well, what it must never do

What an AI agent does well

• Answering instantly, at any hour, the questions whose answer is in your data: where is my order, what is the price, do you have stock, when do you open.
• Understanding the mix of languages the way your customers write it: Darija in Latin script, Arabic, French, in the same message.
• Carrying out simple, verifiable actions: changing a delivery address, sending an invoice, booking an appointment, chasing a late payment.
• Summarising the conversation and handing it to a human with the full history, when it does not know or when a decision is needed.
• Measuring: volume, recurring topics, share resolved without a human, satisfaction. You find out what your customers really ask.

What it must never do

• Invent. An agent that cannot find the information in your systems must say so and hand over, never guess. That is rule number one, and it is a setting.
• Decide alone on a refund, a goodwill gesture, an exception to the contract. It proposes, a human approves.
• Pass itself off as a person. The agent introduces itself as an automated assistant from the first message. It is an obligation, and in practice it does not stand in the way of satisfaction.
• Handle a complaint or an angry customer. It detects the tone, apologises, and steps aside. Insisting is the surest way to lose the customer.
• Solicit without consent. The WhatsApp Business terms and law 09-08 forbid it, and Meta blocks the numbers that do it.

How it is set up

A week on average. We connect WhatsApp, the website and, if needed, the phone system. We import your standard answers and plug the agent into your orders, invoices and appointments. We define together what it may say and do, and what it hands over. Then we test with your team, internally, before opening it to customers.

The three settings that make the difference

1. 1Scope: list the twenty most frequent questions and check that the agent answers them from your data, not from general knowledge.
2. 2Handover rules: keywords (“refund”, “lawyer”, “complaint”), detected tone, missing information, an explicit request for a human.
3. 3Review: for the first two weeks, someone reads every conversation and refines the instructions. After that, a sample each week is enough.

An example

An online shop in Casablanca receives 300 WhatsApp messages a day, half of them asking where the order is. The agent answers those in seconds, changes addresses, and the team only handles real complaints. Time to first reply went from three hours to a few seconds, and the two-person team did not grow.