An AI assistant for customer messages: how to answer on WhatsApp and your website without losing the human touch

A guest wants to know whether your riad has a free room next weekend. A customer asks if your shop delivers to Agadir. A patient wants an appointment on Thursday. These messages arrive on WhatsApp, on Instagram, by email and through your website, often in the evening or on a holiday, and the person writing expects an answer within minutes. If nobody replies, they message your competitor.

An AI assistant can help you answer faster without hiring a night shift. Done well, it handles the repetitive questions instantly and passes the difficult ones to your team with the full context. Done badly, it frustrates customers and damages trust. This guide explains how to do it well.

What an AI assistant actually does

A modern assistant understands a message written in everyday language, including spelling mistakes and mixed languages, finds the relevant information in the material you gave it and writes a reply. It can also take simple actions, such as checking availability, creating a booking request or sending a link. Behind the scenes it is a language model connected to your knowledge and, where appropriate, to your systems.

It is not a person and it should not pretend to be one. The best assistants say clearly that they are automated, answer what they can and make it easy to reach a human.

Which questions it handles well

• Opening hours, address, directions and parking
• Prices of standard products and services
• Availability and booking requests, when connected to your calendar or booking tool
• Delivery areas, delivery times and payment methods
• Order status, if it can read your order system
• Frequently asked questions about policies such as cancellation, returns and deposits

These questions make up a large share of the messages most small businesses receive, and the answers rarely change. That is exactly the kind of work an assistant does reliably.

Which questions should go to a person

• Complaints and upset customers
• Special requests, exceptions and negotiations
• Anything involving health, legal or financial advice
• Large orders or high-value bookings
• Messages the assistant does not understand or is not sure about

Languages in Morocco and beyond

Many businesses in Morocco serve customers in French, Arabic and English, and customers often switch between them in a single message. A good assistant replies in the language the customer used. Test it with real examples from your own inbox, including Arabic written in Latin letters and messages that mix French and Arabic. Language models handle many of these well but not perfectly, so test before you launch and keep reviewing after.

Where it lives: WhatsApp, website or both

WhatsApp is the main channel for many Moroccan businesses, and it can be automated through the WhatsApp Business Platform with an approved provider. There are rules to respect. Customers should have chosen to contact you, and when you want to start a conversation outside the normal reply window you must use pre-approved message templates. A provider or developer familiar with these rules will save you from having your number restricted.

A chat widget on your website is simpler to set up and works for visitors who are not yet on WhatsApp. Many businesses start with one channel, learn from it and add the other later. Whatever you choose, keep one shared record of conversations so your team sees everything in one place.

Teach it with your own information

An assistant is only as good as the information it works from. The usual approach is to give it a knowledge base made of your real documents: price lists, policies, opening hours, product descriptions, past answers you are proud of. When a customer asks a question, the system retrieves the relevant passages and the model writes its answer from them. This is often called retrieval-augmented generation.

This matters because it keeps the assistant anchored to facts you control. It should answer from your material, say so when it does not know and never invent a price, a policy or an availability. Instruct it explicitly to prefer admitting uncertainty over guessing.

Risks and how to reduce them

• Wrong answers. Restrict the assistant to your knowledge base, show it examples of good replies and review a sample of conversations every week.
• Pretending to be human. Tell customers they are chatting with an automated assistant and offer a way to reach a person.
• Privacy. Customer messages contain personal data. Choose providers carefully, collect only what you need, avoid asking for sensitive details in chat and tell customers how their data is used.
• Tone. A cold, robotic style hurts the brand. Give the assistant a voice that matches your business and check that it stays polite under pressure.
• Abuse. Some people will try to trick the assistant into saying odd things or revealing information. Limit what it knows and what it can do, and never put secrets in its instructions.

A practical rollout plan

1. Collect your last few hundred messages. Group them by topic and count them. You will probably find that a handful of topics cover most of the volume.
2. Write the answers once. Turn those topics into clear, accurate material for the knowledge base, in each language you serve.
3. Start in suggestion mode. The assistant drafts a reply and your team approves it with one click. You get speed immediately and learn where it struggles.
4. Move the safest topics to automatic replies. Opening hours and directions can go first. Keep sensitive topics with a person.
5. Set clear handover rules. Define when the assistant must stop, who receives the case and how fast they are expected to respond.
6. Review every week. Read a sample of conversations, fix gaps in the knowledge base and adjust the instructions.

How to measure success

Look at the numbers that reflect real value. Average response time, the share of conversations resolved without a person, the number of bookings or orders that began in chat, and customer satisfaction after the chat all matter. Also track the cases where the assistant handed over, because they show you what to improve.

Do not judge success by how many messages the assistant sends. A good assistant sometimes ends a conversation quickly because it solved the problem, and sometimes passes it on immediately because it should.

Questions to ask a provider

• Which languages has the assistant been tested in, and how?
• How does handover to a person work, and what does my team see?
• Where is customer data processed and stored, and for how long?
• Can I see and edit the knowledge base myself?
• What do conversations cost at my expected volume?
• How do I switch the assistant off instantly if something goes wrong?

The bottom line

An AI assistant will not replace the warmth of your team, and it should not try to. Its job is to give every customer a fast, accurate first answer at any hour and to free your people to spend their time on the conversations that really need them. Start with one channel and your most common questions, keep a person in the loop and improve every week.

If you are thinking about answering customers faster with an assistant on WhatsApp or your website, tell us how many messages you get and what they are about. We will help you decide what to automate first.