Key points in 30 seconds
- A WhatsApp AI agent runs on the WhatsApp Business Platform, not the free app.
- It answers from a knowledge base you control.
- Handover to a human must be simple and fast.
- Measure resolution rate and satisfaction from the first month.
What an AI agent can handle
- Opening hours, addresses, lead times, documents required.
- Order or file tracking, via a connection to your system.
- Qualifying a sales enquiry before passing it to a salesperson.
- Booking appointments.
It should not handle a serious complaint, a negotiation or a binding decision on its own.
Six steps to set it up
- List frequent questions from the last three months of conversations.
- Write the knowledge base: approved answers, tone, limits.
- Get access to the WhatsApp Business Platform, directly from Meta or via an approved provider, with a dedicated number.
- Connect the agent to the knowledge base and, if useful, to the CRM.
- Define the handover: keywords, hours, a queue for the team.
- Test with real customers for two weeks before full launch.
WhatsApp rules to know
Business-initiated messages use templates approved by Meta. Within 24 hours of a customer's message, the business can reply freely. Collect customers' consent before writing to them, and comply with personal data legislation.
Indicators to track
| Indicator | What it measures |
|---|---|
| Resolution rate | Requests handled without human intervention |
| First response time | Perceived responsiveness |
| Handover rate | Limits of the knowledge base |
| Satisfaction | Short question at the end of the conversation |
Writing an effective knowledge base
The quality of the agent's answers depends directly on the knowledge base. A few rules:
- One entry per topic: opening hours, prices, lead times, returns procedure, documents required.
- Short, dated answers: the agent should be able to reply in two or three sentences, and each entry shows its last update date.
- Written limits: what the agent must never promise (discounts, exceptional deadlines, diagnoses).
- Brand tone: formal or informal address, greetings, reply language.
- An owner: one person approves every change.
Handling several languages
In the MENA region, the same client may write in French, Modern Standard Arabic, dialect or English, sometimes in one message. Current models understand these mixes well, but the knowledge base must exist in the reply languages you want, or at least be checked after translation. Set a simple rule: the agent replies in the language of the client's last message, and a person takes over if the request remains unclear.
Example conversation
Client: "Hello, my order 4521 hasn't arrived."
Agent: "Hello, let me check right away. Your order 4521 was shipped on 2 November and will be delivered tomorrow before 6pm. Would you like the tracking link?"
Client: "Yes, and I'd like to change the address."
Agent: "An address change needs one of our advisers. I'm connecting you with our team, who will reply within 30 minutes during opening hours."
The agent handles simple requests on its own and hands over, with context, anything that needs a decision.
The cost of a project
| Item | What it covers |
|---|---|
| WhatsApp Business Platform access | Meta pricing per conversation or per message under current rules, possibly through an approved provider |
| AI model | Usage-based cost, depending on message volume |
| Set-up | Knowledge base, integrations, testing, team training |
| Ongoing work | Updating answers and monthly conversation review |
For an SME, the monthly running cost is usually low compared with the handling time saved; check Meta's current pricing for your country.
Common mistakes
- Launching the agent without an approved knowledge base: it makes up answers.
- Making handover to a person difficult: clients get frustrated and give up.
- Using the free WhatsApp Business app for an automated agent: it is not designed for that.
- Sending promotional messages without consent: the number can be blocked.
- Never reviewing conversations: mistakes repeat without being fixed.
Improving the agent every month
Each month, review a sample of conversations, starting with those handed to a person and those rated negatively. For each one, find the cause: information missing from the base, ambiguous wording, request out of scope. Update the base accordingly. After three months, the resolution rate usually rises clearly, because the base covers clients' real questions better and better.
Connecting the agent to the CRM
Connected to the CRM, the agent identifies the client by phone number, finds their orders or files and logs each conversation in their record. Salespeople see requests received on WhatsApp, and sales enquiries qualified by the agent go straight into the pipeline. The Elafy CRM brings email and WhatsApp conversations together in the client record. Our AI and digital transformation support covers designing, connecting and monitoring the agent.
Frequently asked questions
Can you run an AI chatbot on WhatsApp?
Yes, via Meta's WhatsApp Business Platform, directly or through an approved provider. The free WhatsApp Business app only allows simple automatic replies.
Can the AI agent answer in Arabic and French?
Yes. Current AI models understand and write Arabic and French. Quality depends on the knowledge base provided in each language.
How do you stop the agent giving a wrong answer?
By limiting it to an approved knowledge base, forbidding it to invent, providing a human handover and regularly reviewing conversations.
How long does it take to launch an AI agent on WhatsApp?
Usually four to six weeks, including two weeks of testing with real clients before full launch.
Can the agent reply in Arabic dialect?
Current models understand dialects and mixed languages well. Test the answers with your clients and have your team approve the tone.