WhatsApp AI Agent for Customer Service: Design and Meta Rules

Which guardrails should I set?

Treat the agent as a new employee with limited authority. Guardrails belong in the system around the model, not only in the prompt.

Disclose clearly that the customer is talking to an automated assistant. This is good practice for trust, and some laws and sectors may require it, so check yours. Respect opt-in and opt-out too: Meta's Business Messaging Policy requires consent, and an AI agent that replies to a customer who asked to stop is a policy and trust problem.

How should human handover work?

Build handover as a first-class path, not an escape hatch. A bad handover forces the customer to repeat everything.

Triggers

• The customer asks for a person, in any wording.
• The agent fails to resolve the issue after a set number of turns.
• Sentiment turns negative, or the customer repeats themselves.
• The topic is on a sensitive list: complaints, payments, health, legal.
• A tool call fails or the agent lacks permission.

The handover itself

1. Tell the customer a person will continue and what to expect.
2. Pass a short summary, the customer's details and the actions already taken to the agent's inbox.
3. Stop the AI from replying in that conversation until a human releases it.
4. Show the agent the full history in the shared inbox.
5. Return the conversation to the AI only if the customer's next question is routine.

Our guide to chatbot human handoff goes deeper, and the customer support workflow shows how queues and SLAs fit around it.

How do I handle Arabic and English in one agent?

Support both languages natively, not through translation of an English-only design. In the Gulf, customers switch mid-conversation, write Arabic in Latin letters, mix English words into Arabic sentences and use either digit style.

• Reply in the customer's language, and switch when they switch.
• Test with real messages, including dialect and typos, not only Modern Standard Arabic.
• Write Arabic templates natively. See Arabic customer messaging best practices.
• Check right-to-left layout for names, order numbers and links inside Arabic sentences.
• Keep policies identical across languages, so the Arabic answer does not differ from the English one.

OmniChannels.ai AI agents support English and Arabic language understanding.

What does the 24-hour window mean for an AI agent?

The agent's free-form replies must fall inside the customer service window. When a customer messages you, a 24-hour window opens. Outside it, you can send only an approved template, and Meta charges per delivered message by category, with changes to service and utility messages from 1 October 2026 that you should confirm in Meta's current rate card.

Design accordingly:

• Answer inside the window; ask the customer to reply to keep it open if a task spans days.
• Use approved templates for reminders and follow-ups, not AI-generated text.
• Cap the length of conversations so you do not pay for aimless back-and-forth.

Read our guides to the 24-hour window and template categories.

Worked example: a clinic group's booking assistant

"Noor Clinics" (synthetic), a multi-branch clinic group in the UAE, adds an AI agent for appointment questions.

• Scope: opening hours, branch locations, booking and rescheduling, preparation instructions. No diagnosis, no medical advice.
• Knowledge: approved branch information and preparation sheets, reviewed monthly by an operations manager.
• Tools: read availability, create and change bookings. Cancellations inside 24 hours need staff approval.
• Guardrails: any symptom description or complaint triggers immediate handover with a summary.
• Languages: Arabic and English, tested on real messages.
• Rollout: two weeks on one branch, with staff reviewing a sample of conversations daily, then expansion.

Healthcare has extra obligations; see our patient messaging guide.

How do I measure and improve it?

Review conversations weekly and change one thing at a time.

• Resolution without a human, counted only when the customer did not return on the same issue.
• Handover rate and reason, to find gaps in knowledge or tools.
• Customer feedback after resolved conversations.
• Wrong-answer rate, from a sampled human review.
• Cost per resolved conversation, including message charges.

See support KPIs for definitions.

What are the common mistakes?

• Launching "answer anything". Start narrow.
• Relying on the prompt for safety instead of limiting tools and data.
• No handover, or handover without context.
• Ignoring consent, opt-out and the customer service window.
• Assuming the policy is static. Meta updates its terms; recheck them.
• Skipping Arabic testing with real dialect.
• Limits: no AI agent resolves every conversation, and results depend on your knowledge quality and scope. We make no promise on resolution rates.

How OmniChannels.ai helps

BroadNet is an official Meta Business Solution Provider, and OmniChannels.ai AI agents, delivered through BroadNet AI, understand English and Arabic, hand over to a human agent, keep persistent memory, use tools through MCP and can be deployed as SaaS, private cloud or on-premise, on WhatsApp, web chat, voice and other channels. Conversations land in the shared inbox with the full history. Our team prepares WhatsApp onboarding with you; approvals stay with Meta. See AI agents, chatbots or book a demo.