Hong Kong Retail WhatsApp AI: Meta Rules, Handoff

Meta policy, Cantonese accuracy and human handoff design

WhatsApp has very high penetration in Hong Kong, and for retail it is one of the main channels for customer communication. Combining AI customer service with the WhatsApp Business API can in theory extend service coverage without adding customer service headcount, but real deployment involves three dimensions that each need their own assessment: Meta policy restrictions, Cantonese language context handling and human handoff design.

Meta’s Current Pricing Structure

The billing model of the WhatsApp Business API went through a major adjustment between 2024 and 2025. From November 2024, user-initiated service conversations (initiated by the customer within the 24-hour window) became completely free, removing the monthly limit of 1,000 free conversations. From April 2025, template messages sent proactively by businesses moved to per-message billing, replacing the original 24-hour conversation window billing. From July 2025, per-message billing was fully implemented, and Hong Kong rates for utility and authentication messages were raised again on 1 July 2026.

For retailers, this change means the cost of customer-initiated service queries is significantly lower, while the cost of marketing and notification messages that the business pushes proactively must be recalculated, and the old conversation-based billing logic cannot be reused.

The Reality of Cantonese Accuracy

Modern mainstream AI models such as GPT-4o, Claude and Gemini have improved significantly at handling written Cantonese, Traditional Chinese and English mixed input within a single conversation, and performance in real conversation scenarios is generally usable. However, colloquial Hong Kong Cantonese written input (containing many slang terms, abbreviations and Chinese-English mixing) is still a challenge for models.

It is advisable to test a vendor’s demo version with real customer service conversation records before formal procurement, rather than accepting generic accuracy figures. Test cases should include: pure written Cantonese queries, complaints that mix Chinese and English, emotional expressions of dissatisfaction, and queries involving specific product details. These four are the test scenarios that best distinguish model capability in customer service.

Meta’s AI Disclosure Requirements

Meta policy requires businesses to disclose to customers at the start of an automated AI conversation that they are interacting with AI. In January 2026 Meta went further and introduced structured AI rules, strengthening the related disclosure requirements. Deployments that do not disclose as required face platform policy risk. The disclosure is usually designed as a one-line notice at the start of the conversation and does not affect the rest of the interaction.

Standard Trigger Conditions for Human Handoff

A well-designed AI customer service system should include clear human handoff trigger logic, rather than letting the AI keep trying to handle scenarios beyond its capability:

• Complex complaints or refund disputes: conversations with high emotional intensity, where the AI’s standard reply framework struggles to defuse the situation
• The customer explicitly asks to speak to a person
• AI confidence score below the set threshold (the system cannot find a matching answer in the knowledge base)
• Sensitive topics (involving legal disputes or personal safety)

At handoff, the full conversation record should be passed along so the customer does not have to repeat the explanation. The system log should record the reason for each handoff trigger, as the basis for continually refining the AI’s coverage.

For the design of handoff reason logs, see AI Application Log Design: What to Record, What Must Go to Database.

The Compliance Route for Hong Kong Businesses to Connect

Hong Kong businesses cannot apply directly to Meta for the WhatsApp Business API. They must apply for access through a Business Solution Provider (BSP) approved by Meta, and complete Meta’s business verification process. Local BSPs include several technology providers that already have services in Hong Kong.

For the back-end questions to settle before deployment, see Hong Kong WhatsApp AI Integration: Five Back-End Questions to Settle Before Deployment.

Summary

The practical benefit of WhatsApp AI customer service in Hong Kong retail concentrates on handling standard queries (order status, exchange and return policy, store information). These structured queries make up most customer service volume and suit end-to-end handling by AI. Human handoff design for edge scenarios, and correct calculation of Meta’s pricing structure, are the two key dimensions to confirm before deployment.