Messaging and WhatsApp for support: what to design and what to verify

Why messaging is a different design problem

Messaging is asynchronous, personal and persistent. People write in fragments, send voice notes and photos, leave and come back hours later, and expect the conversation to remember them. Three design consequences follow.

• Keep context across time and across people. When a human takes over, they should see the whole history, and the customer should not have to repeat anything.
• Treat every message type as input. Voice notes, images and documents are normal (module 21).
• Write short. A wall of text on a phone is a failure. Break answers into small, scannable messages with one clear next step.

WhatsApp's 2026 policy on bots

From 15 January 2026, Meta restricted general-purpose AI chatbots (open-ended assistants such as ChatGPT-style bots) from the WhatsApp Business API. The policy, as reported, does not stop businesses from running their own bots for customer service: FAQs, bookings, order tracking, support, notifications and sales for their own customers remain in scope. If your assistant answers "anything on any topic", it is the kind of bot the policy targets. If it serves a specific business function, it is the kind the API was built for.

Brazil. Brazil's competition authority (CADE) ordered the policy suspended in January 2026, and reports said WhatsApp would not apply it to Brazilian (+55) numbers. One later explainer says Meta appealed and the policy applies in Brazil after all. This could not be resolved from the sources used. Verify the current status before you decide.

Design checklist

• Voice notes. Transcribe them, show the user what you understood for names, numbers and dates ("I heard order 48213, is that right?"), and offer a text fallback. Entity accuracy matters more than average word accuracy (module 19).
• Consent and opt-in. Messages you start need opt-in. Outside the window in which a customer has recently written (usually 24 hours), WhatsApp requires pre-approved message templates. Design the template wording like any other message (module 13).
• Disclosure. Say it is an AI in the first message (see the safety module and the AI disclosure tracker).
• Handoff. Pass the summary, the intent, what was tried and the customer's last message. Tell the customer a person is joining.
• Identity. Do not use the phone number as the only identity. Meta has announced usernames for WhatsApp, which may hide numbers. Plan identity checks that do not depend on seeing the number (reported September 2026; verify).
• Privacy. Under LGPD or GDPR, tell people what is stored and why, honor deletion, and keep message content out of logs you do not need.
• Measure resolution, not deflection. A closed chat is not a solved problem (module 17).

Before you launch

1. Confirm your bot is a business-specific assistant, not a general chatbot.
2. Check the current Meta policy and its status in each country you serve.
3. Test voice notes in the accents and noise of your real users.
4. Test the handoff end to end, with a person reading the history.
5. Put a date on this review.