AI Chatbots on WhatsApp: Risks, Limits and Rules to Set First

Businesses connect through the WhatsApp Business Platform, also called the Cloud API, either directly with Meta or through a business solution provider such as Twilio, 360dialog, Infobip or a CRM like HubSpot or Zendesk. Incoming messages arrive at a webhook. A bot, rule-based or AI-powered, decides on a reply and sends it back through the API.

Meta’s rules shape the conversation. Businesses can reply freely within a 24-hour window after the customer’s last message. Outside that window, they can only send pre-approved template messages, which are charged by category such as marketing, utility or authentication. Meta has also changed its policies on general-purpose AI assistants on the platform, so check the current WhatsApp Business policy before designing a bot whose main job is open-ended chat rather than serving your own customers.

• Hallucinated answers: the model invents a return policy, price or delivery date.
• Prompt manipulation: users try to make the bot say embarrassing or off-topic things.
• Privacy: customers share order numbers, addresses, health or financial details in chat.
• Policy violations: marketing messages without opt-in, or restricted products.
• Dead ends: no clear route to a human when the bot cannot help.
• Number quality: if many users block or report the number, Meta can limit messaging.

Each risk has a practical control. None of them is solved by choosing a better model alone.

1. Scope: what the bot may and may not answer

Write a short list of topics the bot handles, such as order status, opening hours, product availability and appointment booking. Everything else goes to a human or a fixed reply. A narrow scope is the single most effective protection against wrong answers.

2. Grounded answers only

The model should answer from approved sources: the FAQ, the product catalogue, the order system. This is usually done with retrieval, where the bot looks up relevant text before replying. If the source does not contain the answer, the bot says it does not know and offers a person.

3. Human handover

Define triggers for handover: the customer asks for a person, the bot is uncertain, the topic is a complaint, or the conversation includes words linked to legal or safety issues. Tell the customer when a person will reply, and staff that queue during the hours you promise.

4. Data handling

Under GDPR in Europe and KVKK in Turkey, chat content containing personal data needs a lawful basis, a privacy notice and a retention period. Do not send more data to the AI provider than the task needs, and check where that provider processes data.

5. Opt-in and messaging limits

Marketing templates need clear prior consent. Keep a record of how and when each contact opted in. Unwanted messages lead to blocks, which affect the number’s quality rating.

Testing should be adversarial, not polite. Ask the bot questions it should refuse, give it incomplete order numbers, write in slang and in a second language, and try to make it promise discounts. Log every reply.

• Prepare 100 real customer questions from past chats and check each answer.
• Add 30 edge cases: complaints, angry messages, off-topic requests, manipulation attempts.
• Check how the bot behaves outside the 24-hour window.
• Confirm every handover reaches a person and the customer is told what happens next.
• Re-run the test set after every prompt or knowledge base change.

Bots handle routine questions well and complicated complaints badly. A refund dispute, a damaged product or a delayed wedding order needs a person. Language quality also varies: models are strong in English and good in Turkish, but local dialects and mixed-language messages still cause errors. Voice notes are common on WhatsApp and need transcription before the bot can respond, which adds another point of failure.

A chatbot is not a project that ends at go-live. Products change, policies change and customers find new ways to ask the same question. Without a named owner, the knowledge base drifts out of date within weeks and the bot starts giving answers that were correct last season.

Give one person the job of reviewing a sample of conversations every week, updating approved sources and re-running the test set after each change. Pair that person with someone from support, because support staff hear the complaints the bot causes before anyone else does.