Customers now expect an instant answer at any hour, on your website, in WhatsApp and inside your app. AI assistants can deliver that, and the good ones resolve routine questions without a queue. The poor ones invent policies, loop endlessly and make it impossible to reach a person.
The difference is rarely the model. It is the preparation around it.
A real warning: your chatbot speaks for you
In February 2024, a Canadian tribunal ruled in Moffatt v. Air Canada that the airline was responsible for incorrect refund information its website chatbot had given a customer, and ordered it to pay damages. The airline’s argument that the chatbot was responsible for its own statements did not succeed. Whatever your jurisdiction, assume customers and courts will treat your assistant’s answers as your answers.
The pre-launch checklist
- Scope: list the topics the assistant may handle, such as order status, delivery times or account help, and the ones it must hand over, such as complaints, refunds outside policy or legal and medical questions.
- Grounding: answer only from approved sources (help articles, policies, product data) and say clearly when the answer is not available.
- Live data: connect to order and account systems through narrow, read-only lookups, with the customer verified before anything personal is shown.
- Handover: one obvious way to reach a person, which passes the conversation history along so the customer never repeats themselves.
- Tone and limits: a short, plain style, no made-up promises, and no discounts or exceptions unless a rule allows them.
- Privacy: tell users they are talking to an AI, minimise what you store and set a retention period.
- Testing: run a set of real past questions, including hostile and confusing ones, and review every answer before launch.
WhatsApp needs its own planning
If WhatsApp is a channel, build on the official WhatsApp Business Platform. It has rules that shape the design: businesses can reply freely within a 24-hour window after the customer’s last message, but messages sent outside that window must use templates approved in advance, and customers must have opted in to hear from you. Plan reminders and follow-ups around those rules from the start.
Measure resolution, not deflection
Deflection, the share of conversations that never reached an agent, rewards an assistant that is hard to escape. Measure instead whether the customer’s issue was resolved, how many came back with the same question within a few days, and satisfaction for conversations the AI closed versus those a person closed. Review a sample of transcripts every week; they are the fastest guide to what to fix next.
