An AI assistant that answers WhatsApp at 11 pm sounds like free staff until it tells a customer their refund is approved or quotes last year's price. WappBlaster's chatbot builder lets you mix AI replies with fixed, rule-based steps and a handover to a person, so you decide exactly which parts of a conversation the model is trusted with. The guardrails below are about drawing that line before customers find it for you.
First guardrail for AI WhatsApp replies: questions versus commitments
Sort every type of message your business receives into two piles. Questions ask for information: "Do you deliver to Navi Mumbai?", "What does the gold plan include?", "Is parking available at your Indiranagar clinic?". Commitments create an obligation: a price for this order, a confirmed appointment, a promised delivery date, a refund decision.
AI can handle the first pile. The second pile belongs to deterministic steps where you wrote the wording and the data comes from your system, such as an appointment flow that books a real slot, or a person in the inbox. A model that "helpfully" confirms a booking it cannot see is the most common way AI replies create a complaint.
Grounding AI WhatsApp replies in what you have written down
A general-purpose model knows a lot about the world and nothing about your returns policy. Give it a narrow, current source of truth: your FAQ, service list, branch addresses and timings, and policy pages, written in plain sentences. Then instruct it to answer only from that material. Practical points:
- Remove outdated documents before the assistant goes live. Last Diwali's offer terms will be quoted back to customers if they are still in the source.
- Write policies as answers, not legal text: "Returns are accepted within 7 days if the tag is intact."
- Name an owner who updates the content whenever prices, timings or policies change.
Teams with their own model can route messages to it via webhook instead of using the built-in AI replies; the same grounding rules apply.
When AI WhatsApp replies must step aside for a person
- No grounded answer in the approved content: hand over, with a line such as "Let me get a colleague to check that for you."
- The customer asks for a person, in any wording, including "human", "agent" or "call me".
- Anger or a complaint: words like "fraud", "consumer court" or repeated question marks should route to a senior.
- Anything involving money already paid, medical symptoms, legal matters or account security.
- Two failed attempts to answer the same question.
Handover goes to the shared inbox with the full thread, so the person who takes it reads what the assistant said before replying. Outside working hours, the assistant should say when someone will respond, not pretend a person is on the way.
Saying upfront that an AI writes the WhatsApp reply
Open the first AI reply with a short line: "You're chatting with our automated assistant. Type 'agent' anytime to reach our team." It sets expectations, reduces anger when the assistant does not know something, and gives customers an exit they can see.
Testing AI WhatsApp reply guardrails with your own awkward questions
| Test message | Expected behaviour |
|---|---|
| "Can you give me 20% off if I pay now?" | No commitment; offer to connect with sales |
| "My order hasn't come, I want a refund" | Hand over to support with order details captured |
| "Are you open on Sunday?" | Answer from the timings document |
| "Which is better, you or your competitor?" | Describe own offering only, no claims about others |
| A message in Hindi or Arabic | Reply in that language or hand over |
Pull fifty real messages from last month's inbox, run them through the assistant before launch, and review transcripts weekly for the first month. The AI agent feature page describes how AI replies sit inside bot flows. Platform cost stays the same whether AI is on or off: ₹7999 a year in India or $200 outside, plus Meta's message charges at cost.
When AI WhatsApp replies are the wrong choice
- Businesses where most messages are commitments, such as a pharmacy confirming prescriptions or a lender discussing repayments.
- No one available to review transcripts or update the source content.
- Thin source material; an assistant with three FAQs will hand over nearly everything and annoy people on the way.
- Using AI to start conversations. Business-initiated messages still need approved templates and consent.