Last month, a dental clinic in Jaipur added an AI chatbot to their WhatsApp Business number. Seemed like a smart move — patients could ask about treatments, pricing, availability, all without tying up the receptionist.
Within the first week, the chatbot told a patient that a root canal costs ₹3,500.
The actual price? ₹8,000.
The patient walked in, phone in hand, screenshot ready, fully expecting to pay ₹3,500. The clinic had two options: honor a price they never quoted, or argue with a patient holding proof that "your own system" said otherwise.
This isn't a made-up horror story. This is what happens when you connect a general-purpose AI to your WhatsApp and let it talk to your customers without guardrails. The chatbot didn't lie on purpose. It did something worse — it guessed, confidently, and got it completely wrong.
And if you think this only happens to small clinics, let me tell you about Air Canada.
When a Chatbot's Lie Became a Legal Liability
In 2024, Air Canada's customer support chatbot told a passenger he could buy a full-fare ticket now and apply for a bereavement discount later — within 90 days of purchase.
That policy didn't exist. Air Canada never offered retroactive bereavement fares.
The passenger bought the ticket, tried to claim the discount, got denied, and took the airline to a tribunal. Air Canada's defense? "The chatbot is a separate legal entity and the airline isn't responsible for its answers."
The tribunal didn't buy it. They ruled that Air Canada was fully liable for what its chatbot said. The airline had to pay the difference.
The legal precedent this set is significant. A court essentially said: if your AI says it, you own it. Your chatbot's words carry the same weight as your employee's words. If it makes up a policy, you're on the hook for that policy.
This was a major airline with presumably a serious AI budget. Imagine what happens when a small business in India deploys a ₹999/month chatbot and lets it loose on customer queries about pricing, refund policies, insurance coverage, or appointment availability.
Why WhatsApp Chatbots Make Things Up
Here's the thing most WhatsApp API providers won't tell you: the AI models powering their chatbots — GPT, Gemini, Claude, whatever — were never designed to know your business.
These are large language models. They're trained on billions of words from the internet. They're incredibly good at sounding human. But they work by predicting the most statistically likely next word in a sentence, not by looking up the right answer.
Think of it like this: imagine asking a very articulate stranger for directions in your city. They've never been there, but they've read a lot about cities in general. They'll give you directions that sound perfectly reasonable — confident tone, specific street names, clear turns. But the streets might not exist. The turns might lead nowhere.
That's exactly what a general AI chatbot does when a customer asks "How much does teeth whitening cost at your clinic?"
The model doesn't know your prices. It was never trained on your price list. But it knows that teeth whitening in India "typically" costs between ₹5,000 and ₹15,000. So it picks a number that sounds right.
And it says it with complete confidence. No hesitation. No "I'm not sure." Just a clean, professional, utterly wrong answer.
This is called hallucination — when an AI generates information that sounds correct but is factually wrong or completely fabricated. And it's not a bug that will be fixed in the next update. It's how these models fundamentally work.
In a general chat with ChatGPT, hallucination is annoying. In a customer-facing WhatsApp chatbot that handles pricing, policies, appointments, and medical information, it's a business risk.
What RAG Is and Why It Fixes This
There's a technique that solves this problem almost entirely. It's called RAG — Retrieval-Augmented Generation. The name is technical, but the idea is dead simple.
Instead of letting the AI answer from its general training knowledge (which doesn't include your business), you force it to search your actual documents first — your price list, your FAQ, your policy documents, your service catalog — and generate an answer only from what it finds there.
Here's how it works in 30 seconds:
Customer sends a question on WhatsApp. "How much does teeth whitening cost?"
The system searches your uploaded documents. It finds your price list PDF where teeth whitening is listed at ₹7,500.
The AI generates an answer using only that document. "Teeth whitening at our clinic costs ₹7,500. Would you like to book an appointment?"
If the document doesn't contain the answer, the AI says: "I don't have that information right now. Let me connect you with our team." No guessing. No making things up.
That step 4 is the crucial part. A RAG-based system has a built-in "I don't know" response. A generic AI chatbot doesn't — it will always try to answer, even when it has no basis for the answer.
The difference between a helpful AI assistant and a liability on your WhatsApp number is whether the system can say "I don't know" instead of confidently making something up.
Generic AI Bot vs. RAG-Grounded Bot: Side by Side
Let me show you what the difference actually looks like in a real WhatsApp conversation.
Customer message: "Do you offer EMI options for dental implants?"
*Generic AI chatbot responds: *"Yes, we offer EMI options for dental implants. You can avail of 0% EMI for up to 12 months on select credit cards. The total cost of a dental implant starts at ₹25,000, and you can split the payment into easy monthly installments. Please visit our clinic for more details."
Sounds great, right? Professional, helpful, specific. Except: this clinic doesn't offer 0% EMI. They don't accept credit card EMI at all. Their implant pricing starts at ₹35,000, not ₹25,000. Every single detail in that response was fabricated. The chatbot pulled "typical" dental implant information from its training data and presented it as this clinic's actual offering.
RAG-grounded chatbot responds: "Yes, we offer EMI options through Bajaj Finserv No-Cost EMI for 3 and 6 month tenures. Dental implant pricing depends on the type — our team can share exact costs after a consultation. Want me to book a free assessment?"
This response came directly from the clinic's uploaded finance policy document. Every claim is traceable to an actual document. Nothing was invented.
Now multiply this across hundreds of customer conversations per month. Every hallucinated answer is a potential refund, a disappointed patient, a negative Google review, or in Air Canada's case, a legal ruling against you.
What to Look for in a WhatsApp AI Platform
If you're evaluating WhatsApp AI chatbots — or already running one and starting to notice problems — here's what separates a safe platform from a risky one:
Document-only answering. The AI should answer exclusively from your uploaded documents — PDFs, spreadsheets, text files, whatever contains your business information. If it can't find the answer in your documents, it shouldn't answer at all. This is non-negotiable.
A real "I don't know" fallback. This sounds obvious, but most chatbot platforms don't have this. They're designed to always generate a response because "unanswered queries" look bad on dashboards. A good system treats "I don't know, let me connect you with a human" as a feature, not a failure.
Human handoff. When the AI can't answer, the conversation should seamlessly transfer to a real person on your team — in the same WhatsApp thread, without the customer having to restart. If the handoff is clunky, customers bounce.
Knowledge base freshness. Your documents change. Prices update, policies change, new services get added. The platform should let you re-upload documents and have the AI reflect the changes immediately — not after a 48-hour "retraining" cycle.
**Source transparency. **Ideally, you should be able to see which document the AI pulled its answer from. This lets you audit responses, catch errors early, and build confidence that the system is actually using your data.
This Problem Is Only Getting Worse
More businesses in India are adding AI chatbots to WhatsApp every week. The WhatsApp Business API is getting cheaper, the chatbot tools are getting easier to set up, and "AI-powered customer support" sounds impressive on a website.
But very few businesses are asking the right question before deployment:** what happens when the AI is wrong?**
Not "if." When. Because every general-purpose AI model will hallucinate. The question is whether you've built a system that prevents those hallucinations from reaching your customers — or whether you've given a confident, articulate liar direct access to your WhatsApp number.
RAG isn't a silver bullet. Your documents need to be accurate and updated. Your system prompt needs to be well-designed. Your human handoff needs to work smoothly. But it's the single most effective technique available today for keeping a WhatsApp AI chatbot honest.
This is exactly the approach we took with Maya, the AI agent inside Instant. She answers customer queries on WhatsApp using only documents you upload — your price lists, FAQs, policies, catalogs. When a question falls outside what your documents cover, she tells the customer she doesn't have that information and routes the conversation to your team. No guessing. No invented prices. No fabricated policies.
If your business runs on WhatsApp and you're thinking about AI — or already running a chatbot that's been giving customers wrong answers — the fix isn't a better AI model. It's a better architecture. RAG is that architecture.
I'm building Instant — a zero-markup WhatsApp Business API platform with Maya, a RAG-powered AI agent that only answers from your documents. We're an Official Meta Tech Provider. If your current chatbot is making things up, see how Maya works.
Top comments (0)