"The bot told a customer the wrong price!" — how we stop AI from making things up

Lovely team · September 2026 · 4 min read

You've probably met that chatbot: you ask if a shirt is in stock, it confidently says yes, then the shop says they don't even sell that shirt. This is called hallucination — an AI confidently inventing an answer it doesn't actually know.

Why it happens

A plain AI chatbot is a language engine, not a business database. It has read millions of web pages, so when you ask "how much is delivery to Khulna?", it produces something that sounds like an answer — a plausible pattern — even though it has never seen your delivery chart. It isn't lying; it's completing sentences. With prices and policies, that's dangerous.

The Lovely approach: answer only from your facts

Lovely is built the opposite way. The AI never answers from its general memory about your business. Before every reply, the system:

1. Searches your knowledge base — the documents, price lists and FAQs you paste in are converted into mathematical "meaning fingerprints" (embeddings), so the bot can find the exact relevant passages for any question.

2. Feeds only those passages + your persona to the AI, with an instruction to answer strictly from them — and to say it doesn't know when the answer isn't there.

3. Falls back honestly. If your knowledge base has no answer, the bot doesn't invent one — it apologizes and points to your hotline/contact, so a human closes the case.

The test playground

Before your bot ever talks to a real customer, you test it in the dashboard's Playground tab. Ask it your customers' actual questions. If it answers wrong, that usually means the knowledge base is missing that fact — you add it, and the bot is corrected for every future customer. Wrong answer → fix the knowledge, not the bot.

What this means for you

The bot is like your best-informed employee who only speaks when they're sure, and hands off when they're not. That's what makes it safe enough to leave answering your page at 1am — and why the customers you gain from instant replies don't turn into refund requests.

Set it up once, test it in Playground, and let it take the night shift.
Start free for 14 days