If your WhatsApp bot has ever told a customer you carry something you don't, quoted a price that's three months stale, or insisted an out-of-stock item is sitting on the shelf, that's not a one-off glitch you patch and forget. It's what happens by default when a language model isn't forced to check its answer against your actual catalog before hitting send. The model isn't lying. It's guessing, fluently, and your customer has no way to tell a guess from a fact.
Here's the part most brand owners don't realise until it bites them: the AI agent doesn't “know” your inventory the way a staff member would. It was trained to sound helpful and confident. If nobody builds a hard stop between “sounds right” and “actually exists in our system,” it will fill gaps with whatever fits the pattern of a normal answer. Ask it about a product that's almost in your catalog, a slightly different colorway, a size you discontinued last season, and it will often just invent the missing detail rather than say it's not sure.
A quick example of what this looks like in practice
A customer messages asking for a blue kurta in size M. If the bot has no live link into current stock, here is a typical bad answer: “Yes, we have the blue kurta in size M available for ₹1,299, would you like to order?” Confident, specific, completely made up.
Compare that to what a properly grounded agent should say: it checks the live catalog first, finds the blue kurta only comes in S and L right now, and replies with that fact, maybe offering the closest in-stock alternative. One of these answers loses you a refund and a one-star review. The other one looks almost boring by comparison, and that's exactly the point.
Why this happens more than people expect
Three things compound here. The model's training data is frozen at some point in the past, so it has no idea what you're selling this morning. Catalog feeds often lag behind the live store by hours or days, so even a system that's “checking the catalog” might be checking a stale copy. And most AI chatbot builders treat the model's raw output as the final answer, with no verification step in between generation and delivery.
Stack those three together and hallucinated product details aren't an edge case. They are the expected outcome on a long enough timeline.
What catalog grounding actually means
The fix isn't a better prompt. It's an architectural decision. Before the agent answers a shopping question, it should run a retrieval step against your live product data, pulling back the specific items that actually match the query. Then, after the model drafts a reply, a second pass checks every product mentioned in that draft against the catalog one more time. Anything that doesn't check out gets stripped before the customer ever sees it.
That second pass matters more than people assume. The first retrieval step reduces hallucination. Only the verification step on the way out the door eliminates it.
Three questions worth asking before you trust any AI agent with your catalog
- Does it verify every product mention against your live catalog before sending, or only at the point of initial retrieval?
- What happens when there's no good match? A confident wrong answer is worse than an honest “we don't carry that, but here's something close.”
- Can you actually see when this happens? If a vendor can't show you a log of stripped or corrected responses, you have no way of knowing how often it's quietly catching mistakes, or whether it's catching them at all.
VritantAI Convert runs this exact two-step process on every reply: retrieval against your live catalog first, then a hallucination guard that checks every product mention again right before the message goes out. Every strip event is logged, so you can see exactly how often the guard is catching mistakes and what it corrected.
Stop hallucinated product recommendations before they reach your customers
VritantAI Convert connects to your Shopify or WooCommerce catalog, runs a real-time hallucination guard on every WhatsApp and storefront reply, and logs every corrected response so you can see exactly what was caught.
See how Convert works →Frequently asked questions
What is AI hallucination in a shopping chatbot?
It's when an AI agent states something about a product, its price, availability, or features, that isn't actually true. This usually happens because nothing checked the claim against real data before it was sent.
Can a WhatsApp chatbot really invent products that don't exist?
Yes. If the underlying model isn't constrained to only mention items it can verify, it will sometimes describe products, variants, or stock levels that have no basis in your actual catalog. This isn't a bug in the model; it's the expected behaviour when there is no verification layer.
How do you stop an AI agent from hallucinating products?
By grounding every response in a live catalog lookup and adding a verification pass that checks the model's draft answer against that same catalog before the message is delivered. The first step reduces hallucination. The second step eliminates it.
Does catalog grounding slow down responses?
A well-built version adds milliseconds, not seconds. The retrieval and verification steps run against an indexed catalog, not a live database scan, so the delay is barely noticeable to the customer.