How to Stop Your AI From Inventing Prices Customers Will Demand a Refund For

If your AI quotes a price you don't actually charge, the customer will screenshot it and ask you to honour it. Most of the time you will, because it's faster than arguing — and that's how a "helpful chatbot" turns into a slow monthly leak from your margin. The fix isn't a smarter AI. It's an AI that's wired to only repeat what's actually in your own product list and policies.

The business problem

You're not in business to argue with customers about whether the $89 hoodie is "actually $59 because your AI said so". You're in business to sell hoodies. Every minute spent backpedalling, refunding, or apologising for something a chatbot invented is pure cost — and worse, it teaches your customers not to trust your prices.

Three painful flavours of this:

• Wrong price quoted. Customer screenshots, demands the lower price, posts it in a group chat.
• Wrong availability promised. Customer turns up at your clinic, store, or restaurant and you don't have what was promised.
• Wrong policy invented. "Your bot told me free returns within 60 days" — your actual policy is 14.

None of these are bugs in the strict sense. They are the default behaviour of any AI that wasn't specifically locked to your real data.

What goes wrong with the usual approach

The usual approach looks something like: grab a popular AI tool, type "you are the sales assistant for [my business], be helpful and accurate", paste a couple of product names in, ship it. This fails three ways:

1. "Be accurate" is not enforcement. Telling an AI to be accurate is like telling a sales rep to "always tell the truth" with no way to check. It sounds responsible. It doesn't actually stop anything.
2. Pasting the catalog into the instructions doesn't work either. As soon as you have more than a handful of products, the AI picks the wrong one or makes up a similar-sounding line.
3. There's no safety net before the message goes out. Whatever the AI typed, the customer sees. If it invented a price, the customer sees the invented price.

How to think about it

A reliable AI sales agent has four pieces working together. You don't have to build them — but you should know they exist, because if your tool is missing one, you'll find out the expensive way.

1. A clear rule: only repeat what's in our actual data

The AI is told, in no uncertain terms, to only state prices, stock, dates, policies, or promises that come from your business data — not from its general knowledge or from earlier in the conversation. If it doesn't have an answer, it asks the customer or pings a human. This rule has to be literal: "be accurate" doesn't count.

2. A live connection to your real data

The AI doesn't memorise your catalog. It looks it up live, one product at a time, when the customer asks. That way, when you change a price tomorrow, every reply uses the new price immediately. Same for stock counts, hours, and policies.

3. A way to inject the answer into the conversation

When the customer asks "how much is the navy hoodie?", the system fetches the real product, hands it to the AI as "here is the real answer, use only this", and the AI builds the reply around it. If the lookup comes back empty, the AI is told to ask or escalate — not to fill in the blank.

4. A checker that reads the draft before it's sent

This is the piece almost nobody talks about. Before any reply goes to the customer, a separate check reads the draft and asks: "did this just claim a price, stock, date, or policy that isn't actually in our data?" If yes, the draft is thrown away and the AI tries again — this time politely asking the customer or handing off. This costs almost nothing per check and stops the rare drift that the first three layers miss.

What this looks like in real life

Customer: how much is the navy hoodie?

AI (looks up the live product in your catalog): The Navy Heritage Hoodie is $89, and we have it in stock in S, M, and L. Want me to grab one for you?

Versus the failure mode:

Customer: how much is the navy hoodie?

Generic chatbot (guessing from memory): Our hoodies are usually around $59, would you like to order?

You, later: Now arguing with a customer who has screenshots.

How UptoNova helps

UptoNova ships the four pieces above as the default, on every reply, on every channel. Specifically:

• Your product catalog is the source of truth. Upload it once (or connect Shopify / WooCommerce). Prices, stock, and descriptions flow in. The AI reads from there or doesn't quote.
• Your written policies are loaded into the knowledge base. Refund policy, shipping zones, opening hours, FAQ. The AI pulls the right one when asked.
• Every reply is checked before it leaves. No invented prices, ever. If the check fails, the reply is rewritten or a human is paged.
• Full trace of every reply. You can open any conversation and see exactly which product, which policy, and which check was applied. No black box.
• Same rules across every channel. WhatsApp, Instagram DMs, Messenger, Telegram, groups, and your website widget all use the same AI sales agent with the same safety rules. See our multi-channel inbox guide.
• Closes the sale, doesn't just answer. When the customer says yes, the AI sends a Stripe payment link or books an appointment right there in the chat, and adds the contact to your CRM with the right details.

In short, this is what a grounded AI chatbot should do: a grounded AI never makes up prices, because it can only repeat what your catalog and policies actually say. For small teams, that turns an AI sales agent for WhatsApp into reliable AI customer service for SMB rather than a liability — and when a question falls outside the data, one-click human handover means a person on your team steps in instead of the bot guessing.

A simple test for any tool you're considering

1. Ask the vendor: "if I delete a product from my list, will your AI still talk about it?" The right answer is no — it's gone immediately.
2. Ask: "show me one reply with the source of every fact in it." If they can't, the AI isn't really wired to your data.
3. Ask: "what does your AI do when it doesn't know?" The right answer is "it asks a question or hands off to a human." The wrong answer is "it does its best."
4. Ask: "what stops it from making up a discount?" The right answer involves a check on every reply before it's sent.

Frequently asked questions

Will the AI sound less natural if it's locked to my data?

No. Tone is separate from accuracy. You set the personality — friendly, warm, professional, casual — and the AI matches it. The only thing it won't do is invent facts.

What if a customer asks something I haven't written down?

The AI politely says it doesn't have that info and either asks a clarifying question or notifies a human on your team. You then add the answer to your knowledge base, and from then on the AI handles it directly. Most businesses fill in the gaps in the first week.

Does this work in other languages?

Yes. The AI matches the language the customer wrote in — Spanish, Portuguese, Arabic, Hindi, French, Russian, English, and many more. Your catalog and policies can be in your own language; the AI translates as it answers.

What about chitchat and small talk — does it refuse to chat?

No. The strict rule applies only to claims a customer might hold you to — prices, stock, dates, policies. For "good morning" and "thanks", the AI replies naturally.

How fast does the AI reply?

A few seconds, including the safety check. Customers experience it as "instantly", which is the speed that wins the sale before they go look at a competitor.

Back to UptoNova