How to stop an AI agent from making up prices and discounts

A customer sends a screenshot: your AI agent told her the jamdani was ৳3,200. Your price is ৳4,800. She has already told two friends. Now you can honour a price that loses money or argue with a customer who has proof.

Every shop that uses an AI agent fears this moment. The good news is that wrong prices are a solvable problem, and the fix is mostly about how the agent gets its prices, not about finding a "smarter" AI. Here are the causes and the fixes, in the order that makes the biggest difference.

Why do AI agents make up prices?

AI agents make up prices because a language model is built to produce a likely-sounding answer, and a likely-sounding price is easy to produce. If the real price is not in front of the model at that moment, it may fill the gap.

In shop AI agents, wrong prices almost always trace back to one of five causes:

Notice that only the last one is really about the AI's behaviour. The other four are data and design problems. That is good news, because data and design are in your control.

How do I keep AI agent prices in one place?

Keep each price in exactly one source, usually your product catalog, and remove it from everywhere else the AI agent reads. This single step removes most wrong quotes.

Go through the AI agent's knowledge and delete or rewrite:

• Price lists in PDFs or Excel files that duplicate the catalog.
• FAQs like "How much is the cotton panjabi? ৳1,450."
• Product descriptions that mention a price ("Now only ৳999!").
• Expired offers and festival price lists.

If a policy needs a number, such as a delivery charge, it lives in one entry only. Our guide to the AI agent knowledge base covers this "one fact, one place" rule in detail.

Should the AI agent look up the price every time?

Yes. Make the AI agent search your catalog for the product before it states any price, option or stock level. A lookup puts the real price in front of the AI at the moment it answers, which an instruction alone cannot do.

This is the difference between a plain chatbot and an AI agent. A plain chatbot reads whatever text matched the question. An agent calls a "search products" tool, gets back the live price, sale price and stock, and answers from that result. If the search returns nothing, the agent has nothing to quote, so it asks the customer to describe the item again or hands over.

Customer: oi maroon kurti ta koto?
[search: "maroon kurti" -> Maroon Embroidered Kurti, ৳1,650, was ৳1,950]
Bot: Maroon Embroidered Kurti er dam ৳1,650 (regular ৳1,950). Kon size lagbe?

Who should calculate totals and delivery charges?

Your system should calculate delivery charges, discounts and order totals from fixed rules, and the AI should only repeat the result. Language models can make arithmetic slips, and a single slip in a total becomes a dispute at the door.

When a customer sees an exact summary and confirms it, any earlier mistake gets caught before it becomes an order.

How do I block amounts that are not real prices?

Add a check that scans each reply for money amounts and catches any that do not match a real price in your catalog. This is the safety net for the rare case where the AI writes a number anyway.

A good check does not just delete the number. It sends the reply back to be rewritten, and if that still fails, it replaces the reply with a safe message such as "let me check the exact price for you". Either way, the invented amount never reaches the customer.

How should the AI agent handle bargaining and discounts?

Write short, firm rules for the moments customers push on price: bargaining, "last time price", and discount requests. Bangladeshi customers bargain, and an AI trained to be helpful may want to agree.

- Never offer or confirm a discount that is not in our written offers.
- If a customer says we quoted a different price before, give the current catalog price
  and offer to connect them with the team.
- Wholesale or bulk prices (10+ pieces): hand over to the team. Do not quote.
- If a customer asks "last price koto?", give the listed price only.

What should the AI agent say when it cannot find a price?

Give the AI agent a fallback message and a handover path for any product or price it cannot find. An agent with no way out is the agent most likely to guess.

Ei product er dam ami ekhon nishchit korte parchhi na. Amader team er ekjon apnake confirm korben.

Then make sure the handover reaches a person who answers. Read our guide to human handover for what makes it work.

How do I test an AI agent for invented prices?

Test the AI agent with questions designed to make it invent a price, before your customers find the weak spot. Run these ten in a test chat:

1. A product you sell, by exact name.
2. The same product, described vaguely in Banglish.
3. A product you do not sell.
4. A sold-out size or colour.
5. "Apni to last time 900 bolechilen."
6. "10% discount den na apu."
7. "3 ta nile koto porbe, delivery shoho?"
8. "Wholesale price koto, 50 piece?"
9. A product whose price changed this week.
10. A question in Bangla script about the same item.

Pass means: the right product, the exact catalog price or no price at all, a correct total from the system, and a handover where needed. Run the same ten again after every big catalog change. If your products are not yet clean enough to pass, start with our guide on training an AI agent on your product catalog.

How does SaleMag prevent wrong prices?

SaleMag's AI order taking uses three of these safeguards together. First, the AI must search your catalog before it states a price, option or stock level, and results include the live price, any compare-at price and per-option prices. Second, any money amount in the AI's reply that is not a real price from your catalog is caught; the reply is rewritten once, then replaced with a safe fallback. Third, the server calculates delivery fees and totals from your settings, and when the order is placed every line is re-priced live. If a price changed, the customer sees an updated summary before anything is saved.

Without order taking, SaleMag's AI agent still answers only from your knowledge base and is instructed never to invent prices, policies or links. If nothing relevant is found, it sends your fallback reply or hands the chat to your team.

Next steps

This week, delete every duplicate price from your AI agent's knowledge, then run the ten price traps above. Fix the data first and the rules second. If you want catalog lookups, a price check and server-side totals built in, start free with SaleMag or see pricing.