How we stop an AI model
from inventing prices in a quote

The problem

Buyers, and increasingly their AI agents, send quote requests as free text: a list of sizes, grades and quantities in an email. A language model can read that. It can also, if asked to "quote", produce a confident price that exists nowhere in your price list. For a supplier that is worse than no tool: a wrong price sent to a customer is a commitment.

The design: the model reads, code prices, a person sends

1. Remove personal data first. Email addresses and phone numbers are stripped from the request before the model sees it.
2. The model only extracts. It returns a list of lines (what was asked for, the quantity, the unit) and any terms the buyer mentioned, such as a delivery date. It is not asked for a price, and it is given no price list.
3. Every item must quote the request. Each line the model returns has to include the exact words from the buyer's request it came from. If we cannot find those words in the request, the line is dropped. A delivery date or payment term is treated the same way.
4. A quantity must appear in the line's own words. If the model says 400 but the line says 4,000 or nothing, the quantity is left blank rather than trusted.
5. Ordinary code prices the lines, from your list. Matching to your price list is deterministic code, not AI. A line is priced only on a product code, or on a clear match where every size and grade in the request is present in the product. Anything else is flagged with the closest item shown, and the unit price is left as a blank.
6. The draft has blanks for decisions. Delivery cost, whether a date can be met, and how long a price is held are left as [blanks] for a person. QuoteDesk never sends anything.

When a request already arrives as structured data (for example, an agent's JSON from a Kavel quote page), the model is skipped entirely: the lines are read directly and every line is, by definition, the buyer's own words.

What this costs

The design prefers precision over recall. It will leave more lines blank than a model that guesses, and a vague line such as "some M8 bolts" gets a question back, not a price. For a supplier checking each draft before sending, a blank is cheap and a wrong price is not.

What it does not do

• It does not check that your price list is right, or that stock exists beyond what a stock column says.
• It does not read PDFs or photos yet; it reads text.
• It is an early prototype, tested on made-up price lists and sample requests. We are running free pilots on suppliers' real past requests to find out where it is wrong.

Written with AI from QuoteDesk's design and code, and checked by a person. The model used for reading requests is Meta's Llama, run by Cloudflare.