Should a language model calculate the total on a quote?
No. A language model generating a total is a probabilistic output, so it will be right almost every time, and the failure mode is a plausible wrong figure inside a document your client has already accepted. That is not a typo you spot on review, it is a commercial commitment you have to unwind while the other side holds a signed copy. The right architecture is to let the model choose which items from your price list the customer is asking for, and to compute every figure in code: unit prices, volume tiers, discounts, deposits, tax and the total.
Asked because you are choosing a tool? DocuDeal is free for 3 documents a month, $149 on Pro for the whole workspace with unlimited users.
Why almost always right is the problem
A calculator that fails loudly one time in a thousand is easy to work with, because you see the failure. A system that is quietly wrong one time in a thousand, in a document that looks exactly like the other nine hundred and ninety-nine, is a system you cannot review your way out of. Nobody re-checks the arithmetic on a document that looks correct.
The line to draw, concretely
Model: which catalog items apply, scope, assumptions, exclusions, the prose, the payment-schedule wording. Code: the unit price, which volume tier applies, line and table discounts, deposits, recurring intervals, tax, the subtotal and the total. Anything that cannot be sourced from the catalog is flagged to confirm rather than filled in with something reasonable-looking.
The question to ask a vendor
"Which half of the pricing is your model doing?" It is one sentence and most demos are carefully constructed not to answer it. A vendor whose model produces the total will tell you it is very accurate, which is a description of the risk rather than a mitigation of it.
The thing that makes it work at all
A catalog that holds cost beside price, not just price. Then a discount can be measured against margin rather than only against the total, and an approval rule can fire on a discount four points past policy before the document leaves, rather than being discovered in the signed copy.
| Task | Model | Code |
|---|---|---|
| Which items the customer wants | Yes | — |
| Scope, assumptions, exclusions | Yes | — |
| Section prose and covering note | Yes | — |
| Unit price and billing interval | Never | Yes |
| Which volume tier applies | Never | Yes |
| Discounts, deposits, tax | Never | Yes |
| The total | Never | Yes |
| A figure it cannot source | Never invented | Flagged to confirm |
How DocuDeal divides it. Worth asking any AI document vendor to fill in the same table for their product.