The AI reads the prices, code does the sums.

A language model writes fluent text, but it can miscalculate. If it adds up five fees and splits the total into three payments, the result is usually right. When it is wrong, the text looks the same.

In a contract, both sides then sign the wrong figure.

The prices come from a source, for example the transcript of a sales call or an email thread. With the prices, the model has one job: find every fee that was agreed and write it down as a list. Each entry holds the amount, how often it is paid, and the passage it comes from.

Code does the adding up. It subtracts discounts and credits and builds the total and the payment schedule. Then it checks two sums. The fees have to equal the total, and so do the payments in the schedule. If one of them is off, the result is discarded.

A fee without a passage from the source does not pass the check. The model then has to read the source again.

Once the list passes, a second check compares it with the source and reports what does not fit.

A separate step writes the text of the document. By then the prices are fixed, and the text has to use them unchanged. The price table in the document is built from the calculated figures.

If one document follows another, for example a contract after a proposal, the second takes over the same figures, including changes a person made. The prices are not read from the source a second time.

Every fee can be changed before the document goes out.

When a document goes out for signature, it is locked, so the version that was reviewed is the one that gets signed.

InkDraft is my product that turns a sales call into a proposal and a contract, with signature and payment. It works as described here and is live on paid plans.