AI quote extraction: keeping the full delivery term

Six weeks can mean different dates

A buyer has received two quotes. The first promises delivery „six weeks after drawing approval”, the second „six weeks from order confirmation”. When a tool writes only „six weeks” for both, the table looks finished and leaves out the difference planning needs.

A good reading records the length of the period and the starting event separately, and shows the passage of the quote with the full wording next to them. These details do not yet give a delivery date, because nobody knows when the drawing will be approved. When a quote gives no start of the period at all, the tool leaves it as not stated. There is no basis for adding „from the order” based on the supplier's other quotes.

What a simpler change gives and what training adds

In this count a shared supplier form gives most of the effect at a fraction of the cost. Training makes sense at a larger scale or when the remaining errors are expensive. Further training will not make the missing part fit into a badly designed summary, so we check the record in the purchasing system before deciding on training.

The set of training examples needs maintaining: when the process, the naming or the base model changes, the evaluation is repeated on the same test set. We include that cost in the quote together with the training itself.

First, the way quotes are collected

If the company can ask suppliers for a shared response form, a separate question about the length of the period and the starting event removes the problem before any reading. Some suppliers will still send their own documents. With a fixed layout, a template-based reading is enough. Microsoft describes custom extraction models: template models for stable forms and neural models for varied documents, adapted on labelled examples. The same vendor separates recognising text from assigning values to fields and describes confidence scores that help pick results to check.

Where to start

The buyer assesses the trial on quotes not used to adapt the model: does the tool keep the starting event, and is the source easy to open? For the call, describe one case: which term was in the quote, what reached the system and what the buyer had to add. We adapt models as part of private LLMs, and build the app where an employee sees the result next to the source as part of AI apps. For the free process review, a few examples of an error that keeps coming back and a note of who corrects it today are enough.