AI quote extraction: keeping the full delivery term

Six weeks can mean different delivery terms

Suppose a buyer receives two quotes. The first states delivery “six weeks after drawing approval,” while the second says “six weeks from order confirmation.” If a tool records only “six weeks” for both, the table looks ready but omits a difference needed for planning.

Useful extraction keeps the duration and its starting event separate. For the first quote, the buyer sees six weeks and drawing approval; for the second, six weeks and order confirmation. Alongside each entry, they can open the passage containing the full wording. They do not need to search the quote again to check whether the condition was copied correctly.

These details do not yet establish a delivery date. The example supplies no date for drawing approval or order confirmation. The table should preserve the conditions so the buyer can continue working with them. The duration alone also does not establish which supplier would deliver sooner.

This task is called data extraction: a tool finds information in a document and transfers it to designated places in a table or system. For the buyer, a useful result means less copying and an easy way back to the source wording.

First check how quotes are collected

If the company can ask suppliers to use a shared response form, include both parts of the delivery term. Separate questions about duration and the event that starts it may resolve the problem before document extraction is needed. Structured responses may need only an ordinary import into the purchasing system.

Some suppliers will continue sending their own documents. A consistent layout may suit template-based extraction. Microsoft distinguishes template models for stable forms from neural extraction models for more varied documents. The latter can be adapted using labeled examples. Source: Microsoft Document Intelligence custom extraction models.

Before commissioning adaptation, try the current tool on quotes whose results buyers regularly correct. If it reads the complete condition but the purchasing system has room only for a number of weeks, the way information is stored also needs attention. Additional model training cannot put the missing detail into a table that has nowhere to hold it.

Reading text and assigning a condition are different problems

A word on a scan may be illegible. In another document, every word may be read correctly, yet the tool attaches “after drawing approval” to the wrong item or drops the phrase when transferring the data. The resulting errors can look similar to the buyer, although different parts of the solution need improvement.

Microsoft distinguishes text recognition from assigning values to fields. It also describes confidence scores for extracted information. A score can help identify a result to review, but it does not replace checking the quote. In our example, the full condition needs to survive even when the number of weeks is read perfectly.

If the quote does not specify when the period starts, the tool should leave that detail unspecified. Other quotes from the same supplier do not justify adding “from order placement.” The buyer can then see what needs clarification with the supplier.

When does a custom model make sense?

A custom quote-extraction application does not always require model training. It can use an existing tool, connect the result to the purchasing system, and show the buyer the relevant source passage. Compare that scope with adaptation before committing to further training.

Fine-tuning is one form of adaptation: further training an existing model using examples with the correct information identified. It is worth testing when the available model repeatedly loses the same conditions across different wording or quote layouts. The purchasing team needs to be able to show what should be transferred and which passage supports it.

The buyer assesses the trial on quotes that were not used to adapt the model. They check whether the tool preserves the event that starts the delivery period and whether the source is easy to open. They also compare how often the existing and proposed extraction methods send a correctly captured condition for unnecessary review.

If the team already has a good comparison table but needs help drawing conclusions from differences between quotes, see our article on AI training for procurement bid comparisons. It explains how buyers examine scope and prepare a question for a supplier. Document extraction provides material for that assessment.

Bring complete quote conditions into your purchasing system

Syntalith builds document applications and adapts AI models. A proposed procurement project would carry the complete condition from a supplier’s quote into the place where the buyer uses it. For a delivery term, that means preserving the number of weeks, the event that starts the period, and a link to the supplier’s wording. We begin by establishing where information is lost: during extraction or when the result is transferred into the system’s fields.

Your purchasing team contributes recurring corrections and explains what the record needs to contain for their work. We use those examples to propose improvements to extraction and its connection to the purchasing system. If an existing tool still drops conditions, we assess whether model adaptation is justified. The buyer should be able to open the source and correct a transferred condition within their working tool.

Describe one case: what the quote said, what reached the system, and what the buyer had to add. That gives us a concrete starting point for discussing which part of the work to include. If documents are needed for further assessment, we will agree on how to use them. See Syntalith’s pricing page for engagement information.