Glossary
What is an extracted value
The difference between the document and what was pulled out of it, and why they are not worth the same.
Extracted value
A specific value — a date, an amount, a batch number — read from inside a document and stored alongside it, without replacing it, so it can be searched, compared and used for alerts.
It is the most misunderstood piece of the system, and the confusion has practical consequences: many people assume that if the value is wrong the document is wrong, or that if the value is right there is no need to look at the document. Both are false.
What each one is
| Aspect | The document | The extracted value |
|---|---|---|
| What it is | The original file exactly as it arrived | A value read from it |
| What it is for | It is the evidence | It is what enables search and alerts |
| If they disagree | The document wins | The value gets corrected |
| Can it be corrected | No, it is what it is | Yes, and there is a record of who did it |
Important
The third row is the rule that settles almost every doubt: **on any discrepancy, the document wins**. The value exists so you can find and control things; the evidence is still the file. That is why correcting a misread value is not "falsifying" anything — you are adjusting the index, not the original.
What having them is for
Watch out
The nuance you cannot deduce: a document **with no extracted values is still there and still works as evidence**. All it loses is being searchable by content and being able to warn about its dates. It is the difference between an archive that stores and one that also works — and it explains why an unread scan feels "absent" even though it displays perfectly.
When to check by hand
- 1
If it feeds an invoice or a mandatory record
Any figure going outward gets checked against the image.
- 2
If the document was handwritten or corrected
There, reading returns plausible values that may not be the written ones.
- 3
And if the value decides something
An expiry that triggers a block deserves thirty seconds of checking.
Worth knowing
Each AI reading of a document is an action and consumes; consulting the extracted value afterwards as often as you like does not. Hence deciding which fields matter per document type instead of extracting everything from everything.
›Can I add a value that is not in the document?
You can note it, but it should be distinguishable from what was read: they are not the same.
›Does correcting the value change the document?
No. The document is never touched.
›What if the document is replaced by a new version?
Values are re-read from the new version; the previous one keeps its own.
A real case
The situation
Someone sees the extracted date does not match the PDF and assumes the document is wrong.
What you do
- Checks the image and corrects the value, which was what was off
What you get
The expiry alert is right again and the document remains intact as evidence.
The situation
A figure is corrected and people think the document changes.
What you do
- Checks the document is unchanged
What you get
The evidence stays intact even as the figure is adjusted.
The situation
A search by amount does not return the document.
What you do
- Checks whether that field was extracted
What you get
You know whether the figure is missing or the document is.
The situation
Two different documents give the same figure.
What you do
- Looks at which one each came from
What you get
You can reach the source from the figure itself.
The situation
An extracted figure is used in a report without review.
What you do
- Approves first whatever was flagged as uncertain
What you get
The report rests on checked data.
The situation
Somebody edits a figure and nobody knows it changed.
What you do
- Checks the figure's history
What you get
The correction has an author and a date.
This article answers
- what is an extracted value
- difference between document and data
- ai extracted fields
- why does the value not match the pdf