Documents and AI
Which values are worth extracting
Extracting too much creates review work and returns nothing.
The temptation when configuring a document type is to tick everything that appears on the paper. It is an expensive mistake: every extracted value is one someone will have to review when it comes back low-confidence, and most will never be used.
The question that decides
For each field: **will I filter, alert or decide with this?** If the answer is no, do not extract it. The document is still stored whole and searchable by content; extracting a value is something else.
| Value | Extract it? | Why |
|---|---|---|
| Expiry date | Always | It is what fires the alert |
| Tax identifier | Yes | It lets you cross-reference with your system |
| Insured amount | If you require a minimum | If you never check it, it serves nothing |
| Insurer name | Rarely | You do not filter on it |
| Full address | Almost never | It is in the document if needed |
Important
Extracting personal data you will not use is storing it twice: in the document and as a loose value. If you do not need it to filter or alert, do not extract it — it is the practical application of asking for the minimum.
The sign you overdid it
- Nobody reviews the low-confidence alerts because there are too many.
- There are fields nobody has looked at in months.
- People correct values that change no decision.
Worth knowing
Start with two or three fields per type. Adding one later takes a minute; unlearning the habit of reviewing forty useless fields takes considerably longer.
›Can I change the fields on a type already in use?
Yes. What was already extracted is kept.
›Can values be extracted without anyone reviewing?
Yes, if the value fires nothing critical. Review is reserved for what decides something.
›What if the document does not carry the value?
It stays empty. That is information: it means that document does not serve what you wanted.
A real case
The situation
A company configures fourteen fields per certificate and nobody reviews the alerts any more.
What you do
- Keeps three: expiry, tax identifier and insured amount
What you get
Low-confidence alerts drop to a handful and get looked at again.
The situation
Twenty fields are extracted and only three are used.
What you do
- Extracts what somebody will actually consult
What you get
Review narrows to the useful.
The situation
A search is made on a field nobody extracted.
What you do
- Adds that field to the extraction
What you get
Search finds what is needed.
The situation
Fields already held elsewhere are extracted.
What you do
- Checks whether the figure already exists
What you get
Information is not duplicated.
The situation
A warning depends on a date that is not extracted.
What you do
- Extracts the validity date
What you get
The warning works.
The situation
Each document type is configured differently.
What you do
- Defines what to extract per type
What you get
Documents of the same type are comparable.
This article answers
- which fields to configure for extraction
- field map for a document type
- extracting too many values
- configure reading per document type