The records exist. They are in filing cabinets, in scanned boxes on a shared drive, on microfilm somebody digitized in 2004. Somewhere in them is the answer to a question your team asks every week.
Nobody looks. Finding one number means a person walking to a cabinet and reading paper for an afternoon, so instead your team works from the last ten years of data that happens to be in a system, and treats the rest as though it does not exist. Decades of expensive fieldwork sitting there, unused.
Why OCR does not solve this
The obvious move is to run everything through OCR and load the text into a database. It is the first thing everyone tries, and it fails in the same places every time.
Handwriting. These records were filled in by hand, by dozens of different people, over decades. OCR was built for typed pages. It returns confident garbage on a field note.
Tables that break across pages. A single record often runs over several sheets. OCR reads one page at a time and has no idea the second page continues the first, so one record becomes three fragments.
Every form is different. The layout changed in 1978, again in 1991, and one regional office always did it their own way. Anything built around fixed field positions breaks on the next box you open.
No sense of what the values mean. OCR gives you characters. It does not know that a number in one column is a depth and the one beside it is a grade, so it cannot catch a reading that is off by a factor of ten.
Damage and degradation. Faded ink, coffee stains, skewed scans, notes written up the margin. A person reads through all of that. Character recognition does not.
What you get back is a pile of text that still needs a person to interpret it. That is the problem you started with.
What works: reading the page, not the characters
Vision models read a page the way a person does. They see the layout, the handwriting, the table structure, and the notes in the margin at the same time, and they interpret them together. A column header two pages back is still context when the model gets to page three.
That changes what is possible. Instead of recognizing characters and hoping the structure survives, the system reads the record, understands what kind of record it is, and pulls out the fields that record type is supposed to have.
Around that sits an agentic pipeline. Different document types get routed to handlers built for them. Records that span pages get assembled before extraction rather than after. Values get checked against what is physically plausible for that field and flagged when they are not. Anything the system is unsure about goes to a person instead of quietly into the database.
In practice
A geological exploration company had roughly fifty years of field records across several regions. Handwritten, multiple form generations, varying scan quality, and no consistent index. The historical data was the company's most valuable asset and the least accessible thing it owned.
The system now reads those records and produces a structured database with the source page attached to every value. Geologists query the whole archive instead of the last decade. Data that was effectively lost is in the same tools as everything else.
We built the system, then turned it into a product: LogDog, our drill log digitization product.
AI drill log digitization for mineral exploration.
The exploration version of this problem turned out to be the same in every company that had it, so we packaged it. LogDog is that system sold self serve to geologists. Every other domain is an engagement, which is what the rest of this page describes.
What you get out
The output is a database rather than a text dump.
Every extracted value carries its source, so anyone can check a number against the page it came from in a couple of seconds. Output lands in the formats your existing tools already read, so nobody has to learn a new system to use it. Uncertain values arrive in a review queue rather than silently in a table. And the archive stops being a liability you keep meaning to deal with.
Whether this fits your archive
The domain above is geology. The shape of the problem is not. Utilities, land records, manufacturing quality files, clinical archives, insurance claims, and municipal permits all have some version of it. Decades of records, filled in by hand, in forms that changed several times, holding information the organization still needs.
If that sounds like what is in your cabinets, the first question is not what technology to use. It is how many document types you actually have and how much of the content is judgment. That is usually a short exercise.