Four steps, in the order your land team would do them by hand.

No person opens a document. The machine reads file names, paths and sizes, fingerprints each file to find exact duplicates, and compares page images to find near-identical scans. It reports back: how many pages, how many duplicates, what the folder names already say about document type, and what is still unknown.

In the rooms we have inventoried so far, between a fifth and two-thirds of the files were duplicates; the jobs on our Work page carry the counts. Nobody should pay to process the same document twice, in our system or anyone else's. This pass tells you what is in the room. No file is opened until steps two and three.

Bulk PDFs become individual documents with their pages and exhibits intact and their order preserved. Each document is assigned a type, in our taxonomy or yours, and the distinctions land people care about hold: a memorandum is not a lease, an amendment is not a ratification, a pooling declaration is not a pooling agreement.

Every call carries a confidence score so reviewers know where to look first. The score is checked against your answers, not ours: before anything runs at scale we run a sample against documents your team has already reviewed and compare. You decide how far down the list your reviewers go. On one scan project, what reached the reviewer was about a dozen pages a box, out of roughly fifteen hundred.

We pull the fields that connect a document to everything else: agreement number, API, book and page, legal description, parties, dates. Clauses that change what an asset is worth (Pugh, depth severance, shut-in, consent to assign, continuous drilling) come out verbatim, for the record, followed by a plain reading of what each means, for your team to accept or overrule.

Then we link on the strongest identifier first: agreement number and API, then book and page, then legal description. The lessor's name is the last resort, and any link made on it is flagged. Every document lands on its lease, its wells, its unit and its owner. For a right-of-way file the rungs change but the order doesn't: recording reference first, then the strip or tract description, then grantor and grantee, so each grant is chained through its assignments to today's holder with width and consent language beside it. Where the record won't support a link, we say so rather than guess.

You get files, CSV or JSON, or import rows shaped for your system in your taxonomy, with every extracted value traceable to its page. Your team or your integrator loads them into M-Files, Quorum, Peloton LandView or whatever you run; where we partner with the system vendor, the rows are shaped with them. There is a REST API and SDKs for teams that want to wire it in.

We are not a land system, and we don't try to become one.