One PDF, several documents: how AI can identify them

A contract, an amendment and a message in one file

A law firm receives a bundle: pages 1–4 are a services contract, pages 5–6 a separate amendment, and page 7 a printed email. In the index the whole file is called „contract”, so anyone looking for the amendment has to open it and go through the pages.

A useful proposal shows three entries: the contract on pages 1–4, the amendment on 5–6 and the message on page 7. An employee opens the places named, checks the boundaries and approves the split. The original bundle stays available. The boundary between documents is what matters: if the last page of the contract lands in the amendment, a correct type name helps little. The classifier assigns a document to an agreed type and does not judge whether the document is valid, signed or has legal effect.

What a simpler change gives, and what training adds

In this count, a ready classifier gives most of the effect for a fraction of the cost. Training makes sense at a larger scale or when the remaining errors are expensive. A short cover note does not have to fit the contract category just because it comes before the contract's pages, so the view leaves such parts for an employee to identify.

The set of training examples has to be maintained: when the process, the naming or the base model changes, the evaluation is repeated on the same test set. We put that cost into the quote together with the training itself.

First check the intake and ready-made tools

When senders can send separate, clearly named files, part of the sorting disappears at intake. Your document management system may have features for splitting bundles that the team does not yet use. Microsoft describes the custom classifier in Document Intelligence, which recognises trained document types and their page ranges in a bundle. Judge a wrong document name and a wrongly drawn split separately, because they are different corrections. Customisation makes sense when the tool regularly joins an amendment to the contract before it with similar-looking documents. Further work on organised case files is described in the article on local AI in a law firm.

Where to start

The comparison covers bundles not used for customisation. The person who handles documents checks that every part is in the index and that the links lead to the right pages. We run model customisation as part of private LLMs, and the application in which the employee sees the result beside its source as part of AI applications. For the free analysis, a few examples of an error that keeps coming back and a note on who corrects it today are enough.