Intelligent Document Processing

Uploading a contract and asking questions about it is a demonstration, not a system. It reads one file at a time, produces prose rather than fields, remembers nothing between sessions, and ends where the actual work begins — in the ledger, the case record or the policy system.

The work is the part after the answer. A system that reads ten thousand documents and writes nothing into a business system has moved the effort rather than removed it.

Documents arrive from wherever they already arrive: a monitored inbox, a shared drive, an upload, an API. Ingestion normalises them, and parsing turns scans and photographs into text with positions attached, because a value means little without knowing where on the page it sat.

Classification decides what each document is before anything tries to interpret it. An invoice, a purchase order and a delivery note carry different fields and different rules, and guessing that from the content alone is how the wrong extractor runs against the wrong file.

Extraction then pulls a defined set of fields for that document type, as structured values rather than a paragraph describing them. Those values go to a knowledge store alongside the source document, so every field can be traced to the page and position it came from.

Reasoning is where the business context enters: comparing a document against another, against a record, or against a policy. That is also where an exception is raised rather than resolved — a total that disagrees, a clause that does not match the standard, a field the document simply does not contain.

What follows is an action in a real system, or a request for approval. Both are recorded.

Parsing and layout handling for the document surface, a vector store for retrieval across a corpus, a relational database for the extracted fields, and a model chosen for the reasoning step rather than for every step. Classification and extraction frequently do not need the largest model available, and using one everywhere is how a system becomes expensive without becoming better.

Orchestration sequences the stages, retries the ones that fail, and keeps a trace of each run. The architecture follows the documents, not the other way round.

A document system that acts on what it read needs boundaries before it needs accuracy. Every extracted field carries a confidence and a source location, so a low-confidence value can route to a person instead of into a ledger. Every downstream action runs under an identity with defined permissions, and every action is logged — including the ones that were refused.

Confidence thresholds are a business decision rather than a technical one. Which fields may post automatically, which always need a second pair of eyes, and what happens to the remainder are questions the people who own the process answer, and the system enforces.

It does not remove review from work where review is the point. Legal interpretation, disputed amounts and anything with a signature attached stay with a person; the system prepares the decision rather than making it.

It does not work well on documents nobody can define. If two people in the business disagree about what a field means, that disagreement has to be settled before it can be automated — a system will otherwise encode one of the two answers silently.