What is document intelligence?

Five steps, and most systems perform all of them.

The comparison people most often want, and the distinction is genuine rather than marketing.

OCR, optical character recognition, converts an image of text into characters a computer can read. That is its whole job, and it is very good at it. Given a scanned pay stub, OCR returns the words and numbers on the page.

What OCR does not do is understand any of it. It has no view on which number is gross pay, which date is the pay period, or whether the document is a pay stub at all. It produces text, not meaning.

Document intelligence uses OCR as one step and continues past it. It classifies the document, identifies which value is which, checks whether the figures make sense together, and decides what should happen next.

The practical difference: OCR gives you a searchable document. Document intelligence gives you a figure you can act on, and tells you when it is not confident.

Systems described as IDP nearly always include OCR. Systems described as OCR generally stop at text.

Three adjacent categories get confused with it.

It is not a document management system. A DMS stores, versions and controls access to documents. It is concerned with where a document lives, not with what is inside it. Many organizations run both.

It is not robotic process automation. RPA automates the clicks and keystrokes a person would perform in an interface. It follows a script and it breaks when the screen changes. Document intelligence reads content and handles variation, which is the opposite problem.

It is not a general-purpose language model on its own. A model that can read a document is one component. A working system also needs to know which documents matter, what to extract from each, how to check the result, and what to do when confidence is low.

The common thread: each of these solves a real problem, and none of them answers the question of what a specific figure on a specific document actually is.

The technology is the same across industries. What changes is the document set and what happens after the reading.

In lending, the documents are pay stubs, W-2s, tax returns and bank statements, and the output feeds an application and a condition list. Mortgage document automation exists because the same figure is typed from a document into a form and then compared back against that document later.

In tax and accounting, the documents are prior-year returns, income forms and supporting records, and the output populates an engagement. Accounting document automation is driven by seasonality more than volume: the same firm handles a year of work in fourteen weeks.

In banking, the documents are identity records, statements and account paperwork, and the output feeds onboarding and monitoring. Financial document automation here is usually about time-to-open rather than headcount.

In healthcare, the documents are intake forms, insurance records, referrals and clinical correspondence, and the output feeds patient records and billing. Healthcare document processing carries privacy constraints the other sectors do not, which shapes how systems are built and where documents are allowed to go.

Enterprise document automation, as a term, usually means a shared capability serving several of these at once rather than a separate system per department.

Document quality is the largest variable and the least discussed.

Layout that moves. Two employers produce pay stubs with the same information in entirely different places. A system keyed to positions rather than meaning fails the moment a layout changes.

Poor capture. A photograph taken at an angle, a scan at low resolution, a page copied several times. The information is present and barely recoverable.

Handwriting. Still substantially harder than print, and common on forms that were designed to be completed by hand.

Multi-document files. A single PDF containing six months of statements, or a scan where three documents run together with no separation.

Language that is technically correct and locally specific. Terms that mean one thing in one institution and something different in another.

Systems that report high accuracy on clean documents and struggle in production usually met one of these rather than a modeling limitation.

Any honest account of this technology includes the reviewer.

Confidence varies by field and by document. A well-formed value on a clean document is close to certain. An ambiguous figure on a poor scan is not, and the useful behaviour is to say so rather than to guess.

Some decisions are not extraction problems at all. Whether an unusual income arrangement is acceptable, whether an explanation is sufficient, whether a document that technically satisfies a requirement actually addresses the concern behind it. None of these are questions about what a document says.

Accountability sits with a person. In regulated work, someone signs. That person needs to be able to see what was extracted and correct it.

What good systems change is not whether review happens but how much reaches the reviewer. Instead of reading every document, a person handles what was flagged.

Any vendor describing full automation of document review in regulated work is describing something narrower than it sounds.

The gain is rarely where people expect.

Rekeying disappears first. A figure read from a document and typed into a form is the most common source of a discrepancy that surfaces weeks later. Removing the typing removes the discrepancy.

Missing items surface earlier. A system that knows what a complete file looks like can say what is absent on the day something arrives, rather than at the point someone needs it.

Reuse becomes possible. A document verified once can satisfy a later requirement, because the system knows what it contains rather than only that it exists.

Review gets concentrated. Attention goes to the exceptions rather than being spread evenly across everything.

What does not change: the documents still have to arrive. Nothing here makes a client send a bank statement faster, and in most workflows that wait is the largest single delay.