Intelligent Document Processing: How to Extract Data With AI

Intelligent document processing is the work of getting software to read a document, find the fields a business needs and pass them on, with a person looking at the values the software is unsure about. The effort sits in five decisions made before and after the software reads anything: which document type, which fields, what check each value must pass, which values a person reviews, and how the team tests it. For the work itself, read the use case on taking fields out of documents automatically. This article puts the decisions in order and says who settles each one.

What Intelligent Document Processing Does to One Document

Four terms come first. A field is one named value on a document, such as a date, a reference number or a quantity. A check is a rule a value must pass before it is accepted, such as a total that adds up. In this article, a doubtful value is one that fails a check or that the software marks as uncertain. A reviewer is the named person who settles each doubtful value against the page itself.

Intelligent document processing works on one page in four moves. To begin with, the software reads the page and finds each named field. It then applies the checks, and it sends the doubtful values to a reviewer. Finally, the accepted values pass to the system that needs them. A team that wants to extract data from documents this way still makes the judgements itself: people choose the fields, write the checks and settle the doubtful values.

Reading the text on a page is one part of this work, not all of it. A page can be read correctly and still give a wrong date to the next system. The decisions below cover what happens after the reading as well. For a firm still weighing how far to go beyond the reading, a side-by-side view is in a table that sets character capture beside field-level handling of each document.

Intelligent Document Processing: Choices Made Around the Reading

Signisys’ working description of the work has five decisions, in this order. It is a description, not a standard; each decision stands on its own here.

The team makes the first three decisions before the software reads a single page. The fourth tells the team whether those decisions hold. Finally, the fifth keeps them true after documents change. The sections below take them in turn.

One Document Type, and the Fields Someone Uses Next

Decision 1 is the narrowest and the one the rest depend on. Ask three questions. Which kind of document does the team handle in the largest volume? Who reads it now? Which values do they type from it? Choose one kind of document for a first run, not the whole inbox. To size up a candidate, how it is handled today is traced in a touch-by-touch look at the hand work on one document type.

For each field, note who uses it next and what a wrong value would break in that step. A purchase order number that feeds a matching step is worth a place on the list. On the other hand, a field that no one downstream uses can come off it.

Then count the layouts. How many different layouts reach the team for this document type, and from how many senders? One layout is a gentler start than many. That is our own reasoning, and a firm with a dozen layouts can still begin; it only needs more test pages later.

Some document types do not suit a first run. Pages with free-flowing text and no fixed fields are a poor start, as there is no named value to check. The separate job of finding answers in a stored pile of documents is covered under searching stored documents with AI.

The Test a Value Must Pass Before It Moves On

Decision 2 turns each field into a check. For each field, put in one sentence what a correct value looks like. For example, a date sits inside a sensible range. A total equals the sum of its lines. A reference number exists in the next system. Finally, a required field holds a value.

The owner of the next step signs each sentence, as that person knows what a wrong value would break. In addition, whoever runs the project keeps the sentences together in one place for a new colleague to read.

Test each sentence on a sample page before it goes into use. If a colleague cannot apply the check without asking what it means, the sentence needs rewriting.

Which Doubtful Values Go to a Reviewer

Decision 3 names who sees what. Two kinds of value need a person: one that fails a written check, and one the software marks as uncertain. The reviewer reads a confidence score as a pointer to where to look first, not as proof that a value is right. The team tests that reading in decision 4, on pages a person has read.

Name the reviewer and a second person who covers when the reviewer is away. Without that second name, doubtful values wait. The reviewer sees the page beside the extracted values and corrects a value in one place. The team keeps the corrections, which lets it see which fields cause the doubt.

In particular, for the fields that matter more than the rest, name a reviewer who is not the person who first read the page. A fresh pair of eyes gives the page a second reading.

Put two questions to the team. Which fields would a wrong value hurt more than the others? Who is the second person for those?

Testing on Pages a Person Has Read First

Decisions 4 and 5 share one working rule. Signisys’ working description of the test runs like this. Gather pages of the chosen document type that a person has already read and confirmed. Run them through, and compare each value with the confirmed one, noting beside each mismatch what caused it: a poor scan, an odd layout, or a check that was too strict. In addition, include a few poor scans on purpose. The authors of a 2019 research dataset of scanned forms describe its pages as noisy and widely varied in appearance, which they say makes understanding forms a challenging task.

Then settle the review rule from what the test shows. For each field, set the point at which a value goes to the reviewer, and change it as the results come in. A fixed number does not fit each firm; the team sets it from its own pages.

Going live is not the end. A named owner watches the results, keeps the corrections, and repeats the test when a sender moves to a new layout or a new kind of page arrives.

A good test on past pages gives no proof that the next page will be read correctly. The owner decides when the next test falls. A written sheet for readiness ahead of the first trial is in nine checks that mark one paperwork flow on dated evidence.

Intelligent Document Processing: Signs a Decision Was Rushed

Intelligent document processing goes wrong in recognisable ways, and a team can spot the early signs. The table sets each decision beside the sign that it was rushed and the person who notices first.

Each sign points back to one decision, which tells the team which one to reopen.

Two Bills of Lading, One Doubtful Container Number

The example below is made up, and no Signisys project lies behind it. A freight forwarder in the Gulf has a documentation clerk who keys the consignee, the container number and the number of packages from bills of lading. Overseas shippers send them as PDFs and photographs.

The team works through the five decisions. To begin with, it picks bills of lading and three fields. Then it writes a check for the container number: four letters, seven digits, and a number the operations team’s list already holds. This rule belongs to the example alone and is no standard. The senior clerk is the reviewer, and the operations coordinator is the stand-in. The test uses pages the clerks read last month.

During the test, a stamp crosses the container number on one page. As a result, the value fails the check and goes to the senior clerk. She reads the page beside the value, corrects it, and the team keeps the correction. Two bills from one shipper then show the same stamp, and the team adds a poor-quality sample to its test pages.

At go-live the senior clerk repeats the test when a shipper changes its layout.

A Document Type to Bring to a First Conversation

One document type, its fields and a handful of pages a person has already read are enough to start a conversation. A team with documents in many layouts, or with fields it cannot yet name, can still begin by listing them. If it helps, Signisys can go through that list with you and work out which of the five decisions is the next to settle.

Signisys works on document automation and IT automation; its document workflow automation solution and automation service pages describe the offer. The box below advertises a Signisys service and promises no outcome.

Intelligent Document Processing FAQ: Points Readers Raise