View all glossary terms

Core concepts

IDP Explained: The Framework That Ties Document Processing Into One Flow, Complete Guide

Korea Deep LearningPublished

IDP (Intelligent Document Processing) is an intelligent document processing framework designed to handle everything from document intake through form identification, recognition, item extraction, validation, and system integration as one unbroken flow.

Why IDP is needed

Simply attaching functions does not reduce throughput.

A common approach when introducing document automation is this: buy a recognition tool, buy an automation tool, and build a review screen separately. Each tool does its job. But staff working hours do not fall as much as expected.

The reason lies in the gaps between the tools. You download the recognition result and check it, reformat it and upload it again, and gather failed cases separately for reprocessing. Each step is automatic, but the work of connecting the steps is done by people. IDP is a view that eliminates these gaps. It asks not how well each function performs, but whether the work flows from intake to registration at once.

The exact definition of IDP: how it differs from a combination of functions

Three axes of difference between a function combination and IDP

First, the unit of design differs. A function combination thinks in terms of tools. IDP thinks in terms of the business process. It first draws where a document comes in, what states it goes through, and where it is registered, and then places functions on that map.

Second, the goal differs. The goal of a function combination is to raise the performance of each stage. The goal of IDP is to keep the processing flow unbroken. Where failed cases go and to whom exceptions are handed over are part of the design.

Third, the performance indicators differ. A function combination looks at accuracy per stage. IDP looks at the proportion of cases completed without human intervention relative to intake volume, and at total processing time per case. If accuracy is high at each stage but the stages do not connect, these indicators do not improve.

The two approaches are not opposed. If the performance of each function is low, nothing flows no matter how well they are connected. However, if you must choose an order, it is better to draw the flow first and then fill in the functions.

The five stages that make up IDP

Five stages that make up the flow

First, intake. Depending on the channel, such as electronic fax, branch scans, mobile upload, or email, the state of the file differs, so it is included in the design from this point.

Second, identification and splitting. When several documents come in as one file, they are split into individual documents and the form of each is identified.

Third, recognition and extraction. Values are pulled out according to item definitions that match the form.

Fourth, validation. Format, relationships, and agreement between documents are checked by rules, and cases that trip them are separated as exceptions.

Fifth, integration. Data that passes validation is registered in the document management system and business systems, and the processing history is kept.

Practical construction of IDP

Draw the flow chart first

The first deliverable is not a technical specification but a processing flow chart. Draw the route documents come in by, the inputs and outputs of each stage, where things go when they fail, and the points where people intervene.

With this diagram, you can see which stage is the bottleneck. In actual cases, the bottleneck was often in form identification and splitting rather than in recognition. If you choose tools without a flow chart, you miss this bottleneck.

Design the exception path as thoroughly as the main path

If only the normal processing path is designed and exceptions are added later, problems arise in operation. From the start, it must be determined where exception cases pile up, who sees them and processes them, and where they return afterward.

The return path is especially important. If values corrected by staff do not flow back through the process and get registered in the system, manual entry remains in the end.

IDP in a domestic environment

In Korean finance and the public sector, the whole flow often must be completed within the internal network. If any single stage requires an external call, that configuration cannot be adopted.

Processing history is also a requirement. Records must show which case was handled at which stage, how, and who corrected what, so that audits can be answered. Added to this is the handling of HWP (Hangul word processor) documents and faxed receipts.

Design around the flow, not the function

Having individual functions in place and reducing the work are different problems. Even if recognition, extraction, and validation each work well, if people connect the gaps between them, the processing time does not drop much.

So the unit of design is taken as the business process rather than the function. First draw the route one case takes from intake to registration in the system, then place the necessary functions along that route. Deciding in advance where people must intervene allows everything else to flow automatically.

Evaluation is also done at the flow level. Rather than looking at each function's accuracy separately, look at the proportion of cases that finish without human hands and the proportion that complete all the way to system registration.

Frequently asked questions

Can we use the existing tools as they are?

It is possible. However, if people are doing the work of connecting the tools, it is more effective to design from that point.

Do we have to build the whole thing at once?

No. It is safer to take one type of work, make it flow from intake to registration, and then expand to other work.

How do we know where the bottleneck is?

Measure the time spent at each stage. In an actual case, a considerable amount of time was spent on form identification and classification.

How should the effect be measured?

Look at the proportion of cases completed without human intervention relative to intake volume, and the total processing time per case. Accuracy by stage alone does not reveal this.

Can it be built in an air-gapped environment?

Yes. A configuration that runs the entire process from intake to integration on the internal network is possible.

Can we introduce the functions one at a time?

Yes, but the reduction effect is limited. If a section where people connect functions remains, the processing time does not drop much.

What should we use to evaluate it?

Look at the proportion processed without human intervention and the proportion of work completed all the way through. Accuracy by function alone does not show the actual change in throughput.

Tell us your work flow and intake channels, and we will organize the bottlenecks together.

Contact us