What Document AX is: a complete guide to the shift that has AI perform document-driven work

Document AX is the application of AX to work that receives, reads and judges documents and registers the results in systems, changing the process people used to piece together so that AI performs it from start to finish.

Why Document AX became necessary

Documents have been digitised, but people still read them.

Much of the work in finance, insurance and the public sector still starts with documents. Applications and supporting documents arrive as faxes or scanned files, and staff open them, find the values they need and enter them into the system. Storing the files in a document management system does not reduce that work.

Even after document recognition technology was adopted, the situation did not change much. Checking recognition results, comparing documents against one another, judging conditions and registering the results were left to people. Document AX takes the whole of document-driven work as its target and designs it so that AI takes over these segments too.

Document work often has fixed forms and handling criteria that are written down. That makes it a good area for an organisation to start AX. Because volume can be counted case by case, it is also easy to compare before and after the shift.

What Document AX actually is: how it differs from document digitisation

Document digitisation and Document AX across three axes

First, the scope differs. Document digitisation ends at scanning paper documents so they can be stored and searched as files. Document AX reads and structures what is inside the files and connects it to business judgment and system registration.

Second, the output differs. The output of document digitisation is an image or PDF file. The output of Document AX is values entered in business systems, judgment results, and a record of the grounds for them.

Third, the role of people differs. After digitisation, staff still open files, read them and transcribe values. In Document AX, AI finishes normal cases, and staff take charge of checking exceptions and managing criteria.

Document digitisation is the starting point for Document AX. Scanned files and document management systems that have already built up become what AI reads, so they are assets to use as they are rather than things that need replacing.

The five stages that make up Document AX

Five stages that carry document work through to the end

First, reading. Documents that mix print, handwriting, tables, seals and checkboxes are recognised. Faxes, low-quality scans and skewed documents are included.

Second, structuring. Titles, sections, tables and document hierarchy are restored and turned into data AI can use. Tables have to be converted with the relationships between rows and columns intact, or values are lost in later stages.

Third, extraction. The fields the work needs, such as business registration number or representative, are found and organised as field and value pairs.

Fourth, judgment. Whether information agrees across documents, whether every required document is present, and whether the conditions are met are checked against the business rules. If the same business registration number is written differently on the application and on the business registration certificate, this stage catches it.

Fifth, execution. Results that have been judged are registered in the document management system, automation tools and the core system, and the processing history is recorded. Without this stage, staff end up entering the results again on another screen.

There is no need to change all five stages at once. If a recognition tool is already in place, the judgment and execution segments can be added first, and for work whose forms change often, the structuring and extraction criteria can be put in order first. The sequence is set to fit the current state.

How Document AX is applied in practice

Start with the work where documents pile up most

Begin by organising which forms arrive in each department and in what volume, and how many types of document one case requires. The more document types and repeated comparisons a process involves, the larger the effect of the shift. At the same stage, check whether several documents arrive merged into one file and whether page order gets mixed up.

In a financial-sector case, the document groups held were listed first, and those with high volume and stable forms were then selected for evaluation to confirm classification and extraction performance.

Look at whether the work finished, not the recognition rate

Even with high character recognition accuracy, if people re-enter values at the registration stage, the shift stops halfway. The results of Document AX are measured by the share of received cases that finished within the flow through to system registration, and by handling time per case.

The classification and extraction performance confirmed at the adoption stage describes the front segment of the flow. Even when those figures clear their targets, the share of cases that finished through registration has to be measured separately to see how far the shift has come.

Design the exception path first

First decide on which screen documents that are hard to read, cases where information does not agree across documents, and cases the rules cannot decide will appear, and who checks what.

Only once the exception path is set can the remaining normal cases be sent through automatically with confidence. Errors confirmed in exception cases are used again as grounds for revising the extraction criteria and judgment rules.

Document AX in the Korean environment

Korean business documents often arrive as a mix of Korean word processor files, government forms and faxed submissions, with seals and handwriting together on the page. Tools built around documents from other countries may perform differently under these conditions, so they have to be validated on the organisation's own documents first.

In finance and the public sector, original documents often cannot be sent outside the organisation, so a configuration that can be installed internally and a de-identification procedure for personal information are prepared together.

Frequently asked questions