Tech Updates

Published on Jun 28, 2022

How to save 500 person-days per year using Intelligent Document Processing

Quick call-back: Cloudflight AI Patterns

We have observed that many already see Artificial Intelligence as a key technology. However, the actual use in companies or the implementation projects currently started or planned do not yet reflect this. Reasons often cited are budgets or lack of time, which is a one-sided view and does not take into account the benefits and possibilities.

The Cloudflight AI Patterns help to systematically visualize, process and evaluate the potential of AI in specific companies. A single pattern is a deployment pattern of a specific AI technology that solves a relatively fine granular problem, such as recognizing the type of a document (order, invoice, delivery bill, etc…). The benefit then results from the overarching use case and the contribution that is made there as a building block.


Case study: Intelligent Document Processing

To make the procedure more descriptive, we will introduce a concrete pattern and its application in a specific enterprise context in more detail. When selecting the pattern, we chose “Intelligent Document Processing” and the corresponding sub-patterns. The basis for this was the broad applicability of this pattern, since documents are ubiquitous and accompany virtually all business processes.

As a company, let’s imagine a large medium-sized production company with several locations. Imagine that 1000 deliveries are delivered to various locations every day.

The simplified process behind this is that the respective delivery bill is first digitalized. This is done by stamping delivery bills after goods receipt and inspection, collecting them in a filing system and taking them from there to a central location for each branch. Here, the bills are scanned and saved in an inbox folder for each branch.

Based on the now digital delivery bills, a manual assignment of the delivery bill to the corresponding purchase order takes place in the software system. For this purpose, each document is opened, and the data is compared with the open orders using simple search criteria such as order number, supplier or goods description. The exact workflow often depends on the experience of the individuals to narrow down the searched order well enough already with as few typed characters as possible, so that the corresponding match is easily recognizable from the remaining list.

As a result of this comparison, items and quantities are checked for completeness and, in the case of partial deliveries, the respective quantities are entered. For later traceability, the scanned delivery bill is stored with the (partial) delivery created in this way and saved centrally.

The data structured in this way is passed on to various third-party systems (e.g. merchandise management) or triggers further actions (e.g. complaints management). Deliveries are not always evenly distributed over time, so that there is often a backlog of cases to process. This backlog dissipates over time – either through corresponding overtime or days with a lower number of new cases.

In the same company, an average of 3 applications arrive daily via the corresponding section of the website. These usually consist of a short cover letter and a resume – both in pdf format. These documents go through a recruiter and are matched to open positions or departments.

When quickly assigning people to possible positions in the company, attention is paid to certain keywords such as various activities, technologies or certifications or other qualifications.



The AI Pattern – Intelligent Document Processing

For example, here we will consider the AI Pattern “Intelligent Document Processing” in the context of the processes briefly outlined above.

Name: Intelligent Document Processing.

Context: Documents contain any information in unstructured form. However, in order to be able to process this information automatically, IT systems need information as data that follows a predefined structure. Intelligent Document Processing can automatically extract and structure specific information (entities) even from heterogeneous documents.


  • Optical character recognition (OCR)
  • Entity extraction (Named Entity Recognition)
  • Text classification


  • Information must be typed from a piece of paper into a software system
  • Information is searched from a digital document (email, pdf, etc…) and transferred via copy and paste to another software system
  • Structured data is derived from continuous text


Sub-Pattern – Document Classification

Name: Document classification

Context: Processes are accompanied by documents of different types (various types of contracts, invoices, receipts, technical documentation, etc.). Often, however, this type is not apparent from file formats or names. Through classification, the type of a document can be determined or verified purely on the basis of its content.


  • For faster access to certain information, the type of a document must be apparent
  • To check completeness along a process, the presence of certain document types must be confirmed
  • The type of a document must be known in order to be able to derive or specify the further processing steps from it.


Sub pattern – Text Classification

Name: Text classification

Context: Individual text fragments can often be assigned to classes. These classes can range from the same topics covered (e.g., news articles) or even synonyms to subject specifics (e.g., assigning posting texts to account classes).


  • Text passages or blocks must be assigned to a topic
  • Free text formulations are harmonized by recoding them into predefined catalog entries


Sub-pattern entity extraction

Name: Named Entity Extraction.

Context: The meaning of texts is often concentrated in a few words, while the remaining ones, for example, are filler words without significant meaning, purely for the better flow of the text. In order to be able to process texts automatically, it is necessary to extract data points from free text. These can be people, company names, place or time information, etc.


  • Text passages or documents are reduced to a uniform table format
  • Predefined data points are extracted from free text


Use case

In line with our approach, we are assessing the relevance of the AI pattern presented to our various business units.



Corporate division – Logistics

In logistics, manual efforts arise at several points, as described above. We basically detect a modality break in the process. We start with the printed delivery bill, which is still stamped before the scan makes the switch to digital processing.

The first potential for efficiency gains is that delivery bills are not scanned separately, but at certain intervals in a batch. The subsequent separation of the individual documents into separate files can be handled by the AI. This can be solved, among other things, through the pattern of Document Classification, by considering each page as a document and classifying whether it is the “start page” of a document or a subsequent page.

Downstream document classification can ensure that the documents are indeed all delivery bills and automatically sort out other documents and submit them for further manual processing.

The next step is to assign a delivery to a purchase order. This is where the entity extraction pattern comes into play. After OCR, relevant entities such as supplier, order number or delivery items are extracted. This usually allows fully automatic assignment to purchase orders and, in simple cases, also checks for quantities or general completeness.

The extraction of entities also enables further minor improvements of the automated document processing. Simple examples here are the consistent naming of files according to order number and document type or the transfer of metadata to third-party systems.

We note the identified potential applications and investigate the next area of business.


Corporate division – Personnel

In human resources, we are not dealing with delivery bills but with application documents as basic documents.

To support all further work, an upstream document classification can also sort files according to cover letters, resumes, certificates, etc. and name them consistently or store a corresponding tool in a suitable place.

The assignment to departments, and in the case of recurring job advertisements, also to vacancies, can be automated in a similar way with a text classification. Likewise, the presence of certain types of information such as education, required certifications, and existing professional experience can be checked.



In following the AI Patterns approach, we now evaluate each use case or AI deployment potential on its own. In the sequel, however, we will now contrast the two environments described above for a more striking example.

First, we evaluate the automation potential through a use case. In the case of logistics, we assume 1,000 cases per working day, as described above. For every minute that the process for a single delivery can be made more efficient, this scales to approximately 250,000 minutes or over 500 person-days per year. This is an enormous amount of work – and means that this time is available to deal more intensively with special cases such as complaints.

In the case of applications, we estimate the efficiency gain using the same formula and arrive at 750 minutes or 12.5 hours per year.

As an aspect of the risk and effort assessment for the implementation of the two use cases in a software solution, we look at the data basis on which we can set up machine learning procedures. In the logistics area, we assume a quarter of a million documents per year for which the historical extracted and relevant data can be exported from the existing software. This eliminates the need for time-consuming manual annotation of possible training data.

This is contrasted with a limited number of suppliers, recurring products supplied, etc. We are talking about a manageable amount of degrees of freedom and therefore moderate complexity and risk.

The situation is different for the application in human resources. The number of cases is significantly smaller, which means that it takes a multiple of time to collect the same number of documents. In addition, each applicant is an individual, so each resume is different, with different structure, level of detail, type of wording and choice of terms, etc. There are therefore significantly more degrees of freedom and a correspondingly higher level of complexity as far as software-supported automation of the work processes is concerned.

In addition to these two aspects – the benefits and the data basis as an aspect of the development effort and risk – a number of other evaluation factors must also be taken into account. Key influencing factors here can be available open source components, models or data, but also aspects such as the acceptance of AI solutions and the willingness of the involved people to also change their workflows.



Practical example

We have successfully implemented and put into production the AI pattern “Intelligent Document Processing” for the Austrian social insurance institutions on behalf of the central IT services of the Sozialversicherung GmbH (see also:

The field of application there is the processing of reimbursements for visits to elective physicians, which patients pay for themselves in advance. Analogous to the logistics environment described above, documents are also differentiated by type, the relevant information (entities) is extracted, billed services are classified on the basis of the service catalogs of the individual insurance carriers, and so on.

This system currently processes around 4,000 documents per day and is still being rolled out to additional units. Even with a relatively small reduction in workload per case, this application thus pays for itself through scaling. In fact, with this technology support, the insurance carriers have succeeded in dissolving the processing backlog that had built up and in offering the policyholders a better, because much faster, service.

The Intelligent Document Processing pattern offers such great potential that we have designed and developed a reusable solution “Picklist” (See also: on which we can efficiently build individual solutions.



In our last article, we presented the Cloudflight AI Patterns as a structured consulting approach around the introduction or expansion of AI in companies. Building on this, we have now looked at a concrete pattern based on two possible deployment scenarios for better illustration. The natural first step is the unbiased identification of all possible use cases. Only in a next step the potentials are surveyed based on different aspects. The reason for this sequence is not to discard topics early on due to their unfavorable cost/benefit ratio. In the overall picture, they can nevertheless become profitable again due to synergy effects between similarly positioned applications.

We are happy to accompany and support you along this path with our experience from numerous AI introduction and implementation projects.