Optical Character Recognition Within Financial Services

There is no doubt that the pandemic has significantly increased the pace of digital enterprise transformation and led many traditional financial service institutions to confront their reliance on analogue processes, outdated business models and rigid customer experiences. While institutions are turning to robotic process automation (RPA), artificial intelligence (AI) and machine learning (ML) technology to jumpstart their digital capabilities, technology such as optical character recognition (OCR) that has been popular since the early 1990s continues to remain an untapped opportunity for automation.

With improvements in optical character recognition offerings over several decades, the results are more accurate than ever and offer a solution to the pervasive domain of undigitized data and the error-prone process of manual data capturing that is particularly present in modern day financial services.

What is OCR?

OCR is a type of computer processing technology that is used to convert scanned images of handwritten, typed, or printed documents into a machine-readable format. Machine-readable implies that the data is in a structured format, in order for it to be processed by computers.

OCR technology leverages off robotic process automation, artificial intelligence and machine learning that allows for unstructured, semi-structured and structured sources of data, which would normally be manually captured into a database, to be captured and processed through an automated solution. OCR effectively reduces the time it takes to not only perform the capturing of data by digitising files within seconds, but also improves the accuracy of the data captured into systems and stored in databases through the elimination of human-user error.

How does OCR work?

OCR is versatile in that it can process multiple sources of data – PDF files, scanned documents of written or typed information, word documents and images can all be scanned and then processed using an OCR tool. [1] Pre-processing of the data source is required to best prepare for the extraction and processing of the required data in order to achieve robust data quality. Pre-processing will convert the document into black and white where the dark areas will be identified as characters and white areas will be considered the background. [2]

Once pre-processing is complete, the model will move on to the feature extraction process. The two main methods of feature extraction are ‘feature detection’ and ‘pattern recognition’. [3]

1. Feature detection focuses on applying rules that detect features of specific letters or numbers to recognise words and numbers in a document. Features of characters include the lines that are angled, lines that are curved, and crossed lines. For example, the letter A has features which include two diagonal lines and a crossed line in the middle.

2. Pattern recognition uses machine learning that is trained on many examples of text in different formats, which is then used to recognise these examples of text in documents.

Once feature extraction has been completed, post-processing is used to enhance the segmentation of characters. Post-processing makes use of natural language processes (NLP) to enrich the OCR process by governing where to separate or group letters, predict missing words from a sentence and enhance the overall accuracy of the output by grouping letters to words as well as establishing sentences where possible. Finally, the output is generated as a string of machine-readable characters.

OCR applications can be trained to memorise the extraction of words and numbers in specific positions of a data source through machine learning (ML). OCR applications are trained on a sample of similar data sources to create a pattern in order to learn what needs to be extracted. Following on from training the OCR tool, it will be able to extract the relevant data from sources, similar to those used in its training, and automatically populate a pre-designed data template that is linked to a database.    

Where CAN OCR be optimised in the Finance industry?

While there is increased pressure amongst banks to digitise and increase efficiencies, the rash introduction of advanced techniques and technology can lead to technical debt and unsustainable technological investments. Management amongst financial service institutions should understand the use cases for OCR applications and how best it can fit into their various finance, risk and customer processes and underlying data infrastructure. In this regard, we have consolidated the foremost use cases for OCR:

Know Your Client (KYC):

One of the most prominent use cases for OCR is improving customer onboarding for KYC operations. Prospective clients are required to fill out sign-up forms containing personal information and provide scanned documents that contain the same personal information, which must then be verified by a team of employees. The procedure of manually uploading data and ensuring quality checks is repeated for each new prospective client and becomes time-consuming for both the client and the employee. This often results in a duplication of the same work and can lead to unintended data mismatches. OCR systems are being introduced to automate this process, as well as being used to verify the data of the client through its AI capabilities. [4]

Paytm, India’s largest digital payments and financial services platform, is using an OCR application developed by Amazon Web services (AWS), called Amazon Textract, to deal with their manual and time-consuming user onboarding process. The AWS application uses OCR to instantly extract data from a variety of document types, which eliminates manual entry and reduces the authentication process of users from days to minutes. By using Amazon Textract, Paytm can extract user data with a very high degree of accuracy and far more efficiently than any manual process, as well as having the capability of integrating securely with other AWS products. [5]

Client Onboarding And Application Processing

Discovery Life, one of South Africa’s largest insurance companies, recently launched an OCR-based AI insurance quote system that allows brokers and clients to upload the required documentation over the internet and instantly receive a quote. Prior to this implementation, data would need to be manually extracted before it could be processed. The OCR quote system can automatically identify, extract and process the relevant text from these documents to generate quotes based on the information uploaded. [6]

As for retail banking, a leading financial services group in Asia, DBS Bank, is simplifying and reimagining their customer experience with ‘Quick Credit’. Quick Credit is an OCR-based application focused on simplifying and speeding up the loan application process for individual customers. Customers simply need to upload a set of photos containing the requested documentation and ID information. Quick Credit then extracts the relevant information and automatically populates the fields of the loan application that would usually be manually filled out by the customer. [7]

In highly competitive industries such as insurance and banking, reducing the time it takes to generate quotes and applications gives a company a significant competitive advantage, as well as providing customers an enhanced and efficient service.

Invoice Processing

Citibank have implemented an OCR solution to digitise and automate part of its trade business. Trade continues to rely on the use of physical documents that often exist in various formats and languages with various parties involved. These documents such as invoices, shipping confirmations and bills of lading are now read and processed through their OCR tool in order for specific financial and logistical data to be digitised, extracted and stored.  This has improved risk controls across the data extraction process and decreased the processing time which overall has enhanced the supply chain experience. [8]

Automated Financial Statement Processing

Professional Bank, one of the fastest-growing financial institutions in America, is optimising its business loan applications through Moody’s Analytics QUIQspread application. QUIQspread is an OCR-based technology that allows users to automate the extraction of data from company financial statements. A crucial aspect of the business loan application is their financial analysis reports – creditors must source and analyse key financial line items, as well as ratios measuring the business’s liquidity, profitability, and efficiency. By using QUIQspread, Professional Bank is able to digitise and extract the financial data from the financial statements into their predefined financial templates in minutes. The process improves its due diligence and audit report process by accurately and consistently extracting and storing the required client financial data, as well as mitigating the risk of human error by no longer needing employees to manually extract the data. [9]

What are the drawbacks of OCR?

OCR technology has proved itself as a useful tool in tackling many of the manual, unstructured data processes that continue to plague financial institutions. Improving data quality and data capturing efficiency across these processes highlighted above, not only saves costs and drives efficiencies but leads to a direct enhancement of the customer experience.

However, the drawbacks of OCR should not be ignored, and a certain level of oversight and review is required to ensure OCR technology is operating as intended. OCR models are only trained on samples provided, meaning if it were to consume a data source that was unlike what its training had prepared it for, results would be inaccurate.[10] An example would be if financial statements uploaded through an OCR application are in a language not part of the sample that was used to train the OCR application. The application then would not be able to extract the data correctly. This makes OCR susceptible to errors when faced with documents that differ from what was used to train it, including handwritten text and different languages. Additionally, OCR is negatively impacted by watermarks, improper alignment of text, as well as imperfections such as smudges.

Before implementing any OCR application into a business process, it is important to consider the underlying data infrastructure, including predefined data templates, as well as databases used to store the OCR processed data. Data templates should be designed and developed based on comprehensive and well-defined business requirements, to capture the required information for each OCR extraction. The template elements will then need to be integrated with the attributes of the associated database. Data quality rules, along with data quality oversight, should be developed to ensure that the OCR sourced results are reliable, accurate and complete.

How can Monocle assist?

At Monocle, we have over 20 years of experience in the financial services sector in the United Kingdom, Europe and across Southern Africa. Working closely with our banking and insurance clients, we have developed unique insight into each of our clients’ respective infrastructure and business processes. We are therefore well positioned to assist our clients in successfully integrating OCR solutions into existing manual processes and infrastructure by sourcing business and data requirements, designing and developing the OCR data templates, integrating the OCR data into the existing databases, as well as designing and developing the data quality solutions that support the OCR process.