Automating the Merchant Onboarding Process Demo

I will begin by speaking about the space we are in and specifying the problem statement. Put simply, the problem statement is converting human-readable data into machine-readable data, ensuring accuracy and the automation required eventually. We want to achieve use cases that are automated in the IDP space and are also viable for the business, along with the technical aspects.

Currently, in the real world, if we look at the human-readable data available which is not easily machine-readable, 80% of all documents exist in a human-readable format. These could be scanned documents, images, scanned images, mobile pictures, and so on. Existing technology relies primarily on OCR-based technologies. Some are traditional, some are AI-enabled, but they have limitations regarding context, levels of accuracy, the amount of integration and maintenance required, and most importantly, the level of automation they can achieve. This makes for limited applicability to use cases in the real world.

This is the space where Hyperscience excels. Hyperscience uses all the capabilities of the machine and the intelligence of the human being to work together so documents can be processed with the required levels of accuracy and automation. We refer to it as human-centric automation. The first step is to understand the type of document we are dealing with, referred to as Classification. The next step is to understand where the different elements of importance on that document are. Documents are of different types: handwritten, printed, fixed formats, varying formats, or completely unstructured. Hyperscience uses AI modeling to create models where prediction or identification of detailed elements is done.

Moving on, once the element is identified, we proceed to the transcription of the actual data element. Hyperscience implements many out-of-the-box models for data types because data exists in different formats, such as dates, addresses, names, and telephone numbers. There could be business-specific data elements as well. Wherever the engine feels the need to request human assistance, it does so by using a confidence threshold. Once the human provides input, this feeds into the training for the model, allowing for constant learning.

While most IDP solutions might stop here, the Hyperscience platform goes beyond. The transcribed data can be enriched, transformed, and manipulated so it is presented in a format required for downstream systems. Along with this, Hyperscience has the capability to have QA tasks used to improve machine performance per feedback. The platform includes the Flow Studio, where you can create business processes to build end-to-end business flows. Supervision is available where required, creating supervision tasks. Case collation exists to create the concept of a case or transaction, defining which set of documents goes together. The engine also has robust reporting tools for accuracy, usage, and performance. Furthermore, the platform integrates easily with external systems like Salesforce, Pega, AWS, Bizagi, and UiPath through out-of-the-box connectors.

We will now move into the use case for this session. The process is in the banking space where a merchant or entity wants to get onboarded with a bank. We are concentrating on the pre-onboarding and onboarding side, where merchant documents need due diligence checks performed on them.

The platform can be used to create a process which traditionally is done manually with eyeball checks. Documents enter the system through a variety of ways: scanned inputs, mobile pics, or uploads. The checks performed include making sure all documents are available as part of the transaction, ensuring information matches across documents (like ID and address information), and performing due diligence checks. For example, doing a nature of business check, identifying risk profile validation by integrating to external systems, or identifying promotional growth opportunities for cross-selling. The Hyperscience Platform can perform all these checks in an automated fashion along with reading and transcribing the data elements, then providing a supervision area where a human can review the recommendations and decide to proceed.

Here we see a way to manually upload a set of documents into the platform. I will upload a set of multiple-page scanned PDF documents, which have different sets of information along with application forms, ID forms, and registration documents. Once processing starts, the platform performs classification, identification, and transcription. For this demo, the documents were already trained, which does not take much time using the Hyperscience Platform. In many structured document cases, we can use just one blank document to train the platform.

Once classification is complete, the platform identifies each document on a page-by-page basis to understand where the different elements are. Then it moves to the transcription phase. In this packet, we have the application form, the ID document (which in the Indian context is an Aadhaar card), and a government-generated registration document. The checks performed include the document completeness check, ID check, and address checks for the KYC documents.

The engine has completed the job and is requesting the human to supervise the output. When we look at the output, the different validations and recommendations are given. For example, success on the document validation tells us required documents were found. There was a promo code validation applied which resulted in success. The address on the application form (which is handwritten) and the address available on supporting documents has been validated. There is an industry risk profile validation done based on the nature of business identified on the document, and a risk profile done based on external data availability where the risk rating of the industry code is mentioned.

The individual supervisor can come in and complete the task based on the information provided and transcribed by the engine. We can have this data available in a JSON format for downstream processing. The entire data set, including file names, pages, elements, and validations, is available. You can also use the UI to examine how different elements were transcribed, such as the organization name, address, and Aadhaar number. For example, this registration certificate issued by the Indian government does not follow a single format; there are variations. Similar to this, we have the ID document where the name, year of birth, gender, and Aadhaar number are available.

For this demo, we demonstrated how the platform can be used to perform an end-to-end business use case. By automating a use case like this, there are many business benefits. Many of these checks in banks are performed at the branch level; a process like this allows checks to be done at a centralized location. Processes are streamlined so any change can be easily implemented rather than conducting training sessions for each branch. On documents like this, the platform is able to achieve 100% automation.