Return to page

Keynote Session by Commbank with Sonal Surana






Sonal Surana, General Manager at Commonwealth Bank of Australia shares recent innovative ideas in this second segment of the CommBank team keynote.

Talking Points:



  • Sonal Surana, General Manager, Chief Data & Analytics Office, Commonwealth Bank of Australia
  • LinkedIn

Read the Full Transcript

Sonal Surana:


Good morning everyone. It's been a rollercoaster of a ride this first year of our partnership with H2O AI and the momentum continues to get even more exciting. We've heard from Matt about our AI ambition and how front and center it is for CBA strategy, and we've heard from Andrew on how we are mobilizing and delivering an AI culture across all levels of the organization. Now it's time for Dan and me to bring it to life for you with a few solid examples of recent innovative ideas turning into cool capabilities and starting to drive real differentiation and delight in what we are delivering for our colleagues, our customers, and our communities. So first up is document AI.


Document AI


Document AI is H2O's product, and we are really leveraging how we can reimagine and transform core processes and operations here at CBA. It is helping us unlock new dimensions of value as we stretch ourselves from structured data to unstructured data, from analyzing billions of transaction data points to now analyzing equally massive previously unexplored opportunity with documents, text and image data. From moving from automation to now intelligent automation. New technology is now enabling us to get the brains added to previously available arms, legs, and eyes of OCR and RPA solutions. Leading edge developments in the space of machine learning, deep learning, natural language processing and generation enable us to sort of unlock a lot of radical transformation in the way we previously thought of dealing with documents. And thanks to our strategic partnership with H2O, we've gotten early access to this document AI capability as they were rolling this out to the market. And in the short four months of time, we've stuck this up as a capability for the group where today we are having this scale and ability to process millions of documents per day on this platform. This really enables us to think about, holistically, the processes we can transform within CBA, leveraging this capability. Ignoring unstructured data and the treasure trove of insights that are available and hidden in it creates knowledge gaps and missed opportunities. And that is where we want to get focused to really unlock value for us.


Opportunity Landscape


So where do we start looking for opportunities? To give you a sense of scale of possibilities, CBA has over 16 million customers and with a few hundred pages of documents per customer that already creates hundreds of millions of documents where we can start to unlock value, starting from account origination and customer onboarding, related documents, to ID documents for customer verification, to loan applications and income verification documents for credit assessment to ongoing servicing requests. And add to that non-customer documents such as supplier contracts, invoices, purchase orders, just massive, the scale of opportunity that exists and the hidden insights that are waiting to be mined and productivity gains that are waiting to be unlocked. From a customer lens, let's zoom into the credit assessment example a bit more. Any loan or mortgage application requires numerous documents, be it bank statements, credit card statements, pay slips. All of these documents come from wearing different formats and templates depending on the source institution they come from.


Even if you take the examples of pay slips, they hold a wealth of information from gross salary to net salary to names and addresses of the employer and employee, salary period, pay range, tax details. Traditional OCR is not able to keep up and has limitations on wearing document templates and also the intelligence that can be applied to really extract relevant insights from that. So a lot of possibility as we look and zoom to unlock value from document AI there. Even as we look at non-customer focused lens, a lot can benefit the procurement function and the accounts payable and accounts receivable process. As we think about invoices and the vary invoice formats, applying technologies like document AI can really inI invoice processing, reduce operational costs massively, make us reduce payments 10 times faster and also create sort of a three-way automated match between purchase orders, invoices and receiving reports, driving, tremendous efficiencies in there. So that's just a few examples of where this technology can really unlock potential for us.


KYC Application


Let's now zoom into KYC and what we've been able to achieve there. Commonwealth Bank is one of the largest banks in Australia, and we have hundreds of thousands of customers getting on-boarded every year. To ensure compliance with risk policies and regulation, we of course, make sure there's a proper due diligence and every customer goes through the KYC process. Bank colleagues today spend a significant amount of time first extracting and second verifying information from customer ID documents such as passports, driver's licenses, birth certificates, et cetera. And as we look at this really posed wonderful area for us to sort of pilot document AI, document AI is able to automatically read and extract this information from customer ID documents, name details such as name, address, place of birth, and other critical information and store and structure it for us to really then leverage it downstream.


This really opens up possibilities on how we can automate and simplify, improve accuracy and productivity, and create delight for customers through faster onboarding and for colleagues through simplifying and automating core processes. Legacy OCR does not really hold up in scenarios of varying document types, especially photocopies are provided or faxes and emails come through of scanned ID documents. This creates poor resolution. Equally you have IDs coming in with various sizes, shapes, orientation, and that's where traditional OCR is not able to drive as the accuracy uplifts that is possible with this new technology that's now available to us. So we've made solid progress on this effort. H2O is really helping us bring a combination of models into this automation exercise, classification models that help us identify which sort of document is being supplied with passport, driver's licenses or others. Noise classification models that help us understand the quality of the scan document, noise reduction models that help us do the pre-processing and get the documents ready to then finally apply OCR and an advanced NLP, AI models to classify and manage that text.


So with some stats, every week we have a few thousand ID documents coming our way. 75% of them are passports, 25% driver's licenses and others. We have a few dozen colleagues who are manually doing the quality verification. And that is where we are beginning the automation. And we will have of course, the potential to sort of take it further upstream at the extraction stage itself, we've already achieved accuracies and automation of 50 to 85% on wearing document types. And this really is just the beginning of where we can sort of start to embed and relook and radically reimagine our day-to-day operations and make lives better for our colleagues and our customers.


Signature Detection & Verification


Moving from natural language processing to deep learning and what's possible there. Moving from structure data to text documents, and now starting to experiment with reading imagery of documents. With new deep learning capabilities that are available, that is where we are now. Also static get focused. So Dan Jermyn, CBS Chief Decision Scientist will be up here right after me sharing with you some more powerful examples, and he graciously allowed us to use him as a trial and demo case for this capability. So what you see here is we are picking up his signature from a document, trying to identify where exactly on the document it exists, whether it exists and where exactly it exists. Sometimes you might need to understand if a document is signed, other times you might want to compare it to a signature you have on file. So here it is matching, trying to match whether the document is a genuine one. A signature is a genuine one or a forgery. Don't worry, Dan, you're safe. Both the signatures are forgery here, we just graciously use his name, not a signature. But as you see this, it brings to life the possibilities of where this technology can take us into what we can do for KYC straight through processing and identifying and preventing scams and frauds.