Banking is a document intense industry. Documents that are not always based on the same format, but need to be handled quickly, such as corporate action notifications. JP Morgan wanted to automate the handling of these document. Enter: RPA and AI. During the AI Summit by UiPath, Frank Chen of JP Morgan explained how non-tech users build and trained the model handling these documents.
‘Our use case of automation is specifically about the extraction of data attributes around corporate action announcements. It’s a difficult area for automating extraction since there are no standards in form and formats for these announcements’, says Frank Chen, head of intelligent automation APAC at JP Morgan. In the past JP Morgan used OCR tools and automation to extract information, but the constantly changing formats and need to process the data fast, made it less useful for this type of information.
Operations builds and trains the model
‘To automate document understanding for this type of information, we’ve used models from UiPath Studio to create bots to go to the websites and pull out the corporate action notifications. Then we used AI Center by UiPath for document understanding.’ This is the key thing for the success of this project, Chen says. ‘We could just pull someone from operation, a non-technical person, to build and train the model.’
Since corporate action notifications need to be processed quickly, not having to engage a data scientist for the models is a major benefit, says Chen: ‘If you need a data scientist to build a model, you need a much longer process.’ It’s true that the initial implementation took a bit longer: ‘But this was the first time for us to work with this platform. Now that we have set it up and understand how it works, we are seeing much quicker times.’
Gaining confidence and increase accuracy
The third component of this project was using Action Center by UiPath, a platform where robot and human meet each other. ‘In the action center critical elements can be reviewed by our users. And that also gives our business users the confidence that we have built what we’re expecting. It also allowed us to retrain the model for the elements that were not correct, which increased the accuracy.’
Moderator Amit Kumar, VP industry practice at UiPath, was wondering, since JP Morgan no longer needs a data scientist to train models, have the flood gates opened for other use cases? Is Chen planning any future automations? Chen says that the current use case is looking at structured data and their interest is leaning towards unstructured data. ‘We’re working together with UiPath to develop these models.’
Handling non-English languages
Chen: ‘A use case we’re looking at for example is invoice processes, particularly in the APAC region where there are non-English languages. A use case that’s just gone live extracting data from Chinese invoices. And here as well, it’s a massive game changer to not have to explain to a data scientist who might not understand the language and therefore have barriers to build the model.’
Kumar asks how JP Morgan started with RPA, since implementation, training and deployment might not always run as smoothly as one wants. Chen says that it helped to start small and to really understand what capabilities they were looking for. The small scope helped to really see if the solution met the requirements and to gain the trust of the users. ‘We could show that we can do it with this tool and that way we could bring people with us.’
The goal is not automation
But also, Chen says, it’s important that you really understand what you’re trying to automate. ‘The goal is not automation. Sometimes it’s even better to look at a process and stop the process all together. But you can also see: why are we using a particular format? Can’t we use a structured format where we don’t need document understanding? You need clarity before you embark on an automation project like this.’
According to Nitin Purwar, director, banking industry practice at UiPath, banking is an industry where lots of benefits can be gained because it is extremely document intensive. ‘Even when there is no large-scale standardization of this documents, or when the documents are either structure or unstructured.’ At the same time, Purwar says, the information from these documents is used in a lot of rule-based repetitive tasks. Ideal for RPA, so to speak.
Handling all data around mortgages
He shows various use cases, for example: document understanding mortgages. Not just to extract the right kind of data from all the documents that are handling in such a process, but also to intelligently match party, loan, property, title, and other details to ensure data are accurately recorded and reduce errors in the closing process. ‘There is a large document volume you can process with RPA and AI’, he says.
To gain the benefits, however, you first need to set the right expectations, Steve Tegeler, senior director solution engineering at UiPath. In a very practical session, he shows how setting the right expectations and mapping the current journey, helps you to set the right metrics. ‘For example: when an e-mail comes into the company, how much time does the person handling the e-mail need to process the document.’
A North Star to guide you
‘Let’s say it takes 10 minutes per e-mail and 20.000 hours per year if you do it manually and we’re completely automating the process, you’ll save that time. However, 100 per cent is not realistic, you might need validation by a human, but even if you do that, you’ll still save a lot of time. And over time, when you train the model, the results will be even better. By setting a North Star, you’ll know exactly when you’ve reached it.’
Photo: JPMorgan Chase & Co
Learn more about machine learning in key areas of banking here
You can watch a recording of the AI Summit here