Automating the Account Opening Process Demo

I’m going to spend some time just on one slide before pivoting into the demo to position how Hyperscience works as opposed to conventional technology that turns documents into data. At Hyperscience, we have a philosophy called human-centered automation. We are fundamentally a machine learning and artificial intelligence platform. We recognize it’s not perfect science, and at times we need to ensure that machines and humans are working together to deliver an accurate outcome.

Everything begins with human-readable data on the left-hand side. We need to get that into Hyperscience through input connectors. This could be via Hyperscience’s API, scanners, RPA, or other integrations we have out of the box. The first thing the machine needs to do is classify the document. What is this document I see in front of me? Is it an invoice, an application form, or a supporting document like a bank statement or driver’s license?

At any point, the machine may not be confident based upon your target accuracy to classify that document. In which case, it puts its hand up and asks a human to intervene. This is the notion of human-in-the-loop. The human will go in and help the machine only on that task, identifying it as a driver’s license, for example. Then we move on to the identification stage: knowing what the document is, where is the interesting data located on that page?

For structured forms like application forms that never change, the machine would do 100% of the work. But with semi-structured and unstructured documents, data moves around, so we need to teach the machine to locate that data. If the machine is not confident enough to find a particular field against the target accuracy, it asks for a human to identify that field on its behalf. Once we know where the data is, we transcribe it. Whether handwriting or machine print, it doesn’t matter to Hyperscience. We’ve been trained upon millions of different data points and apply natural language processing to read with context. We can ascertain what an email address is, a phone number, or a social security number. If there is varying quality at the document or writing level, again the machine asks a human to transcribe that on its behalf.

Finally, we can validate this interesting data against local business rules, internal databases, or enrichment through third-party APIs. Hyperscience can surface logic to humans to approve or reject a claim, for example. Once done, we have a completed submission and can output that to your downstream process via API, RPA, etc.

You may wonder how the machine learns over time. We use QA, or quality assurance. Think of it as a teacher marking a student’s homework. The human marks the machine’s work. If the human agrees with the machine, the machine’s confidence grows, making it more likely to automate that task in the future. Conversely, if the machine got it wrong, the human penalizes it, lowering confidence. This ethical use of AI keeps the machine honest and ensures we maintain target accuracy whilst increasing levels of automation.

I’m going to pivot to the demo. I’m going to play through a scenario around account opening. We have a structured form and need to provide supporting documentation. As a customer looking to open a savings account, my local branch is shut down, so I have to print out the form, fill it in by hand, and send a copy of a recent utility bill and driver’s license. I’m unhappy with the experience, but I send it. The bank has been set up with Hyperscience to listen to an email inbox and ingest that email body and attachments.

Here is an example of the structured form sent in. It’s upside down, has a slight skew, mistakes, and the writing has been squashed in. I’ve even added a pair of glasses for effect. Not necessarily compliant or conducive to technology solutions. The supporting documents are a driver’s license and a utility bill. The utility bill and license are going to be checked for consistency in a faux KYC process. We check the name and address match across the two documents. We also check that the statement date is within the last 90 days. In this case, the statement is from last year, so that’s going to kick off an exception process.

Let’s move into the platform. It is straightforward to set up new use cases. You may be used to coding in different layouts and coordinates. With Hyperscience, we can upload the blank copy of the form, drag and drop over where the interesting data is located, and give that a name. We have a number of out-of-the-box data types with natural language understanding, but equally, you can teach the machine your own data types, like a specific branch code. Once the form is set up, we describe how to process this effectively from left to right, like a digital assembly line. First, we specify our input, such as the email listener.

Here is a flow through Hyperscience. It can be as simple or complex as you want. Here we do classification, identification, transcription, and then enrichment and validation. We look into an external API to ensure customer data is correct, check the document has been completed properly, read the email searching for keywords and sentiment analysis, and provide the ID verification check between the license and utility bill. Then we complete that and output downstream.

Let’s dive into the submission. We have a new submission with three documents in a status of manual identification. The documents have been split out and correctly classified: application form, driver’s license, utility bill, and the email body treated as unmatched. We moved to identification. The machine isn’t confident where the interesting information is on the utility bill. I assume the role of someone in the account opening team and help the machine. I drag and drop over the account address, and that’s all the help the machine needs.

Next, we check transcription. The application form was handwriting and upside down, so there might be a scenario where the machine is unable to read the writing to the required accuracy. While it’s doing that, we are also transcribing data within the license and utility bill and invoking a full read of the email body. The machine surfaces up manual transcription tasks. It zooms me directly into the field the machine needs help with. This is my handwriting, but I can understand why the machine might be confused where the ‘I’ and ‘R’ look like an ‘N’. I type the value on the machine’s behalf. Everything else within that document, which is now right-sized, the machine has transcribed with confidence.

Next, we ensure these forms are filled out correctly. I’ve contrived a scenario where I created an error in the application form that doesn’t meet local business logic. We surface a task called flexible extraction. Hyperscience tells me there is an error: multiple checkboxes have been selected in title fields. I can overwrite the machine’s view and say actually ‘Mr’ wasn’t checked, but ‘Mrs’ is correct.

Now we go down the flow doing the sentiment analysis and ID & V check. Recall the license and utility bill address match, but the utility bill date exceeds the 90-day limit. We’ve set up a faux contact center function which will email and place an outbound phone call to the applicant to say there is an error. We also send a flag downstream to an escalation team. Hyperscience can take an output and inject that into a downstream process. As the applicant, I’ve received an email and phone call, enabling me to reach customer outcomes much quicker.

Equally, we flagged downstream that the bill validation check failed and there were negative sentiment alerts. I can review the email body and see the negative sentiment, giving me as a Customer Service Agent information at my fingertips. Finally, looking at the forms themselves, the application form began life upside down with glasses on it. It has now been rotated and right-sized thanks to the blank copy we hold up to the submitted document. We can see an example where I’ve handwritten and squashed things inside the box, with numbers right on the line. The machine has no problems identifying that with handwriting. We can see the hairdresser field I helped the machine with from an audit trail perspective, and things like email addresses and National Insurance numbers have been pulled in correctly.