Insurance Claims Processing Demo

In this short video we are going to see how Hyperscience can help insurance agents process travel insurance claims in particular. We will see how the Hyperscience Platform can process claims documents of varying different structures, both with very high levels of accuracy as well as automation. Ultimately, we are looking to help reduce the time it takes to process these claims without sacrificing the accuracy of the data that is being sent downstream.

In this demo, we are going to see a customer email. In their claim, they are going to send an email to their insurance provider along with all of the different claim documents that are required. Hyperscience will listen into that mailbox, automatically ingest the email body as well as each of those attachments. The machine will then classify each document, extract the data out of each of those documents, and then apply some custom business logic to validate whether the claim is actually valid or not according to the insurer’s particular policy conditions.

The insurance agent will then be able to review all of the information and make a decision about whether to actually accept the claim or not. This data will then be passed to the downstream claims systems whilst we’re being very confident that the data is accurate and complete.

So here we have an email ready to send to our insurance provider. Our customer, Sarah, has written in the email, included her policy number in the body of the email, and attached several documents here to support her claim. The first document is a structured handwritten claim form. It’s a photo of the document, not very straight, with messy handwriting across the page. It’s a two-page form. The second one is a completely unstructured document: a handwritten note from a doctor providing some details about the medical condition which caused the cancellation of the trip. The third document is a screenshot of the customer’s cancellation invoice for the trip that was canceled.

I am going to send in this email and Hyperscience is going to listen into that email box and pick up that email along with all of the attachments. If I open up the Hyperscience UI, we can see that we have a new submission which is currently processing. The machine has taken the body of the email and each of the attachments and is classifying the different documents that we are interested in. It has classified that travel claim form and the cancellation invoice. It also has the additional documents attached that are being processed.

Now that the machine has classified the documents, the next task is actually transcribing the data out of each of those fields within the documents. We have a manual task waiting for an insurance agent to complete here. This is a transcription task. This means that there is at least one field on the page where the machine isn’t confident enough that it can hit our accuracy target. In this example, we have an accuracy target of 99%, and instead of potentially sending an error downstream, the machine is asking a human to have a look instead.

In this interface, an insurance agent can come and complete any manual tasks. The machine is just highlighting the one field that it needs an agent to complete. I don’t have to scroll through the document trying to find the field. All I need to do is enter that date, 20th of June, 2023, and submit that field. That was actually the only field on the page that the machine needed any help with. The rest of these fields, the machine is very confident that it can transcribe and hit our 99% accuracy target.

Now that the machine has all of the data transcribed, it can run a number of validations and the results will be shown to the insurance agent so that they can decide whether to approve or reject the claim. I have a supervision task available to me now. I have a new case of documents created here, and I have each of the documents assigned to that case. On the right-hand side, we are seeing the results of the validations.

The first one here is actually a warning, telling us that there appears to be a mismatch between the policy numbers given across the different documents. The second warning is about the start date of the policy. According to the policy terms, the policy had to be active before any symptoms of the condition began. We’ve done a database lookup to the policy management database, saying that the policy didn’t begin until the 31st of March. However, on the claim form, the customer told us that the symptoms began back in February. Based on this, the machine is recommending that we reject this claim. It is still up to a human agent to decide whether to actually reject it or not.

The last results of the validations are the results of this keyword search. We had our note from the doctor, and what the machine has done is transcribed all of the data from that note and searched through it to see if any keywords we are interested in are present. We’ve asked the machine to look for some particular medical language. Based on those two results—the warning about the policy number and the mismatch between when the policy was active as well as the start dates of the symptoms—I’m gonna go ahead and say that yes, actually let’s reject this claim.

This is now the claim complete within Hyperscience. We can be confident that all of the data has been accurately captured to be sent downstream. We can take a look at what data the machine has actually extracted. If we look at our claim form, we can see individual fields like name, address, and email address have been extracted correctly. The way these extractions work is based on the type of data we have told the machine to expect. We told the machine to expect an email address, and instead of trying to read character by character, it looks at the context of the entire field. Even though part of it is crossed out and bits fall outside the box, the machine is still able to accurately capture that.

If we continue down, we can see some messy fields here which have still been transcribed accurately. Even though this has bits crossed out and uses the abbreviation of October, the machine has correctly identified the date. We can see longer paragraphs of text as well, detailing the reason for cancellation and diagnosis. If we look at the cancellation invoice, this is a semi-structured document. We know exactly what data we want to transcribe, but the format varies. The machine has identified where the data is and transcribed it, including line items from the table. This document didn’t require any human help; the machine was confident it could accurately transcribe all data at a minimum of our 99% accuracy target.

Hopefully, over the course of this video, you’ve seen how Hyperscience can help insurance companies process travel claims with very high levels of accuracy in a very short period of time. We’ve seen how Hyperscience can ingest customer emails, classify document types, extract required data, and apply business rules and validations before sending complete data downstream. Thanks very much.