AI in Insurance: Resolution Life's AI Journey with Rajesh Malla
Rajesh Malla, Head of Data Engineering - Data Platforms COE at Resolution Life insurance takes the stage at H2O World Sydney 2022 to discuss AI transformation within the insurance industry. Resolution Life is the largest life insurer in Australasia. Malla discusses the use of H2O Driverless AI to predict claim triage and other insurance AI use cases.
Rajesh Malla, Head of Data Engineering, Resolution Life
Read the Full Transcript
Who is Resolution Life?
I would like to discuss who we are to start with before, I open up with the business case as well as our H2O journey. So we are the largest life insurer in Australasia, we manage about $27 billion of assets, within Australasia. And we are part of a larger group called Resolution Life Group which was founded in 2003. So, we operate across various continents as you can see. And our focus is to become a number one insurer across the globe by 2024, right? So we are customer focused and a data driven organization. So we would like to place all our decisions on data, and our focus is to get a right outcome, on humanitarian grounds, to our customers. So today I'm going to talk about our journey with H2O. We recently started our journey with H2O about six months ago. And I'm going to discuss a business case that we have implemented H2O on and the benefits that we see today.
What Problems is Resolution Life Facing?
So the business, as an organization and as the largest insurer, since 2019 we have paid approximately 1.1 billion in claims since 2019. So the claims processing is largest and most complex in processing, which has its own problems, right? When a claim is lodged, we would like to assess the claim at the right time and provide a customer with the right outcome. So there are quite a few problem statements in assessing a claim. All our case managers are really working hard for a right outcome for our customer. Then again, right? Because of the complex nature of the claims processing, they may need to look at historical data that is well and beyond their limits to understand the claims, understand the customer claims, understand the demographic claims 360 degrees of a claim that has been submitted and whatnot.
So, there are quite a few complexities that we have listed on that. So that could result in not a right outcome for our customer. So we were looking at how to solve this issue and how to reduce our capital as well, because when we spend more time on processing a claim, that means it's an expense to the organization as well as not a right outcome for our client. We want our customers to be paid as quickly as possible and return to work for them to be successful in their life. So our journey with H2O, as I mentioned, started six months ago. So before processing any of our claims, there are multiple steps before we process a claim, right? So the first step is to identify the segment of that particular claim.
Using H2O to Segment Insurance Claims
So that means, is it an easy claim for us to move forward? Do I really need to spend any time on it? Or is it a complex claim where we really need to talk to our customers to understand their situation and whatnot? So was there a claim made before? Is it an extension to the claim and whatnot? So the segmentation of the claim makes a critical part of the claims journey. So what we have done is using H2O, we have implemented a claims triage model. So what it does is eventually it segments out the claim soon after a claim is launched. So previously when a claim was submitted, a claims manager needed to be assigned to the claim, open up the claim, look into the claim manually, spend about a day or two, or a week in some cases, to identify the segment of that particular claim before we can start processing the claim. So we made a big leap from there. So by implementing Driverless AI, now within 15 seconds the claim is actually lodged, and we can segment the claim into a bucket. That means we know the outcome of the claim within 15 seconds since the claim has been launched. And this is possible through H2O Driverless AI, right? What does it mean? It means a lot, right? It reduces the capital. As I said, that means profit to the organization and a right outcome for our customers.
Snowflake Implementation on Azure and H2O
How we have implemented it. We use state of art technology at Resolution Life. We are in an organization where our entire platform is on cloud. So we don't have any on-premises systems. We are a fully cloud organization and we are running Snowflake on Azure and H2O, some of the state of the art technologies. So, soon after a claim is submitted through our claims application, an API gets triggered onto our Snowflake environment, and a module model is actually deployed. A module H2O model is actually deployed onto our Snowflake environment, where the model is constantly running on the Snowflake environment, giving a predictive outcome within 15 seconds. So basically the outcome is sent back across to CMS, which is our, which is our Claims Management Solution, and also the outcome is actually, stored in our container storage like Blob Storage where we can use it for our future purposes. So how do we handle H2O? So, as you guys know, H2O is heavy on development and light on Productionisation. So that means all the development is done outside training the model, making sure the model is good. All that is done outside and once the model is ready, we deploy the model across onto the Snowflake environment to utilize the compute power of Snowflake.
How H2O is more Accurate than In-House Claims Models
Okay? I would like to show you this quite interesting slide. Maybe Sri will like this slide as well. So, we were using this claims triage model in-house build previously before we got across onto H2O, right? We spent quite a bit of our time and the accuracy pre-H2O was 71%, where when we moved across onto H2O, the straight accuracy was 77%. That means 77% of our claims we are able to identify. We were able to segment them correctly. So that's one of the biggest improvements, that's about a 6% improvement there. And then the estimated value, these estimated values, guys, don't quote me on these values, these are just there as indicator numbers. So these numbers can go anywhere from, if there are 3 million there. The benefit that you see is 6 million. So it just talks about the ratio of the benefit that we are getting. So previously, all the in-house build, if it is giving us a value of 720K soon after we've implemented H2O, it is 1.35 million. So that number is there, just to show you the ratio, one is two. The most important thing, which is not in the slide, which I actually wanted to talk about, is the time taken to implement this, right?
H2O is Four Times More Efficient than In-House Models
So the time taken to implement this particular model in-house versus H2O is one to four, so one H2O is four in-house models. So we need to have our own data scientists working on it, right? Where H2O is going to give us quite a few models for us to choose and predict from. So what are the benefits that as an organization we achieved by implementing such a model? That 15 seconds outcome, the right utilization of our case managers, right? Reducing the resources and focusing on where we actually need to focus on our business. That's the right outcome for our organization.
Future Implementation of H2O in Call Centers
So, with this implementation, we are not going to stop here eventually, right? We are looking at quite a few implementations going forward, and that will be based on the prioritization of the organization. One of the interesting use cases, which could be relevant to every other organization is that call center, right? So we would like to use H2O to assess the right call reason. So we get, as of today, over 40% of our calls are through self-service, but still 60% of the calls that we get are answered by the call center people. So we manage more than 1.1 million customers. So that means quite a large volume of calls that we manage. Due to some restrictions, a customer may call for various different reasons. There could be a primary reason, there could be a secondary reason. Or the secondary reason could be more problematic rather than the primary reason for them.
So, because of these applications, the way that they are built, we can only capture one particular reason for the call. So we are not able to understand or identify where the emphasis should be going forward in order to bring our customer service levels up. So what we are trying to look at is, in the future, we are looking at H2O to really understand our cues, the call generation, why is a customer calling us and what is the predominant reason for a customer to call? And based on that, we can improve our business, right? The other thing which is also not in this pack, which is very close to my heart, is the product library where we are, at the moment, looking at Hydrogen Torch and assessing it. So, the plan for us is to build a product library that contains the information about the product.
Using H2O to Understand Product Disclosure Statements
The insurance business is a complex business as we know, right? Even twins living in the same household may not get the same product, right? Your premium will be slightly different. So that means individual customers are having individual products. All those products are actually embedded in the product disclosure statements. So a product disclosure statement is a 30 page long statement in order for us to really understand. So what is a product in relation to claim? What is a product in relation to finance? What does the product mean for an organization? It's very difficult for us. So there is quite a hierarchy for a product. So what we are looking at going forward using H2O is to identify that product library and using Hydrogen Torch or going forward, we would like to build that capability within our organization that's not there in this pack.
Actuarial and Financial Use-Cases for H2O
But, that's something that we are working on, which is forecasted to be in Q1 2023, which starts from January. The other important use case that we are trying to look at is extending H2O over and beyond income production claims. So we would like to extend it to the other parts of our claims business. We would also look at extending it towards the marketing for campaigns and whatnot. In our finance space and also for our actuarial, right? In predominantly actuarial, H2O may play a key role, within our actuarial analytics going forward. That's all I have, guys. Any questions? I'm happy to take questions. Thank you. Yes, sir.
What Things Will Be Implemented In The Future?
So segmentation is the one where we have implemented H2O on. So, going forward, we will be implementing the processing of the claim as well. So at the moment, because we have only adopted the journey six months ago, right? The first step is to segment it and after that it's processed. The second step is processing. So yeah, we'll definitely utilize H2O to process our claims as well. End to end.
I'd like to remind people, if you do have questions for Josh, we've got a couple of minutes, so feel free to scan the QR code up there on the top and then we can address your questions. We'll give it a moment.
How Does H2O Deal With False Positives and Negatives?
And so you mentioned that your accuracy can reach around 77%, right? So are you interested in false positives or false negatives in a way? And how much will it be, and then what would be your next section to address this problem?
We are definitely interested in false positives and false negatives. You're talking about this slide, right? Yeah. So then again, right, when we are looking at the 77% accuracy, we assign that once we segment into a bucket, right at the moment, the process is, someone looks into that bucket and manages it. So if there is a false, a negative or a positive, right, they're going to pick it up from there. But we want to automate that going forward as well. We are only segmenting the claim as we speak as of today, right? But our process is to take it forward and make it an end-to-end seamless journey going forward. At the moment we are not there, but that's where we want to go.
Audience Member 2:
How do you deal with the sparse data of life insurance compared to other sectors?
I'll take this offline and I'll come back to you. Yeah.
What Machine Learning Platform Is Best For What Use-Case?
Audience Member 3:
Thank you for the talk. That might be a bit of a controversial question. You mentioned that you use Microsoft Azure as a platform for your ML, so why not using Azure ML Versus?
We use Azure ML as well. So the right use case for the right business case. Right tool for right use case.
We do have a data sciences practice. And we do use multiple products right? Within the organization. So what products work for it, depending upon how it works and the best outcome that we are getting, right? So if I'm getting a better outcome, I presented this claims triage. We have spent quite a bit of time building our own claims triage model, where we are able to get to 71% accuracy. Where, in less than a couple of months, we are able to get 77% accuracy. So we are adopted to that. There are a few models that are still built, in-house and we are still using those models.
What Can H2O Do About The Remaining 23% of Inaccuracy?
Yeah, there is definitely a risk of it. There is a saying "how do we mitigate the risk of this 23% accuracy?" Yeah. Obviously there is a risk of 23%, right? But also, depending upon the barring that we do with Snowflake, with H2O, right?How accurate do you want your model to be? We can increase those accuracies that are there on this slide. How accurate do you want, what timelines, what the interpretation is and whatnot. So basically we are happy with that 76% for the time being. We are looking for a 100% outcome going forward. But I think what we have now is what we have now. So definitely, there is a bit of a risk for those 23% where, if we can't segment those, they will stay unsegmented.
So that means one of our case managers will actually need to go into that 23% and look at those and segment it manually, right? But we are not going to leave them unattended, so they will be attended to. So, but 76 of them are automatically segmented. Have you seen change in using data scientists with the implementation of H2O? Yeah, there is definitely a change in H2O because H2O provides multiple models and you can pick and choose on the model. Documenting actually, interpretation is a big thing, right? Understanding the model, understanding the lineage, explainability of a model, explainability of a model to an audit, right? Or to a third party or through a governance body is very important. And that's what H2O provides in this particular context.
How do the important factors, discovered by the model and things that you can change in your business, improve actual outcomes going forward? Yeah, there's definitely the outcome. We have improved the outcome, right? If you're able to segment these 76% of our models, that means there's a lot of cost saving in there, straight cost saving. That is a capital expenditure saving, which is good for any organization. Look, if there are no more questions, thank you very much. Thanks for listening. Thank you very much.