382 Little Collins Street
Melbourne VIC, 3000
(03) 9008 5922
11-31 York St
Sydney, NSW, 2000
Jonathan: That’s interesting. Can you give me an example?
Hannah: Well, even if you are the best driver in the world, that doesn’t stop someone from running into the back of your car. Then you’ll make a claim, and really see the value of the insurance product. That aspect of the insurance industry really resonates with me — that we show up for our customers and impact their experience when they need us the most. That concept drove me into the claims space and eight years later, I still find it fascinating.
Jonathan: If IAG is one century old, then the claims process must have evolved over that time.
Nowadays, you must see many opportunities to apply new technologies to it, such as AI. At Silverpond, we see a big trend towards AI, because the technology is now mature enough for enterprises to apply it in the real world — they can take it beyond writing a report or building a proof-of-concept model. But that doesn’t always happen. We are particularly interested in the operational side so I’m keen to know how you managed to take your claims AI project from concept to execution. What’s the story?
Hannah: I am passionate about activating AI by building it into our operations because that’s where we’ll see real impact. The story began with a vision to switch from a product-centric organisation to a customer-centric organisation. To achieve this, IAG created a Customer Journeys program to map out the end-to-end customer experience. This starts from when the customer first buys an asset, like a car or a house, realise they need to insure it, get a quote and buy a policy, through to making a claim and starting the process all over again. At IAG, the process of getting a customer back on their feet after a claim is called the ‘Recover Journey’. During that Recover Journey, we want to make the experience as effortless as possible for the customer. We could see there was immense opportunity to improve the customer experience in this area.
Jonathan: So having realised that, what did you do?
Hannah: We pulled together a cross-functional team to examine the entire recovery experience following a claim. We wanted to see it from the customer’s perspective, not look at it from the point of view of IAG processes and structure. We started by interviewing our customers to see what their pain points were. Then we validated these insights with data, which helped us prioritise which areas to look at. As a result, we delved into the motor total loss claims experience – more commonly known as a ‘write-off’. A total loss occurs when your vehicle is too damaged to be repaired. As they are heavier hits, it can be a horrible experience. All you want to do is to get a new car, forget it ever happened and move on with your life. These types of claims can often drag on as they take much longer than repairing a ding.
Jonathan: Could you tell us something about the pain points?
Hannah: One was around the communication blackout. When you have a claim, you might not hear anything for a while. As a customer, you don’t know what’s going on, so you call IAG to check in on progress and that clogs up our call centres as well as frustrating our customers. The second major pain point was time to resolution. Customers want to get back to normal quickly.
Jonathan: How did you approach improving these pain points?
Hannah: We examined the end-to-end total loss process and hypothesised that if we were able to identify a total loss earlier in the claims process, we could then communicate it earlier to customers, reducing the communications blackout.
The previous process was that the customer would call up then take their vehicle to a repairer, who would do a quote. Our assessors would have a look and find it was uneconomical to repair because the damage was too great. Getting to this point would take 7 to 14 days, with the customer not knowing what was going on. We would then tell the customer their car was a total loss and often they would be surprised at the decision. After that, there would be more time spent settling the claim. The customer would have to dig out supporting documents and the consultant would enter the information manually then calculate the settlement offer. Often, the customer didn’t understand why deductions were being made and would ask questions, which would lengthen the process.
Breaking down the whole process from the customer’s point of view helped us to identify how we could use AI and process automation to make the whole experience smoother.
Jonathan: So what was the outcome?
Hannah: We combined AI with business process automation to create a quicker, less stressful end-to-end experience. Now, IAG uses information collected at lodgement to predict total loss pretty much straight away using that data. This then triggers an SMS to the customer to let them know their vehicle is a potential total loss and provides a link to total loss FAQs. It’s an early warning system that mentally prepares them for that decision.
If our assessor determines the vehicle definitely is a write-off, we automatically calculate the settlement amount and notify the customer. To accept our offer, the customer simply clicks a button, which triggers the payment. And that is how we are now able to settle things so quickly. Instead of taking 3 weeks, the whole process is reduced to as little as 3 days, sometimes quicker.
Jonathan: What do IAG’s customers think?
Hannah: Before we scaled this to all our personal lines customers, we ran a trial on a subset. We put some customers through the new experience and others through the old experience, then sent out surveys to measure customer satisfaction. We saw a 10-point uplift in the Net Promoter Score (NPS) for customers who were part of the trial. They were a lot happier with the way we were managing their claims, which was an exciting outcome for us.
Jonathan: I read that you worked under an AI ethics framework and did an impact assessment. Did you work with legal and compliance teams as part of this cross-functional team?
Hannah: Yes. We worked very closely with our Data Governance team and Director of Algorithmic Ethics to apply IAG’s robust AI ethics framework. Before we even started using the data, we got sign-off for its usage on the basis that the intent of the model was to improve customer experience. It also helped that we weren’t using any personal information in the model.
One of the key considerations from the AI ethics framework was around potential harm to customers if the prediction was inaccurate, considering false negatives and false positives. For example, what if we predicted someone’s vehicle was a total loss and it ended up not being a total loss? What was the experience for the customer there? That is why, when we SMS the customer, we are very careful. We make it absolutely clear that the customer vehicle is a potentialtotal loss, and we make it clear that the assessor will be making a final decision.
With the false negatives, where the vehicle was a total loss but the model didn’t predict it, customers would miss out on getting the SMS. This was considered to be less of an issue because the customer would be just going through the standard claims experience and was no worse off.
Jonathan: How much did you have to consult with the rest of IAG, and did this affect the timeframe?
Hannah: This was truly a cross-functional delivery. We partnered with many parts of IAG to get the total loss use case activated. As I already mentioned, Data Governance and Ethics were some areas we consulted. Additional to that, we had to work very closely with our digital and technology teams to deploy and integrate the model. The Claims team was also heavily involved as the process owners, and we also had to bring in expertise to compose the customer communications.
By taking a tactical approach, we were able to commence the pilot in just six months. We didn’t want to implement a full production release initially — we wanted to test that it would work and get feedback first before putting in the effort. Once we saw the great NPS results, we decided to scale it. It’s been incrementally delivered and continuously improved over the last 18 months. The reductions in claim cycle time are a result of the automations that have been introduced incrementally over time.
Jonathan: With the pilot, did you have any challenges collecting the data?
Hannah: Not really. Sourcing robust data is often the first challenge for many industries, but we are lucky that the quality of data for insurance is fairly high. Obviously, as an insurance company, IAG is in the business of assessing risk, so we have to collect good data to do that and feed it into the pricing algorithms. Also, IAG has been consolidating and simplifying our source systems in our Enterprise Data Hub for several years. So it is actually fairly easy for us to gather data for modelling.
Jonathan: So the current model is based on mostly structured data?
Hannah: Yes. The model mostly comprises structured data, like vehicle make and model, sum insured and damage areas. That said, the claims description also plays an important role in the model. We used natural language processing to generate features from the claim description as well.
Jonathan: Tell us more about the deployment and maintenance.
Hannah: We deployed the model on our internally developed Analytics Platform, but we needed to integrate it with some of IAG’s technology to be able to send out the SMS, talk to our claims systems and record the model predictions back into our Enterprise Data Hub.
Also, the work doesn’t stop once you’ve deployed and activated your model. You still need to maintain it, monitor it and refit it. At the moment, our team of data scientists manages the AI lifecycle from end-to-end. That includes the model build, as well as monitoring and continuous improvement.
Jonathan: Do you find it hard to add people to your teams? Talking to other people in the industry, they’ve often found it very hard to find data scientists, particularly those that might come from a statistical background, because a lot of data scientists come from more of an analytics or business intelligence background.
Hannah: Forming a diverse team within your data science team helps with that cross-skilling. We’ve got people with actuarial backgrounds. We’ve got people with stats backgrounds. We’ve got people who might have come in from more of the computer science side of things then have developed a passion for machine learning. I think having diverse backgrounds is really productive, as everyone brings something to the table.
Jonathan: The AI industry is moving pretty quickly. Back in 2018, a lot of people were asking what AI was. But the industry is more sophisticated and experienced now. We have got a few years under our belt applying this technology. What is your observation about this?
Hannah: I think you are right about awareness of AI now. A few years ago, everyone was talking about big data. That was the latest buzzword. Then it was AI. Now, I think people have a better grasp of AI and businesses are starting to wonder what they can do with it. Informing them about what other business, like IAG, are doing is really important. So is building trust in AI, letting people know it’s more than killer robots.
There is an element of whether AI can be trusted, like what happens if it goes wrong? Part of how long it takes to deploy a new AI model is building trust. Before the pilot, we used data to validate the model and the data scientists were comfortable with the accuracy. Our business doesn’t have that level of trust yet, so we did things like score the model on a daily basis so they could see the results every day, and that built trust. Conducting a small pilot also built trust, and gave us the confidence to gradually scale it, and scale it some more. Hopefully, the more times we do that, the more trust we can build in AI and what it can do.
Jonathan: Talking of trust, did you produce a mechanism for the end customer to be able to ask or challenge the decision made by the model?
Hannah: That is definitely an issue if you are using the model to automate the decision fully. What we have done is start off using AI to augment decisions, but we are not fully automating that decision. We tell a customer that their vehicle is a potential total loss. The assessor is making the final decision, and they are a person. We are not replacing a person with AI here. We are more in the space of augmenting decisions at the moment and supporting and validating decisions, not fully automating them.
Jonathan: That is what we are seeing in the other industries Silverpond participates in, like medicine and utilities. I think any of these heavily regulated fields are going to want to see this transition period.
Hannah: It will take time, but it is worthwhile because AI outcomes can be transformational. Supporting people to do their jobs better is really the angle we are coming from and trying to alleviate some of the tasks that people do not want or need to do and letting them do what they are good at, like the human empathy side of things. AI should let us spend more time talking to actual customers instead of doing some of those manual tasks.
Aurelie Jacquet who chairs the Standards Australia committee that is representing Australia as ISO develops international standards on AI where we discuss the upcoming AI standards undergoing the drafting process at ISO and how our industry can embed best practices and meet the expectations from the rest of our society.