Use Cases for Machine Learning in Insurance

Data science might be getting more attention these days, but it has always been a key part of the insurance industry. Actuaries were the original data scientists, and insurance companies have used advanced statistical models to price and assess risk for decades. And the insurance industry is changing fast. Disruptions from advancing technology and computational capabilities are allowing for wide adoption of machine learning (ML.) This is complemented by the type and nature of data insurers are now able to acquire. Increasingly, insurance companies are not only relying on decades old data like mortality tables, but also using dynamic models leveraging data collected in real time.

How much value ML can bring to your organization depends greatly on the foundation – people, organization, availability and quality of data. Before implementing advanced ML models, organizations need to make a fundamental shift in how employees are trained, how assessments are made, as well as how data is collected and processed across all aspects of operations. Beyond ML, it’s developing organizational skills and processes that will make a real difference.

This Power Read will offer an overview of ML, how it impacts the insurance industry and a look at some of its applications to date. Let’s look at where and how data is acquired and how this, in turn, reshapes the models and processes that currently anchor the industry.

Past and Future

Historically, the information insurance companies have on customers was captured just once, usually when the customer first purchased the policy. This information was not necessarily refreshed, especially for life products, even with decades having elapsed.

Compare this with a bank’s relationship with their customers. Like insurance companies, bankers need to get fairly detailed information when customers apply for a credit card or a loan. The captured information is used similarly to insurance: to assess the risk associated with the customer, and this information is then stored as part of the customer’s profile. Due to the nature of services offered, banks engage with their customers on a regular basis. Even if the relationship is largely transactional, they continue to learn a little more about the customer’s behaviour and preferences with each interaction. As insurers, we may not see customers for years and usually reconnect only as they need to claim, move to a different life stage, or seek to reassess their protection coverage.

Traditionally, and for the most part now too, insurers use details from these infrequent interactions -- a combination of the static history submitted years ago and other details such as demographics tying a customer’s risk to their generic profile – to predict future outcomes.

ML allows insurers to stop looking at customers solely through the lenses of the past data, even with a relatively small number of interactions. With more information available, insurance companies can move away from just providing indemnification. Instead, we can develop predictive models that can be used in prevention. Through the growth of ecosystems (links to healthcare, for instance) more interactions and information can be captured, to build better quality and more accurate models.

Data Sources

Where will the data come from? For one, first party data gathered through insurer owned applications and provided by customers themselves, for instance profiling information and needs analysis. While this data may have been gathered, in most cases systems to extract this data were not established or evolved to capture the information in a structured manner that would allow us to maximize its value.

Data from third parties and open sources can also be used to complement the data you have on hand. This could be data from government bodies, ecosystem partners, service providers or devices from health and wellness to Internet of Things (IoT) devices used in the industry. With a projected 1 trillion connected devices in 2025, IoT will only grow and propel the shift towards insurance 2.0 where predictive models relying on real time data will be widely used in insurance.

With the increased reliance on third party and open source platforms, infrastructure to support these exchanges must also be developed and governed. This includes government-led partnerships to anonymise data exchanges, assurances of data protection best practices, and implementing industry wide data standards that can enable portability of customer information between insurance providers. With this, customers will have more freedom to switch between providers without being penalised, should they make the choice to do so. There also needs to be clear policies from regulatory bodies such as the European Union (EU) to set the right standards to ensure that resulting new models driven by newly accessed data remain fair and unbiased.

Thanks to ML, you’ll have a better understanding of customers as well as new levels of insurance protection customization to better serve their individual needs. Traditionally, one-size-fits-all products were offered to customers due to the limitations of legacy systems and the fact that only coarse risk profiles and experience data were available. Today, we can offer tailored protection to an individual, and specialized models are being created at every stage of the value chain.

The measure of an organization’s sophistication in applying ML lies in how you train, improve and combine specialized models, delivering higher impact and achieving automation at scale. This section highlights key applications of ML that could aid your organisation.

Assessing Claims

In instances where the cost of processing a claim is higher than the actual value of the payout, companies have traditionally set a threshold of accepting most of the submitted claims with only a coarse level of sampling to minimize the costs. If the threshold for claims in a specific category was $100, a claim request for $90 would be approved without going through rigorous vetting. While this might lower processing costs, it opens insurance companies to the risk of fraud and abuse, where an individual could be awarded multiple $90 claims that may not necessarily be legitimate.

Instead of basing the decision to reward the claim on a fixed amount for everyone, ML offers the option to use the historical data and the behavior of policy holders to determine personalized criteria instead. A policyholder with a good track and no negative signals record could have a higher threshold, say $1,000. Another who has previously exhibited red flags might have a threshold of $40. If both individuals wanted to make claims of $120 each, the company would have needed to conduct rigorous assessment for two claims with the previous fixed threshold of $100. Using ML reduces processing costs and time by only assessing claims in ‘higher risk’ cases to reduce the chances of fraud and at the same time providing benefits of quick turnaround time to good customers.

Improving Customer Experience

To process medical claims, insurers need supporting technical medical details. For traditional claims, policyholders are expected to provide specific details, despite not being medically trained experts. Naturally, this breeds frustration on both sides. Insurances aren’t able to process the claim without sufficient information. Policyholders have no clue what these details actually are, much less where to get them.

With ML applied in Optical Character Recognition - OCR, that can identify text from images of the documents, specific medical details can be pulled from a form completed by a healthcare professional. The insurer gets the information they need, and customers don’t have to stress about deciphering medical jargon making it much more hassle free.

Computer vision technology can also be used in other fields like home and motor insurance claims. Instead of continuing to send traditional surveys to assess the damage, customers could send videos and pictures for insurers to get a sense of the severity and provide the first level of cost assessment.

One more example of application of ML in combination with computer vision is for surveyor activity. One of the most popular use cases is in car insurance where customers can submit videos of the car before getting insurance instead of this task being done by a surveyor. The model will detect all the body scratches, bumps etc. automatically dramatically reducing cost and time required. Similar models are used also for damage cost assessment. The ML model will be able to provide an indication of the severity of car damage and estimate the cost of repairs as direct input to the claim process, speeding it up significantly.

Detecting Fraud

Perhaps the most advanced and successful application of ML to date lies with fraud prevention. Given the complex nature of fraud and the inevitable lack of cooperation from customers, identifying clusters and trends have been an ongoing challenge for the industry. It’s the proverbial game of cat and mouse.

With ML, insurers can now conduct a more detailed analysis of outliers tied to potential fraudulent incidents. From here, we can build more models to detect fraud and aid investigations.

Advanced fraud and abuse models can even detect more elaborate fraud schemes like organized rings of fraudsters where individual customers make infrequent claims, which won’t raise red flags per se. However, by analyzing patterns and connections, ML models are able to catch these complex frauds. In many countries, with support of the regulators, anonymized claim data is being shared across the industry to more effectively detect and combat complex fraudulent activity across markets.

1 Enabling Predictive Models

Traditionally, insurance models have been built using historical data to approximate future outcomes at the population level. With new sources of data that capture individual behaviours and preferences more frequently and accurately, predictive models are being used to drive prevention.

2 Expanding data sources

Insurers can increasingly use third-party and open data sources by being part of larger ecosystems to better understand their customers – even the ones they interact with infrequently. Infrastructure and regulation must evolve to ensure ethical use and security of even anonymized data exchanges in order to create models that are fair and unbiased, especially since our ability to understand why an ML model made a specific decision is extremely limited.

3 Improving Customer Experience and Preventing Fraud

By enabling automation for key processes like claims, ML allows companies to offer improved customer experience by taking away the burden of submitting technical details and improving the speed of service. ML also opens new possibilities in offering highly customized protection solutions. The biggest potential and gamechanger is the ability to use ML in combination with new types of data to deliver a true win-win solution: helping prevent situations that result in claims from happening in the first place.