How machine learning is transforming insurance claims
Machine learning (“ML”) has been one of the most prolific areas of high-impact use cases for the insurance industry. And within insurance, claims management offers one of the most promising areas to apply this technology due to the large amount of data available to train algorithms and the consistency of principles applied in the process of loss assessment. Here, we look at some of the use cases for ML in claims and the challenges limiting adoption.
Focus on fraud
Fraud detection is perhaps the area where we are seeing the most advanced adoption of ML among insurance companies. Startups such as Shift Technology, Friss and Owl Labs have seen strong demand from operators and have attracted significant capital from investors to support their growth. These tools work by applying state-of-the-art data science to large sets of historical claims data, augmented with third-party data.
Insurers have been quick to adopt ML-based fraud detection strategies because they can often deliver the most immediate and tangible return on investment. Many fraud teams within claims rely on rules-based approaches (often in the head of a claims adjuster) and tend to focus on claims where the scope of fraud is the most obvious, potentially overlooking larger and more complex fraud cases. In contrast, ML can often detect more subtle fraud patterns that might not be visible on an individual claim. For example, is a single body shop regularly overcharged for multiple claims? Does a doctor’s office routinely diagnose patients with whiplash in low impact crashes? Such patterns may only appear when reviewing all historical claims associated with an individual policyholder or provider.
An additional area of significant innovative activity has been damage assessment. The development of computer vision to analyze claim photographs has been the central innovation of several startups, especially in personal lines.
In auto claims, companies like Tractable and Snapsheet allow policyholders to submit photos of their damaged vehicles and generate a loss estimate without a human assessor viewing the image. Their models are trained on historical data for similar vehicles and are often accurate enough to settle a claim instantly or make an informed decision on how best to move a claim forward. Computer vision assessments speed up decision making, reduce claim leakage (paying too much to settle a claim), and improve the customer experience. Expert evaluators still have a role but are directed to more complex cases where confidence is low.
We see similar products emerging in property damage assessments from start-ups such as Flyreel (acquired by LexisNexus), Hover and Hosta Labs.
A more nascent area that we believe has huge potential is the use of ML to automate the processing of complex documents. Claims professionals today are buried in paperwork. Even simple claims can include damage reports, doctor’s notes, multiple invoices, emails, and text messages, all of which contain crucial information in unstructured or semi-structured formats.
Natural language processing (NLP) and computer vision technology have the potential to significantly reduce manual data entry. ML applications can look up an invoice, for example, and extract individual lines, payment information, and invoice numbers. At the end of the process, an accounts payable professional only needs to approve the invoice with all the necessary context at their fingertips. Small claims that meet certain criteria and without a “red flag” can be approved without any human interaction.
Startups often train their models on a single class of documents (at least initially). Hypatos is a leader in invoice processing. DigitalOwl focuses on reviewing medical records. Groundspeed Analytics started as a loss specialist. Insurers can put these tools together to extract all the necessary information from documents so claims professionals can focus on decision-making, not data entry.
Barriers to Adoption
In short, we are seeing new companies forming that are using ML to bring efficiency to claims management. Perhaps the biggest challenge these companies face today in adoption by incumbent insurers is integration into existing systems.
Data is the cornerstone of any machine learning model. ML applications need historical data to train and tune models and once they’re up and running, they need quick access to new claim data to be effective. Currently, only a small number of carriers have the technology to deploy ML models in their claims operations.
There are dozens of potential use cases for ML that can bring efficiency to the claims management process. However, many start-ups get so bogged down in carrier integrations as part of the initial pilots that projects never get past the POC stage. To fully unlock the potential of ML in claims, insurers will need to transform core IT systems.
Based in London, Jack Prescott is a senior partner at MTech Capital, a venture capital fund focused on the insurtech space.