by Joanne Butler, Head of Product Marketing and Pre Sales, Charles Taylor InsureTech for Insurance Day, first published on 14 October 2021: Industry trust in AI models essential to the future of claims management

Artificial intelligence (AI) is changing the game in the claims space, bringing increasing speed and efficiency to the claim process, enhancing the customer experience and helping insurers improve their claim ratios.

According to a 2021 survey of Europe, the Middle East and Africa chief information officers by Celent, many insurers have adopted AI use cases, with 41% investing in natural language processing, 37% in anti-fraud systems and 33% in image recognition or geospatial analysis to improve their claim management capabilities.

AI is enabling the automation of back-office processes such as triage and fraud detection as well as helping automate customer journeys in high-volume retail classes. This allows insurers to settle genuine claims faster than in the past, although there is still room for improvement, such as using AI to tailor an increasingly personalised triage response or mine a variety of data sources for missing information rather having to ask customers addition questions.

With self-service increasingly popular, chatbots will play a central role in processing basic claim notifications, verifying policy numbers and even detecting sentiment in human language. Voice analysis software will also increasingly be used to predict whether a caller is genuine or not, with Celent finding 26% of insurers have invested in some form of speech-to-text processing tool.

The next big area of development for automation is in claim settlement. Instant approval and payment of claims would take the customer experience to another level but requires a high degree of trust in AI models. The industry is only at the beginning of exploring this application of AI, starting with the initial focus on automating the most straightforward low-value claims.

As well as driving claims costs down by reducing fraud and human error, AI is also helping insurers and claim managers to build a more holistic understanding of claims and claim portfolios, drawing a growing range of data sources into decision-making and using historical claim data and predictive modelling to mitigate future risks – creating a virtuous cycle that should continue to improve claims ratios over the long term. The benefits to customers and insurers are plain to see, though not everyone understands how AI works in practice.

Under the bonnet
The two key components of an AI solution are the rules engine and the machine learning model (see graphic). Rules are a sequence of yes or no questions that determine the fundamental characteristics of the claim. The machine learning model is essentially an algorithm written in code (via Python or JavaScript, for example) which uses historical claims data to simulate scenarios and predict the likelihood of a particular outcome in the future.

The quality and relevance of the outputs generated by the rules engine directly reflect the quality and relevance of the questions asked. Similarly, predictions made by the machine learning model are only as reliable and relevant as the data fed into it, so a high volume of high-quality data is critical for success. It takes a significant upfront commitment of resources to build and train a machine learning model and ensure it has the quality and volume of data it needs – and as data quality varies by line of business and geography, AI is much easier to apply in certain parts of an insurance business than others.

Raw claim data must be enriched with as much useful data from additional sources as possible to provide more detail and context on the claim, from text analysis of unstructured data sources such as police reports and photos using optical character recognition (OCR) to publicly available geo data on the claim location and environment and real-time data from connected devices.

Insurers will increasingly use OCR to verify and better understand claims, whether quantifying property damage or assessing a claimant’s body mass index or health habits from their selfies. Meanwhile, internet of things data feeds, from home appliances to industrial machinery and premises, provide accurate, timely data on why a claim occurred as well as ongoing insights that help predict the likelihood of future claims. If seamlessly brought into the claim process, this data may one day trigger automated claim settlement in certain classes of business.

Data gathering
Data is also gathered from internal and external sources providing background on the claimant including historical insurance relationships which may identify, for example, a high volume of similar or unsuccessful claims made by the claimant, or multiple identical claims made under different names – all of which could indicate fraud.

Having absorbed all this data, the machine learning model’s algorithm delivers its output in the form of a score or probability ranking (one to 10 or one to 100, for example) – in the case of the example flagging high-risk fraud cases and triggering either a referral or rejection based on the insurer’s agreed tolerance threshold in a particular class of business. Investigation and verification by human adjusters then close the loop on each claim.

Crucially, those findings are fed back into the model, helping it learn and improve its accuracy over time. Information generated by the tool is also continually fed back to the insurer, helping mitigate future fraud risk, improving underwriting decision-making and delivering continued improvements to the insurer’s bottom line.