Learning Machines outlines the business case for AI and insurance
Fred Senekal, Head of R&D at Learning Machines
Artificial Intelligence and machine learning are significant drivers of change across various industries. Simply put, many tasks that are currently undertaken by humans can be executed by algorithms and robots with much greater speed and efficiency and at lower cost. Learning Machines, an AI, machine learning, big data engineering and cloud services consulting firm, believes the insurance industry is advancing.
It is not to say that humans will be replaced by machines any time soon, but many organisations are rushing to automate their production, distribution and service due to the enormous benefits associated with it. With the future being highly competitive, only those organisations which are adapting with great speed to the current technological revolution will remain relevant and ultimately survive.
Learning Machines says Insuretech companies are gaining market share at the expense of larger, more established enterprises. These enterprises are usually slower to adopt new technologies, often due to constraints imposed by legacy systems, outdated business models, excessive regulations and company culture.
However; there are many opportunities to adopt machine learning and automation across the insurance value chain, whether in product and service development, marketing and sales, underwriting, contract and policy management or claims management and customer service.
Leveraging its many years market experience and solution knowledge, Learning Machines has identified several use cases that demonstrate the value machine learning and automation offers the insurance industry.
Insurance pricing and underwriting
The correct pricing of insurance products is vital to managing risk and the ultimate profitability of insurance companies. Traditionally, pricing and premiums are determined according to statistical modelling approaches, such as generalised linear models.
As we are seeing a rapid increase in the availability of customer data, large benefits can be had by applying machine learning techniques to the data of individual customers. In fact, machine learning becomes necessary to be able to detect patterns in large volumes of data as these approaches are able to exploit non-linear relationships in datasets with many features.
Consider for example the use of telemetry to gather data about driving behaviour. Already, various insurance companies have created smartphone applications that measure and monitor various driving data points and can detect risky driving behaviour. Based on this data, car insurance premiums can be adapted on an individual basis.
Not only do customers tend to perceive such systems as fairer than traditional approaches, but it also offers the advantage that people change their driving habits in order to receive the benefit of lower premiums, thereby reducing risk for the insurer.
Claims Reserve Optimisation
Insurers need to set aside claims reserves to pay policyholders who have filed or are expected to file legitimate claims on their policies.
In some cases, such as permanent health insurance, settlement occurs over a large number or months or years, presenting an ongoing liability for potentially long lasting claims. Reserves are set aside, typically on a conservative basis, to account for future liability.
As claims in force grow over many years, these reserves can become fairly large and even a small improvement in reserving basis can represent a large amount of actual monetary value that could be invested otherwise.
In many cases, traditional approaches are based on life tables or survival analysis, using data accumulated over many years and often pooled among insurers. Newer machine learning approaches can incorporate survival analysis in machine learning frameworks and can outperform traditional approaches with substantial margin. This can have a significant impact on reserves that need to be kept in hand.
Claim Fraud Detection and Prevention
Claim fraud is an unfortunate reality that costs the insurance industry (and ultimately consumers through increased premiums) billions of rands each year.
In South Africa, the Insurance Crime Bureau estimates that up to 20% of the R35 billion paid in short-term insurance claims could have been fraudulent in 2019. In the United States, non-health insurance fraud costs an estimated $40 billion per year.
Detecting claim fraud is a difficult and labour intensive process, requiring humans to investigate claims and data at an individual level. Where systems are in place, they often rely on rule-based algorithms that are difficult to update with new data sources and quickly becomes irrelevant as new methods of conducting fraud evolve.
Machine learning algorithms are ideally suited for fraud detection and can identify claims that should be subjected to further investigation. For example, voice analytics can be applied to voice recordings of claim submissions to detect how nervous a person appears to be when submitting a claim. Other systems can correlate social media data with submitted claims data, for example to identify whether a person has consumed alcohol prior to a car accident claim.
Accelerated Claims Processing
The insurance industry has been around for quite some time with many insurers having existed for decades and even centuries. Unfortunately, some are slow to adopt new technologies and are hampered by bureaucratic and outdated processes. Many companies still require slow, time-consuming paperwork to fill out claims and are lagging behind in digitalisation.
In the digital age, customers expect on demand, real-time and efficient submission and processing of claims. Insurers now offer new ways of submitting claims, for example through submissions on smartphones or web portals. AI systems now assist customers with the submission of claims, by guiding them through the claims process.
Where scanned documents are involved, document capture technologies and optical character recognition can efficiently capture typed forms. In fact, systems can now read handwritten text (for example doctor’s letters) at a level exceeding human capability.
Various parts of the claims process are also being successfully automated, from claims routing to approvals. Insurtechs have gained market share by offering systems where claims are submitted digitally and are approved within minutes.
The use of automated inspection is becoming more prevalent to validate underwriting and claim decisions. This may be during pre-cover, post inception or during renewal.
Drones equipped with cameras are being used to identify potential problems with the structural integrity of buildings or structures. For example, technology now exists to capture and analyse images of rooftops to identify tiles that may be broken, which could potentially lead to water damage after a storm.
Deep learning applied to geospatial imagery provides risk and value information related to properties. In agricultural insurance, drones and unmanned aerial vehicles are being utilised to access damage to crops, allowing accurate assessment of the extent of damage.
With the abundance of camera-enabled smartphones, insurers now offer apps to snap and submit photographs of vehicle collisions to optimise and fast-track claims. These apps contain AI photo analytics to assist users to adequately take photos during the occurrence of an accident.
Damage and Repair Cost Estimation
Closely coupled with automated inspection, is the automated estimation of damage and repair costs.
There have been enormous advances in the field of computer vision over the last decade, driven by advances in deep learning algorithms and the availability of large image datasets.
These algorithms are now being applied in the insurance industry in areas such as automatic vehicle damage inspection.
Computer vision algorithms are able to detect defects such as scratches, dents, rust and breakages as well as the parts of the vehicle that are damaged and with what severity.
After an inspection a report can automatically be generated, containing a list of damages and an estimated repair cost. Such an approach standardises damage and cost allocation, as current approaches based on the use of human assessors could vary greatly.
Efficient Customer Support
The application of machine learning to customer support can drastically reduce the amount of labour intensive effort, which saves time and reduces costs significantly.
For example, natural language processing (NLP) is now commonly applied to incoming emails to route it to appropriate departments and people, while extracting relevant claim information and analysing sentiment. This allows for quicker response times and improved customer satisfaction.
The use of chatbots and conversational AI has also become standard practice. While being somewhat overhyped a few years ago, chatbots are now widely adopted in the market, with customers expecting to be able to interact with an on-demand and intelligent virtual assistant to answer common questions, submit information or to follow up on the progress of their claims.
From the point of view of the insurer, this allows human effort to be allocated to queries that are more difficult to be handled, enabling call centres to be run efficiently.
Personalised Recommendation Systems
Customers have different needs, preferences and lifestyles. They expect personalised policies, loyalty programs and recommendations, based on their individual preferences and attributes.
Insurers are starting to provide tools that offer personalized insurance plans, based on machine learning models trained on individual customer preferences. These tools may provide machine-generated insurance advice, to ensure that customers have adequate cover.
In health insurance, wearable devices are now being used to track health vitals, with automated reports being generated to provide healthy living advice and risk reports.
Marketing and Propensity Analysis
In marketing, propensity analysis allows the use of data to predict whether a person will take a particular action, for example make a purchase or accept an offer. Propensity analysis may be used to find the most likely candidates for a new product or offering, allowing marketing and sales efforts to be directed efficiently.
Machine learning models may also be used to predict the most appropriate times to contact customers, for instance after work or during lunch.
The acquisition cost of new customers is substantially higher in the insurance industry than in many other sectors. It is much less expensive to keep an existing customer than to acquire a new one.
Insurance companies are utilizing churn prediction to predict when customers may churn, enabling them to take proactive measures to keep their clients.
Machine learning algorithms can pick up on leading indicators, such as changes in utilisation of apps and rewards programs, a change of frequency of interacting with customer support, changes in income or changes in life circumstances such as becoming pregnant or getting married.
Similarly, machine learning algorithms can predict attrition in employees, by monitoring changes in work habits and employee satisfaction.
In conclusion, experts at Learning Machine say the insurance industry has enormous potential for the application of machine learning, due to the fact that it is inherently data driven.
The automation of labour intensive processes lowers costs, saves time, increases operational efficiency and improves the quality of customer service.
Clearly machine learning allows insurance companies to be faster, cheaper and more accurate.
To remain competitive, insurance companies need to adopt machine learning approaches fast, or risk losing market share to competitors.