**The Insurance industry is characterized by predictions, insight, and assessment. Machine Learning delivers exactly that, making business operations easier and more efficient, from both companies and users sides. **

Swish analyses three ways Machine Learning can enhance the Insurance industry.

Product Discovery & Recommendations

To obtain insurance, one of the first steps is to fill out many long forms then have an advisor review them and make suggestions on what the client would like to buy. While this information is necessary, machine learning can help reduce the friction of the process and increase the quality of results.

First, forms can be digitized into an easy-to-use wizard where all the information is automatically saved for the user. Once the profile is complete, a machine learning model can then access the client’s risks and make suitable recommendations.

These recommendations are generated after training the machine learning model with other users data. The model is able to recognize patterns of products selected by people of a similar profile. This can be modified to also include items such as client satisfaction with the product chosen to make the recommendations more accurate. If the client profile changes throughout the lifetime of the policy, the same model can determine if the policy is still the best fit.

If it is part of the insurance offerings, the potential client can have multiple accurate insurance recommendations, without any work needed from an advisor. Of course, the help of an advisor can still be used to verify the accuracy of the client information, the recommended products and discuss potential options not listed.

This whole process can be surfaced in the form of a digital assistant/chatbot (more on them later). All the information can be gathered in a conversation, and recommendations can be presented much like a real advisor would.

In many industries, digital assistants have proven to be successful at increasing the level of engagement of users already on a website/app.

Pricing & Credit Risk Assessment

Analyzing a client's credit risk and assigning an appropriate price is one of the most important and difficult aspects of the insurance process. Two key risks to consider are the clients’ likelihood to pay their premium and whether they are going to cause a large loss payout. Both risks can be better analyzed using machine learning and big data.

Traditional credit risk analysis uses statistical models and client data to predict the likelihood of default. While this analysis is effective, there are some flaws, often leaving out quality borrowers who lack traditional credit scores linked to geography or work history. With clients’ consent to provide additional data (such as bank statements), machine learning models are able to spot patterns of creditworthiness. The advisor can then grow their book and the client can get the coverage they deserve & require.

A small percentage of clients result in the vast majority of large loss payouts. Knowing which clients are likely to cause these payouts is critical to any insurance company. Much like predicting the risk of default, machine learning models are better able to predict payouts by spotting patterns not visible to the human eye. This allows companies to offer more competitive pricing or add a necessary risk premium.

Machine learning models have the potential to reflect bias driven from the data. Although these biases may be the correct business principle, it could penalize one client segment beyond what is deemed morally correct. Many prediction algorithms are leaving out data points such as gender and race to avoid these. This is why it is also important to still have a human’s input on the model's output to make any necessary adjustments. This isn’t just theory, a proof of concept by AXA shows an accuracy of 78% in predicting large loss payouts.

Automated Claims Process - Chatbot + Fraud Detection + Adjustment Prediction

Submitting and processing claims can be a lengthy, unenjoyable process. It requires clients to find and fill out multiple forms, often with multiple back and forths with customer support. Once submitted, claims go through inspection for potential fraud and get reviewed by a claims adjuster for payout. All of this process takes time, potentially causing undue financial stress on the client, and increasing costs to the insurer.

There are multiple ways machine learning can improve this process for both parties.

Starting with the submission, the process can be entirely digitized with an automated chatbot. The chatbot asks the client a series of questions, which replaces the need for filling out forms or calling customer service. The client can also send photos of the item/incident and record a video of themselves describing the claim. Chatbots have benefited greatly from deep learning advancements. Traditionally they have only been a series of pre-programed questions and responses (iterative models) but now they can be trained to take on a more life-like conversational approach (generative models). Combining the two approaches makes for a very capable customer support agent. This type of bot comes in both web or mobile versions, making it more convenient for the client to submit and reducing support costs for the insurer.

The claim can then be automatically sent for fraud detection.

Fraud detection is a “perfect problem” for machine learning to solve. To assist in the classification of fraud a model type called anomaly detection is used. Training a machine learning model on verified claims allows it to spot claims that lie outside the realm of what is normal/common. Then the claim can be escalated to the proper employee for further inspection. If the claim looks good, it can be moved to payout.

Not all payouts can be fully automated, but the process can still be improved through machine learning. Data that is usually not in the claim, such as on-site inspection and additional interviews, is often required to fully process a payout. A machine learning model predicts payouts, by training it on previous payouts done by the company. An advisor can fill in any missing data in the claim to get a range of potential values, providing the client with clarity along the way. Once all the data is collected the model can give a recommended payout that the claims adjuster can approve or modify.

These machine learning models can be applied in similar areas throughout the insurance process. The chatbot described can be used in general customer support and policy servicing requests. The fraud detection model can also help detect false KYC filings. The automated claims service is already showing a lot of market traction with the startup Lemonade raising $120MM.

This analysis was brought to you by the Machine Learning team at Swish.

We build tech solutions that bring an edge to businesses. Do you want to take your Insurance business to the next level? We can make that happen. Let’s chat.