Solutions

Challenge & Pain Points:

Data-driven decision-making and execution are key to the underwriter’s success. In today’s world, the models driving the execution of the business must be dynamic and based on basic units. Execution models towards decreasing costs and increasing revenue, should be built upon these units. Unit economics is the discipline of measuring profitability on a per unit basis to facilitate this approach and to achieve profitable scale up. It is rooted in the in-depth understanding of the units, which are the customers in the retail insurance business. A practical approach is the accurate estimation of Customer Lifetime Value (CLV) and Customer Acquisition Cost (CAC), as well as their covariance.

Traditional actuarial approaches treat profit as a pure function of cost of risk and a simple margin over cost which are almost always modeled for single products and a single coverage period. However, earnings generated by a customer are neither limited to a product nor to a single policy term. Customers may cancel policies and churn, or renew and even buy other products over a sustained period of time spanning multiple years, which are possibilities overlooked by the traditional actuary. Hence, profitability per customer is a function of the ratio of CLV to CAC. The higher the ratio is, the higher the profitability. Including various operating costs, such as retention, maintenance, etc., in the picture will give an analyst or decision maker the CLV/CAC ratio threshold for smart decision making. At the end of the day, existing customers with subthreshold ratio, as well as missed (or churned) customers with high ratio will return as a loss, either through wrong acquisitions or missed opportunities, respectively.

Insursight.ai Solution:

These underlying mechanics of unit economics make it a perfect subject matter for AI. Both CLV and CAC need to be predicted. Who is likely to buy what (acquisition, up-sell, cross-sell), where (distribution channels) and how? Who is likely to churn or renew and when (average lifetime)? And if possible, what is the expected cost of retaining a customer belonging to a specific segment? Is it worth it?

Traditional approaches are based on simple statistics of past data. The aggregate model uses average past revenue per customer to come up with individual CLV estimates, overlooking inter-customer covariance. The cohort model attempts to address this weakness of the aggregate model by manually defining cohorts, providing single CLV estimates for each cohort, skipping per customer assessment. In fact, there are inter-customer covariates that can provide invaluable insight into personal traits; such covariates are just not evident at first sight, hence are impossible capture by manual rule-based groupings. That’s where AI gets into play. Modern AI approaches utilize high-dimensional predictive models fit to whole past data, as well as external factors such as seasonal dependence, stock-market indices, etc., to predict individualized CLV and even CAC. State-of-the-art learning strategies even allow us to push the limits of what can be done with available data. Unlike early predictive models such as GLM or models based on specific probability distributions, these models are capable of learning not only the nonlinear relations between covariates but also the covariates themselves and/or complex, non-parametric probabilistic models.

Challenge & Pain Points:

Insurance is about calculable risks. Irrespective of the insurer’s business (pricing) objectives, it is not possible to sustain a viable operation without an accurate estimation of the monetary risk (cost) and its distribution over the insurer’s liability period.

Traditional approaches such as statistical frequency-severity modeling, or GLMs are limited with inherent assumptions, manual feature engineering and low dimensionality due to their ability to exploit limited number of features. In practice, imbalanced data distributions invalidating the assumptions and limited features used result in sub-optimal cost estimations leading to adverse selection of risk instances.

While costing is a regression problem, pricing is an optimization problem towards a usually complex objective function that considers multiple business aspects. Pricing has never been as simple as adding a profit margin over the estimated costs and interests. Although technical profit is still an important factor, it is not the only component of pricing objective function in today’s sophisticated world of insurance. Other components that must be covered by the pricing model include reinsurance, portfolio, customer lifetime value, operations, cash flow and reserve strategies.

Insursight.ai Solution:

Statistical models start by assuming a distribution model for a given dependent variable, such as frequency and severity (or their product), and estimate the model parameters from prior data while selecting the features to use through some hypothesis testing strategy. In a similar fashion, GLMs start by assuming a non-linearity (the link function) through which a linear relation can be built between dependent variables and manually selected features.

Challenge & Pain Points:

Claims processing has multiple aspects including claims data extraction, fast vs investigative track (fraudulent claims) segmentation and actuarial policy compliance to name the foremost ones. Accuracy throughout the claims underwriting is the key to profitability. Though accuracy in costing and pricing is a prerequisite, profitability cannot be achieved without fast and accurate claims processing. While accuracy ensures planned technical profit realization by ensuring compliance with underwriting processes and the actuarial assumptions that were used in costing and pricing decisions, fast processing improves customer experience, increasing customer satisfaction and company reputation. The conventional pipeline involves manual processing which is not only prone to errors (over/under-coverage) but also increases processing times leading to customer dissatisfaction. Despite these conspicuous pain points, current AI applications in insurance almost exclusively target operational problems, more specifically data extraction via document processing, and underutilize the potential of AI in solving core problems by means of both improving conventional rule-based decision making via discovering efficient and effective rules, and going beyond it where simple rules are insufficient and/or require substantial human expert involvement, such as health status assessment in health insurance claims processing.

Insursight.ai Solution:

Insursight.ai provides full-fledged AI-driven automated claims underwriting. The solution provides fast and accurate claims processing, not only speeding up the process but also ensuring compliance to actuarial principles at the time of costing via determining the claims to be compensated automatically and optimally for both the beneficiary and the insurance company. Insursight.ai solutions do not only utilize the claims data but also the policy itself and the external data such as public health, weather, economy, insured’s credit score and social network, the claimer’s characteristics, etc.

Challenge & Pain Points:

Attracting new customers is much more costly than keeping existing ones. Accurately predicting whether a customer is likely to cancel their policy, or not renew it, is thus a highly valuable capability for an insurance provider. If a crystal ball could show which customers are considering leaving, one could attempt to change their minds by offering promotions, or could do targeted surveys to determine common causes of strong dissatisfaction in order to improve products and services.

One common way to attack this problem is via expert analyses of past data, trying to find clues, with the benefit of hindsight. Which aspects of a customer’s history or demographics could be predictive of their nonrenewal of a policy? This kind of manual analysis is a difficult task that requires domain knowledge, and painstaking trial and error. Methods of traditional statistics like logistic regression take some steps towards automating this task, with the added limitation of strong assumptions on the way the likelihood of churn depends on the observed behavior and demographic data.

Insursight.ai Solution:

The Insursight.ai Churn Predictor product takes a machine learning approach to churn prediction, taking in all available data including policy characteristics, demographics, external factors such as seasonality, and crucially, various forms of recent customer engagement, to predict the likelihood of churn in a specified time period. Without strong mathematical assumptions such as those of logistic regression or GLM, our models are able to “listen to” what the data actually says, rather than trying to put the rich “square peg” of customer behavior into the “round hole” of a parametric model. In addition to making accurate predictions of churn, the model also gives  insights into the factors resulting in churn via modern techniques of machine learning model interpretability.