We are often asked by clients how much data is enough for AI analysis. People may have low confidence on the data they have. After all, insurers are not Google or Amazon, who has very large amount of digital and high frequency behavioral data on the customers.

Let’s face it. Insurers have legacy systems, which are often disconnected. A lot of the processes are paper based. While most insurers have embarked on the path to full digitalization, it will take a very large amount of investment in time and money.

Is it that we must wait that long? We at Cutters have thought long and hard on this question too, coming from knowledge of insiders in the insurance industry where we have utilized insurance data for analytics for over two decades.

The imperative for us is therefore that any practical solution we come up with must work within the confine of the current state of insurance data.

We are happy to report that Relacio does not need to consume very large amount of data to be effective. In fact, we don’t have a requirement for a specified list of data in terms of size and shape.

In some instances, the data we worked with was only several thousand records, while we also worked with datasets with millions of records in other cases. The good news is that the performance appeared very consistent across various datasets that we have applied. This is all due to the unique way we approach the problems.

Just don’t underestimate the data you have. Let Relacio work on it!

In the age of big data and AI, companies are re-discovering the value of their data. British mathematician Clive Humby proclaimed that data is the new oil.

But companies must come to terms with the quality of the data they have. Most insurers have disparate systems which were not designed with a vision how data can be worked together to yield analytic insights.

Yes, data is the new oil, but it is not, if the new oil is dirty!

While most IT departments do enforce data quality checks in their databases, those checks are typically rule based. It is primarily designed to look for violations of data definitions. For example, if it is a numeric filed, it shouldn’t have text input. Or if it is an interest rate, it shouldn’t, normally, be greater than 100% or less than 0%. Even so, it is not possible to anticipate every possible type of violations.

Worse yet, there is one type of data errors that is virtually undetected in these detection methods – logical inconsistency. One example of logical inconsistency is a claim for cervical cancer from a male policyholder. While male and cervical cancer separately are legitimate in its data definition, putting them together presents a logical inconsistency that is quite likely caused by data entry errors. Logical inconsistency between data is exceedingly hard to detect.

This type of data problems inevitably biases the analysis because the analysts cannot control for it, unlike the more conventional data error types, which can always be excluded if needed in the analysis since its existence is known.

Relacio possesses unique ability to detect logical inconsistency errors in the data. Deep sweep on the data allows us to spot those errors. With this capability, combined with conventional data error checking, Relacio provides a good handle on how good the data really is.

Companies may be surprised by how bad their data error problems are. In one case, we estimated that 10% of the claim data has data error problems, as measured in terms of claims amount. In another case, as much as 65% have data errors in it.

Accumulating large amount of dirty oil doesn’t do much good. Data is absolutely a competitive advantage. With Relacio, meaningful, automatic and real-time data quality monitoring is doable.

Insurance product mis-selling is one of the key and consistent themes that regulators everywhere place a hawkish eye on. Mis-selling compliance is very complex for insurers as it involves many agents and large number of products. The buyer/seller dynamic is exceedingly difficult to monitor.

While insurers place great emphasis on education, training, and process standardization to minimize mis-selling, it is no doubt a labor-intensive and elusive process. Breaches often get reported in the news, which often result in regulatory fines and irreparable harm to reputation.

We believe one smart way to go about detection is through the analysis of data. If events are recorded in the data, we can help to discover it.

We have had numerous conversations with insurers on this topic. While the amount and type of data varies in each company, typical insurers do record sales and agent behaviors.

In one application, we applied Relacio to the data on a portfolio of newly issued policies. The data only contains policy information. Yet with such a small size of data, our findings are interesting.

Examples of what we found are: 

Both are potentially mis-selling cases.

Relacio can be placed as a safeguard check before the policy is issued.

Health insurers incur large financial losses to fraud, waste, and abuse in its claims. According to The National Health Care Anti-Fraud Association (NHCAA) in the United States, annual cost of fraudulent health claims is more than USD 300 billion. While there are no reliable statistics on the size, it is believed that up to 10% or more claims are lost to FWA activities. This figure is believed to be much higher in some Asian markets.

To detect FWA, most insurers rely heavily on rules from experience or expert judgements. These static, reactive rules need to be manually reviewed, maintained, and updated regularly to keep pace with constantly evolving fraud schemes and patterns. This maintenance process is labour intensive and is rarely timely enough. In recent years, while machine-learning based predictive models have been used to detect FWA, these types of models frequently rely on learning from historical fraud cases. However, one key problem is that most insurers haven’t accumulated meaningful list of historical fraud cases for the machine to learn from. A natural consequence of these approaches is that it always lags behind new FWA schemes.

Relacio represents an innovation of AI tool that breaks through the limitations of existing tools. We look for FWA activities by studying the data directly without human intervention. Virtually all types of FWA can be potentially detected, whether you have seen it or not in the past.

Some of the examples Relacio uncovered and brought to human’s attention are provided as follows. Relacio found these cases without any prior knowledge or human intervention.