According to the Association of British Insurers, a total of 112,000 general insurance frauds valued £1.28bn were detected in the UK during 2017. Over the years this number has grown consistently as shown in the following chart:
Industry experts believe that insurance frauds have typically been inherent in the process of policy underwriting for long. However, the value of frauds was significantly low before the 2000s and hence did not demand the immediate attention of the industry. Since then insurers have started adopting various rule-based methods of fraud identification. However, with the advent of digital technologies, these rule-based systems are grossly inadequate and inefficient in identifying fraudulent claims.
Thus, we believe that there is a need for Artificial Intelligence and Machine learning models to predict and manage insurance frauds. In the following sections, we present our proprietary framework for Fraud Management based on AI & ML technologies.CONVENTIONAL FRAUD MANAGEMENT
Conventional fraud management systems inclined to focus more on “soft frauds” also called opportunity fraud, when a person takes advantage of paying a lesser premium or exaggerated claim. On the other hand, “hard fraud” is executed by more organized crime groups and so the intensity of attacks is higher.
Insurance companies have created business rules to identify frauds, but the fraudsters are smart enough to exploit the gaps in these rules (mostly static) by manipulating information
Few of the challenges faced by insurance companies are mentioned below:
1. The static fraud identification process, rules are based on experience & simple business triggers.
2. Fraud management process is manual in nature.
3. The process takes a long time for identification as it is reactive.
4. There is a no structured inventory of fraudulent triggers & feedback loop for continuous enhancement of process.
5. Absence of a platform for real-time information consumption.
6. Majority of activities are temporary & ad-hoc, without continuous monitoring & involves incident reporting only.AI/ML Backed Fraud Management Framework Identification:
Foundation of our framework is a data-driven machine learning-based solution. We leverage sophisticated techniques for pattern mining such as Link Analytics, Sequencing, Associations, Discriminant Analytics, Geographical data mapping, K-means Clustering and support vector machines. These techniques unfold the latent behaviour, exceptions and unusual activities.
Streamlining data eco-system, addressing data quality and accessibility will further ease out the fraud identification process.Classification & Validation:
Identified unusual patterns are tagged as suspicious. Ensembled machine learning techniques may be utilized for scoring & classification. This score & classification further validated using k-fold cross-validation. Business benchmarking can achieve through back testing & pilot run of the model in a futuristic time frame.
Intelligent Consumption & Fraud Management components are explained with an example in Business Impact Section.Key Differentiators
1. RPA enabled Claim Processing: Claim Processing is highly data & document-intensive. Robotic process automation facilitates smooth collation of data from various data sources to be used at a centralized document so that claims can be processed faster. The Key is to smartly automate an existing set of workflows to improve agent productivity.
2. ML-driven Fraud Management: This helps to predict patterns of claims, prior assessment of claims, claim classification and provide various triggers related to frauds.
AI-powered interactive Chat-Bots: This facilitates real-time Q&A through text & voice. These intelligent bots are efficient and serve as a virtual assistant. Natural Language Processing backed Chat Bots are helping customers to resolve their issues & concerns. Bots are trained in a manner to deliver personalised solutions that result in customer delight.Advantages
Machine learning models result in dynamic rules. These rules are managed & validated through robust model management practices. Some of the best practices we follow as mentioned below:
- Validate data distribution for rules
- Corroborate Algorithms
- Auto Assignment of Weights
- Monitor & Compare model accuracy
- Score Testing & comparison over the period
- Automation of the entire monitoring task
- Allow configurable Multi-level degradation threshold
- Notification to model managers when threshold crosses Business Impact:
- One claim requires 9 hours of processing on an average if there is no ML model in place
- Post model implementation, we can process ~70% of the claims STP (Straight Through Processing)
- STP of Claim requires only 3 hours of efforts
To assess Business Impact Analysis, we have shown the complete flow of activities.
The example shown of processing 1000 Claims is illustrative in nature:
Above illustration is based on below mentioned assumptions:
The framework clearly shows low risk customers can be processed faster and hence reducing the time for overall claim processing (about 70 % of claims are STP). Our model uses reinforcement learning through feedback loop (traced through red lines here) to improve accuracy (reduce false positive).
As a result, our framework serves customers quickly and cost effectively.Conlusion
It has become indispensable for insurance companies to develop a robust & effective strategy to proactively combat mounting fraud problems. Our framework enables insurers to Identify, Assess and Manage frauds across multiple lines of business by using components of Data Management, AI/ML Models & BI Systems for faster delivery.
** Framework & Numbers shown in the paper are only for illustrative purposes.