Award Applications

FIRST EVER INNOVATION TO DETECT MOTOR CLAIMS FRAUDS USING ARTIFICIAL INTELLIGENCE

Ground-breaking motor claim fraud detection system including:
1. Natural Language Processing (NLP) technique, customised for the Thai language – one of the more complicated languages in the world.
2. Ultra-fast path-finding algorithm to detect fraudulent claims in seconds versus possibly days or hours through manual methods
• Allianz’s proprietary innovation, a patent has been filed
• Fully scalable – to handle an increased number of parts and thousands of claims, without performance degradation or user experience, laying the foundation for motor claims straight-through processing

5,000 ways to defraud

With some 5,000 parts in a car, and various touchpoints in the claims process including clients, garages, surveyors and assessors, a motor claim is highly susceptible to fraud.

Two of the most common methods a dishonest garage submits fraudulent claims are through “parts upsell”, where claims are made on parts which were not damaged in the accident, and “severity upsell” where damages of different parts are made out to be more severe than they actually are.

Conventionally, a motor claims assessor has to manually check the claims details submitted by garages and input his assessments into a central claims system. This is followed by a visual assessment of damages through the photos submitted. If there is missing information or if clarification is needed, the assessor has to get in touch with garages.
This time-consuming process means that an assessor spends on average an hour to process a single claim. Whether a fraudulent claim is successfully detected or not largely depends on the judgement of each assessor, individual experiences, understanding of car parts, and judgement of the potential impact of accidents, all of which increase the subjectivity of the assessment.
Customised for the “most complicated” language

When the Allianz Data Science Team was approached to radically improve the motor claims process in Thailand to prevent frauds, it was clear from the onset that a truly innovative solution had to be found.

The Thai language – with some 44 consonants and 32 vowels, no grammar or tenses, and with no punctuation marks to denote the end of a sentence – is often described as one of the most difficult languages to learn.

With the names of some 5,000 car parts often written in free text on claims forms by the fragmented car garage industry, a solution had to be created to match against a standardized parts list, which has never been done before in the Thai market.

Once the data is processed and parts standardized, the data is modeled with a graph network technique, where parts are linked together as nodes in the network through patterns learned using machine learning from thousands of historical claims records.

Enter AI and Deep Learning

NLP’s deep lexical ranking of free text w.r.t standardized parts list

Allianz’s data science team formulated a proprietary algorithm through the use of deep lexical ranking.
Other than listing physical parts, garages also include notes written by hand, on labor costs and services provided. Therefore we had to ensure the machine was able to recognize the free-form written text through NLP, and then “train” the machine to group the different descriptions of the same parts to match against the standardized parts list.

As an example, analysis of past data showed that a simple damaged “side mirror” could be written in various ways in Thai, such as “side mirror”, “side mirror joint”, and “side rear-view mirror”.

Graphical Networks and Ultra-Fast Path-Finding Algorithm

The key to outsmarting fraudulent claims is to be able to accurately scrutinize the potential damages and impact from an accident to avoid “parts upsell” and “severity upsell”.

The team had to use graphical networks to pour over some half-a-million claims data of the past. Using AI, the team constructed linkages between the potential severity and parts damages from an accident, taking into account the direction of impact and sphere of damages.

Allianz’s data scientists were able to develop a ground-breaking, heuristic path-finding algorithm which speeds up the process by a thousand times as compared to the traditional algorithms such as breadth-first search, depth-first search, or dynamic programming.

One-second average processing time

This fraud detection process has already been successfully deployed in Thailand for proof of concept, with a success rate of 76% accuracy and a truly digitalized process, and also reducing the process from 1 hour to 1 second.

As it stands, traditional algorithms do not match up to Allianz’s proprietary algorithm’s speed. Traditional algorithms used in the market still take up to one hour where there are more than 16 parts, whereas Allianz’s proprietary algorithm only needs one second to achieve the same.

Raising the industry’s bar

The success of Allianz’s data science team has gone a long way in raising the industry’s bar, through a system that is highly scalable and transferrable.

Rather than depend solely on the experience and expertise of claims assessors, which may lead to inconsistent results, the Allianz Data Science team’s fraud detection system helps to standardize the process and greatly reduce the claims leakage rate. The fraud detection system can also be easily integrated into existing claims systems, allowing for easy adoption. The system will soon be deployed in other markets.

Early results in Thailand show a cost avoidance of more than 15% from this solution, compared to previous methods. The speed of detection will cut down manpower costs, which will translate into more time for investigation of detected claims, leading to better user/customer experience. More importantly, it has also resulted in higher levels of customer satisfaction and trust in Allianz, given our ability to fight fraud and a quicker turnaround time for claims approval.

Kick starting Straight-through-processing (STP) for claims

This solution has also laid the foundation to kick-start the creation of STP for motor claims, as only high risk or anomalies that require human intervention will be flagged.

We are now looking to roll out the solution to other parts of Asia, starting with Malaysia.