AI vs Fraud in the UK and Japan

Fraud is a constant and ever-changing threat in the global insurance industry. In partnership with Aioi Nissay Dowa Insurance, Aioi Nissay Dowa Europe and The Aioi R&D Lab Oxford, Mind Foundry has developed powerful AI solutions to combat it.

In the UK alone, insurers detected 72,600 dishonest insurance claims in 2022, valued at £1.1 billion, with motor insurance claims representing 59% of that total. However, that number only shows the fraud that was detected; the true cost of fraud, including fraud that goes undetected or unreported, is likely two to three times that amount.  Economic pressures have exacerbated this issue, potentially increasing the number of fraudulent claims and raising premiums for honest claimants worldwide. 

Types of fraud 

The types of fraud insurers see every day go by many names, including: 

• Opportunistic fraud: When policyholders exaggerate legitimate claims, such as inflating the value of stolen property.  

• Hard fraud: Deliberate fabrication or staging of events. For example, staging car accidents or arson. 

• Internal fraud: Fraud committed by employees or agents within the insurance company itself, including embezzlement or falsifying claims to meet quotas. 

• Claims fraud: Submitting claims for losses or damages that never occurred, such as false medical claims for nonexistent injuries or claiming phantom passengers in car accidents. 

• Underwriting fraud: Providing false information or using identity fraud during the application process in order to obtain lower premiums. 

• Premium fraud: Intentional manipulation of the premium calculation process. For example, falsifying business revenue or payroll data. 

Fraud Techniques 

The techniques for committing fraud are constantly evolving in order to evade the traditional ways of detecting them. These include: 

• Staging accidents: Intentionally causing accidents in order to file claims. This is most common in motor insurance. 

• Exaggeration of claims: Overstating the extent of damages or injuries. These are very difficult to distinguish from legitimate claims. 

• False documentation: Using fake or altered documents to support claims, such as false repair invoices or medical reports. This is getting much harder as generative AI makes it easier to create deep fake photographs or injuries and accidents. 

• Identity theft: Using stolen identities to file claims or apply for policies. This is often linked with broader criminal activities or groups, such as fraud syndicates. 

• Inflating repair costs: Collaborating with service providers to inflate repair or replacement costs. For example, overcharging for car repairs or medical treatments. 

• Multiple claims for the same incident: Filing multiple claims with different insurers for the same incident. Also known as double-dipping. 

• Misrepresentation: Providing false information about the cause of damage or injury, such as claiming weather damage for pre-existing damage. 

Data Challenges in Fraud Detection

Finally, significant data challenges necessitate advanced solutions that can handle large datasets efficiently and in real-time. Challenges with data include: 

• High-volume of claims: Insurers process thousands of claims daily, making it challenging to scrutinise each one. Robust data management and analysis tools are required. 

• Data from multiple sources: Claims data comes from various sources, including policyholders, medical providers, repair shops, law enforcement, etc. Integrating and cross-referencing data is complex and becomes more challenging as data volumes increase each year. 

• Unstructured data: Many claims include unstructured data, such as handwritten notes, images, and videos. Analysing this data requires advanced technologies like natural language processing and computer vision. 

• Historical data: Insurers need to analyse historical data to identify patterns of fraudulent behaviour. This requires large-scale data storage and efficient retrieval systems. 

• Real-time data processing: Detecting fraud often requires real-time data processing and analysis. Implementing real-time systems is technically challenging and resource-intensive. 

• Privacy and security concerns: Handling sensitive personal and financial data necessitates stringent security measures. Insurance is heavily regulated and fraud detection must be balanced with compliance to data protection regulations (e.g. GDPR). 

• Data quality and consistency: Poor data quality can lead to false positives/negatives in fraud detection. Data accuracy, completeness, and consistency must be ensured across systems.

Enhancing Fraud Detection Through Human-AI Collaboration

Insurers are increasingly using AI to address fraud detection. One example of this is in the work Mind Foundry has done recently with Aioi Nissay Dowa Europe (AND-E) and the Aioi R&D Lab – Oxford. 

Greg Cole, UK Claims Director, Aioi Nissay Dowa Europe, describes the problem they were facing. “Originally, we were using a traditional rules-based model to flag and triage potential fraudulent claims. But fraudsters are smart; they're really good at finding new ways to avoid detection, and they're really good at evolving what they do. That caused our model to deteriorate over time, and the lack of explainability meant that we could no longer trust or understand the predictions it was making. So we knew we needed something better. We worked really closely with Mind Foundry to develop an explainable AI solution that was perfect for our unique business requirements and most importantly, would gradually get better and better at learning fraud detection patterns.” 

As more insurers utilise AI to improve and automate fraud detection, AI should not be seen as a silver bullet solution that can replace humans. While AI can quickly process large volumes of data and identify patterns indicative of fraud, it may lack the nuanced judgment and contextual understanding human investigators provide. Additionally, there are concerns about transparency and accountability of AI decisions, the potential for bias in algorithms, and the risk of over-reliance on technology, which might overlook subtle or complex cases of fraud that require human intuition or expertise. 

“It is impossible to replace the human expert,” says Tadeo Corradi, a Senior Research Software Engineer at Mind Foundry who was actively involved in building AND-E’s unique fraud solution. “It will always be a collaboration between an expert and AI. The tools we build at Mind Foundry help them prioritise the investigation of claims so the expert can investigate that claim more efficiently.”  

At the heart of Mind Foundry’s Fraud Investigation solution for AND-E is a powerful, continuously learning predictive model trained on a vast dataset, including over 20 million unstructured documents, handwritten notes, and historical claim data. The model looks at all incoming claims and, at any given point in time, can assign each claim a fraud score based on a set of features jointly developed by AND-E UK’s fraud specialists and Mind Foundry’s data scientists and engineers.  Crucially, the model also learns continuously. This lends a critical advantage to counter-fraud efforts because “The behaviour of people trying to commit fraud is ever-changing and ever-adapting to any countermeasures that you apply,” says Corradi. “And therefore, a successful model is one that stays ahead of those changes, or at least it adapts to them quickly.” 

How Mind Foundry's fraud solution works

AND-E UK has deployed this predictive model in a live Claims Investigation Dashboard, which their claims team uses to discover and investigate high fraud score claims and carry out a claims similarity search. AND-E's Chief Product Innovation Officer, Phil Norris, says, “Our claims fraud team have really embraced this tool, with a key feature of this dashboard being the explainability of the score.” 

Jan Martin, Head of Third Party Claims, says the AI solution is an important addition to his team and helps AND-E keep premiums down for their customers. “Last year, the solution increased automated referrals retained by our fraud department by 800%, meaning that handlers work on much fewer false positive cases. This led to a 120% improvement in overall fraud detection that not only helps us defend our business but also improves the customer journey, with faster claims handling and reduced claims cost. It's also helped us reduce false positives by over 50%, meaning we now work much fewer claims and close them faster. And we reduced the overall claims cost to our business by recording a 4% saving on our capped Third Party indemnity spend, nearly double than the previous year.” 

As impressive as those numbers were last year, the results are expected to improve each year because of the unique way the model has been designed and deployed to continuously learn. 

Model Decline and Continuous Learning 

Research shows that 91% of Machine Learning models decay after their first year, mainly due to new data. The rapid evolution of fraud techniques exacerbates this decline. Risks of model decay include: 

• Increased fraudulent payouts: Ineffective detection leads to more fraudulent claims being paid, increasing financial losses. 

• Customer dissatisfaction: Legitimate claims could be wrongly flagged as fraudulent, leading to delays, disputes, and dissatisfaction among policyholders. 

• Reputational damage: Perceived unfair treatment due to false positives can damage the insurer’s reputation and erode customer trust. 

• Missed fraud cases: Genuine fraudulent claims might slip through undetected, undermining the effectiveness of the fraud detection system. 

• Regulatory and compliance risks: Failure to accurately detect and manage fraud can lead to regulatory scrutiny and penalties. Insurers could face legal actions from policyholders whose legitimate claims were unjustly denied. 

• Operational inefficiencies: Increased manual review of flagged claims due to declining model accuracy can impact human resources and operational efficiency. 

• Competitive disadvantage: Competitors with more effective fraud detection models could gain a market advantage and attract more customers. 

Maintaining Machine Learning models' performance and accuracy is crucial for insurers to effectively manage fraud risks and uphold their operational and financial integrity. As more models move out of the test phase and become a critical part of the operations of mature insurers everywhere, maintaining and governing AI responsibly becomes an increasingly time-consuming and resource-intensive task.  

Graph showing how machine learning models decay over time

Operational Efficiency in Fraud Detection
 

When AND-E set out to build their fraud detection tool, they knew they would need something that didn’t just perform well on Day One. It needed to maintain, or even exceed that performance, as time went on. They needed something that would automate and reduce some of the most time-consuming aspects of model governance so their team could focus on other challenges.  

Greg Cole says, “When you go past pilots and proofs of concept towards actual deployment of AI, in-house data scientists can spend half their time on AI governance, including retraining a model or trying to understand its health and performance. But this solution has simplified all of that for us, meaning we can focus more of our time on building models rather than managing existing ones.”  

Cole and his team measure model performance and ROI of their fraud solution against numerous KPIs and business goals. They also use the solution’s powerful Performance Dashboard to benchmark the latest results against simulations of what the same model would have achieved in a static, non-learning deployment. The results have been exciting. In 2023, the model’s ability to continuously learn generated a 52% increase in value creation over a 12-month period and is tracking to maintain an increase in 2024.  

Motor Repairer Fraud in Japan 

The same fundamental technology that powers AND-E’s fraud solution in the UK can create and maintain similar value, even in an environment where completely different techniques for committing fraud are used. One recent example of this occurred in Japan last year when a major scandal emerged involving the improper exaggeration of claims and inflation of repair costs by a group of large car dealerships used by many Japanese underwriters. The scandal inspired a nationwide discussion and highlighted the need for new methods for detecting fraud. 

In response, Aioi Nissay Dowa Insurance (ANDI) collaborated with Mind Foundry and the Aioi R&D Lab – Oxford to further develop an award-winning AI technology that could provide this type of solution for all of Japan. At the heart of it is a new, continuously learning predictive fraud model that is trained on 4.2 million pieces of data on repair estimates provided by the insurer. The model powers a Fraud Detection System that looks at actual fraudulent insurance claims and repair costs and then predicts the likelihood of encountering fraud at individual repair shops.

Onsite investigations by human inspectors can then make the final determination about fraudulent activity. The solution’s strength is in the way it gets better over time and in the way it creates powerful human-AI collaborations by bringing  AI-generated insights together with human judgements, leading to more accurate detections of previously undetected fraudulent claims. Aioi Nissay Dowa Insurance recently won two awards from the Insurance Asia Awards for the development of this solution. 

ANDI General Manager Daisuke Kodama says, “Measures against fraudulent insurance claims are one of the key issues that the entire Property and Casualty (P&C) insurance industry needs to address from the perspective of customer protection and the operation of a healthy and stable insurance system. The recent inappropriate automobile insurance claims made by a major used car sales company have necessitated further strengthening of the damage inspection system for insurance companies.” 

Always Staying Ahead of the Game 

Fraudsters around the world are crafty and constantly on the lookout for new opportunities. Insurers wanting to use AI to assist them in their fraud detection capabilities should look for solutions that combine continuously learning AI with human-centric domain expertise and insight.  “If there's one thing we know about fraud,” says Oxford Professor and Lab Advisor Michael Osborne, “whether it’s claims fraud in the UK, repairer fraud in Japan, or warranty fraud in Europe, it’s that people don't stand still. This is why it’s essential that our models continue to adapt to the world as it changes.”   

You can watch the full video here 👇