While AI in claims has improved customer interactions and increased efficiency and decision effectiveness, the human touch cannot be removed entirely, especially when AI falsely identifies a claim as fraudulent and when deepfakes are involved. Asia Insurance Review spoke with Swiss Re’s Mr Ali Shahkarami and Cover Genius’ Dr Peter Phillips.
According to Accenture’s 2022 report, Why AI in Insurance Claims and Underwriting? Improving the insurance experience, AI has “emerged as the transformative technology and critical differentiator in the insurance industry, especially when applied in tandem with humans”.
For instance, the report said that AI could be “transformative in the enablement of improved customer interactions, increased efficiency/automation and decision effectiveness”, although it would have to be “applied responsibly”.
“The emergence of fully integrated claims platforms is a revolutionary shift,” Cover Genius CTO Peter Phillips also said, speaking to Asia Insurance Review.
Ensuring fairness
To mitigate the risk of biased outcomes in the use of AI in insurance claims, according to Swiss Re global head P&C solutions Ali Shahkarami, it is crucial that “the underlying data is responsibly used and that adequate controls are in place”.
He said that potential bias could be identified by using mitigation techniques at three different stages of AI deployment:
- Pre-processing: Applied to training data
- In-processing: Applied to an AI model during its training
- Post-processing: Applied to predicted labels
“Through such techniques, AI algorithms can be continuously monitored and evaluated to detect biases and remove them from the predictions,” he said.
Addressing false positives
It is important that AI tasks are “clearly defined and that humans always have full control of the decision-making”, Mr Shahkarami believes, as predictions “are prone to false positives which may lead to wrong claims decisions”.
“Some of the ways to address this issue include threshold adjustments, feature engineering, feedback mechanisms and manual reviews combined with adaptive learning,” he said.
According to Dr Phillips, while technologies could “help catch bad actors and potentially automatically deny a claim, direct human decision-makers [would] still (be) needed to spend time on the right problems, manage complexity and try to ensure that it does not happen again”.
Catching deepfakes
While it is technology that assesses the online evidence provided to determine fraudulent activity during a claim, Dr Phillips said, “further action needs human involvement”.
When fraudulent or deepfake activity is determined, he said, insurers would need to:
- Decline the claim
- Stop it from happening in the future. It is important to rely on human touch to manage the complexity and make a final decision on fraudulent claims
Fraud prevention
According to Dr Phillips, when protection is embedded directly in the platform at the point of purchase, the condition of an object a customer has purchased is already known. In that sense, he believes that by working directly with a partner to offer embedded protection, a partial vet has already been performed with the data gathered at the time of purchase.
Mr Shahkarami said that AI-led led image algorithms helped claims managers make faster and smarter decisions during and after Nat CAT events. The platform was built on customer research and is supported by a partner ecosystem providing satellite imagery, property characteristics or damage estimations, he said.
Improving fraud detection rates
According to Mr Shahkarami, there are several methods in which the detection of fraudulent claims could be improved:
- Training detection models with data. The use of more comprehensive and higher fidelity data is an essential component of improved model predictions, including the identification of fraudulent claims
- The implementation of systems that continuously monitor incoming data and flag potentially fraudulent claims would allow for a prompt response to suspicious activities
- Feature engineering, which would help extract relevant features from the data that signal fraud
- Retraining models via feedback loops as more data becomes available
- Insurers could also use insights from industry fraud patterns to develop more targeted and transparent models
Dr Phillips also said that while AI has been used for a whole range of applications, such as dynamic pricing and product recommendations, “it still has to be carefully applied to claims assessment”.
Protecting personal data
As the insurance business “is reliant on strong data and analytics capabilities”, Mr Shahkarami said, insurers are well-equipped with state-of-the-art systems that ensure secure storage of data.
“Additional important factors in this context are effective governance structures, constant vulnerability assessments and patching, as well as regular trainings to ensure employees stay vigilant,” he said.
According to Dr Phillips, the environment is also highly regulated, and compliance is at the core of the business.
“We are bound to some of the most stringent data requirements and always operate within the appropriate regulatory framework,” he said.
The future
Using AI in claims could streamline the flow of claim information and result in improved customer interactions, faster settlements and reduced claim handling expenses, according to Dr Phillips.
“Results like this can help distinguish companies in a competitive environment to help build customer loyalty and retention,” he said.
On the other hand, Mr Shahkarami said, due to rapid advances in the data and technology spaces, the “nature of fraudulent claims as well as the tools to confront them is constantly changing”.
While he does not expect any “extreme one-off developments within a short span of time”, he predicts a constant evolution over the coming years.
“As more sophisticated fraudulent activities are possible, there is a clear need for insurers to adopt new innovative approaches using technology and deep risk expertise to counter such attempts,” he said. A