AI Is a Tool for Decision-making and Not a Decision-maker — Great American Insurance Group Data Science Manager

(US & Canada) Nasser Mooman, Data Science Manager at Great American Insurance Group, speaks with Gavroshe Chairman Derek Strauss, in a video interview about data-centricity, AI use cases, the challenge of acceptance of AI applications, and the categorization of AI.

Mooman begins by stating that the insurance sector has a wealth of legacy data. He establishes that the type of insurance varies, as it could be health or automobile-focused, and each has its own structure and domain.

Adding on, Mooman appreciates Great American for truly being a data-centric organization. Delving deeper into the kind of data the company has, he explains that it is structured around 36 unique business units, each functioning like a company on its own. For instance, one might have health, car, property, or casualty insurance, and each of the units comes with its own data sources, users, and applications. This diversity makes Great American Insurance stand out, he says, especially how this data is used by data scientists to train AI models.

These models are tailored to the specific data types, sources, and customer needs within each unit to improve prediction accuracy and optimize outcomes, says Mooman. For example, in crop insurance, data is collected from sources like satellite and drone imaging, along with historical weather data, to predict events like hail or drought.

Moreover, Mooman states that Great American not only collects its data to train AI models but also utilizes third-party partner data to enhance its models. He notes that the company tries to get a model that is not biased but generalizes in a way that utilizes diverse data to benefit customers.

From predicting weather patterns for farmers to assessing driver safety for fleet operators, Great American Insurance Group’s diverse and rich data creates an impact, and Mooman feels fortunate to be a part of the data science team.

When asked about the challenges in choosing the right use cases to pursue, Mooman states that balancing between ensuring data security and maintaining a legacy ecosystem is challenging. Also, the customers and agents are hardly on the same page.

However, the organization has made significant investments in applying AI and leveraging its existing data, leading to numerous use cases that benefit the end consumer.

One such use case is “medical record formation,” says Mooman. These records often contain years of detailed information from nurses and doctors, which traditionally would require a human to analyze. Now, with AI, insights from these valuable records can be extracted without needing manual interpretation.

Furthermore, Mooman mentions summarizing long medical histories, spanning hundreds of pages, making the information accessible for doctors, underwriters, or patients. This capability speeds up processes such as insurance claims and fraud detection, benefiting the insured.

Nevertheless, Mooman maintains that it is imperative to have a human in the loop to verify and validate the AI’s output. He adds that the company does not rely on AI autonomously, and AI is a tool to provide enough information for decision-making and not a decision-maker.

The next use case Mooman highlights is predicting property loss. For instance, he mentions using historical and image data to assess potential risks of whether a house in Florida is vulnerable to hurricanes or if a property in California might face fire threats.

By predicting these losses in advance, Great American can help customers prepare, ensuring their property meets insurance standards, and can also expedite claim processing when damage occurs.

Shedding light on the challenge of social acceptance with AI applications, Mooman states that it is similar to that faced by other technologies. He continues that AI is being used in many industries, with social media and search engines quietly collecting data behind the scenes.

In this scenario, Great American Insurance Group has put safeguards in place to ensure responsible data usage. For example, when training AI models, the organization relies on reinforcement learning, which involves collecting feedback from users, like when they correct grammar or paraphrase text. This feedback helps fine-tune the models over time.

However, Great American prioritizes the security of its customers' data, ensuring that no data leaves the secure environment, even when working with partners like Microsoft Azure. In addition, the organization does not allow external retraining of models and strives to keep the models bias-free and ethically sound.

Mooman stresses that if end-users are shown the benefit of utilizing AI in a way that improves human life, the application will be eventually accepted. He considers it critical for companies to implement guardrails around data usage. At Great American, while using AI in any application, every AI model is validated by a human to ensure its accuracy and fairness.

Elaborating further, Mooman categorizes AI into three segments:

1. Autonomous AI

2. Mixed-initiative AI

3. Adjustable AI

According to him, autonomous AI operates without human intervention and has no direct impact on people. Next, mixed-initiative AI involves shared decision-making between AI and humans, wherein some tasks are automated while others require human input.

Finally, Mooman states that Great American uses adjustable AI, which helps accelerate decision-making without fully replacing human judgment. In conclusion, he says that while it optimizes tasks for adjusters, underwriters, and customers by providing insights and speeding up processes, the final decision-making is done by humans.

CDO Magazine appreciates Nasser Mooman for sharing his insights with our global community.