The GenAI impact on insurance: Part 1

The benefits of GenAI

• One type of GenAI, large language models, can be useful to help explain complex data sets in terms that humans understand. For the insurance industry, there is a huge benefit in supercharging the efforts of underwriters, strategists, and others. For example, the rich amount of information that can be gleaned from aerial imagery on a single property can be complex. The information across an entire portfolio is even more so. Language models can help reduce this complexity and distill what is most relevant to bring to the attention of decision-makers.
• A particular style of GenAI most relevant to Nearmap is the idea of large vision models. Nearmap has more than 50 petabytes of imagery data. These models are designed to explore the patterns in extremely large data sets and give more traditional modeling techniques a helping hand. They’re an exciting way to get value from a rich historical archive of 15 years of high-resolution imagery and open up a dizzying array of new possibilities.

Finding the right GenAI technology

1. Data accuracy: When used incorrectly, GenAI may appear to provide confident results that are very misleading — it fundamentally attempts to generate data that looks “plausible,” and doesn’t have a built-in way of expressing uncertainty. Tried and tested techniques must be used to create accurate data and provide precise insights. Consider how an insurer assesses property risk. The AI needs to identify and categorise risk based on precise variables such as structural integrity, environmental threats, and historical claims data associated with similar properties. Often the most important thing can be to acknowledge uncertainty on a data point and allow a human to dive in and explore in more detail. This is equally important for both text and imagery.
2. Imagery insights: GenAI techniques can enhance imagery – greatly increasing resolution through “super-resolution” techniques, or otherwise enhancing the quality. There is a risk, however — much of the research is focused simply on what looks good and plausible (for creative industries), and not on ensuring that enhanced fidelity holds to the “truth on the ground.” An artificially enhanced image with more details, and a paragraph of detailed text from a language model, can do equal damage – providing the illusion of certainty through increased detail — when the truth is far less certain. At Nearmap, our mission is accurate truth on the ground, so we must take extra care when working with GenAI in both these scenarios.
3. Handle large datasets: The insurance industry deals with data at a colossal scale — spanning millions of client records and potential risk indicators. A language model that you can interact with may seem very intelligent, but a system that attempts to reason about large data sets must be querying a high-quality database in the background. GenAI provides a great interface to reduce the need for users to do things like write SQL queries to interact with a database, but they are not a replacement for the database itself.
4. Adaptability: The insurance risk landscape is often influenced by factors including climate change, economic shifts, and reinsurance rates. The nature of risk is changing, and models that are a decade old may no longer be applicable. Various aspects of large vision models can be used to help digest the evolving patterns in the petabytes of imagery as we monitor the livable world.