A focus on machine learning,
deep learning, and generative AI
Whether you are an early adopter of AI technologies in your workplace,
trying to avoid them, or somewhere in between, learning more about AI can help.
By Lisa A. Gardner, Ph.D., CPCU, AIC, AIDA, API
Few subjects are grabbing headlines faster than artificial intelligence (AI) or technologies that support computers and machines in performing specific, complex tasks usually done by humans. While AI technologies cannot fully mimic most of the human brain’s capabilities, they can mimic some, like understanding and responding to human language, learning from new information, and seeing and identifying objects, to name a few.
Some of the better-known examples of current AI usage in insurance and risk management include using chatbots to respond to customer inquiries, using creative tools to develop content for product marketing, relying on drone images to evaluate storm damage to buildings, creating algorithms to identify potential fraudulent claims or likely insurance buyers, and using sensors to engage anti-collision devices like anti-lock brakes.
As AI technologies evolve, so shall the applications and devices that incorporate them. Whether you are an early adopter of AI technologies in your workplace, trying to avoid them, or somewhere in between, learning more about AI can help.
This article gives an overview of AI technologies and applications in insurance. The focus here is on the most prominent subset of AI fields affecting insurance: machine learning, deep learning, and especially generative AI, which are rapidly transforming insurance work.
Machine learning
Many of the earliest applications of AI in insurance arose from machine learning, a subfield of artificial intelligence. Considered foundational to many AI capabilities, machine learning requires large historical datasets to train algorithms. These datasets may contain text (e.g., customer records), numbers (e.g., premium payments), sensor information (e.g., smoke detectors), images (e.g., pictures of windstorm damage to roofs), audio recordings (e.g., recorded claims statements), and/or clicks, depending on what you want to accomplish.
Developing machine learning algorithms often requires following a fixed set of instructions or procedures, also called training. Data scientists train algorithms using one dataset and then apply them to another to make forecasts or predictions, classify items, or make rational decisions.
Examples of insurance applications of these capabilities encompass forecasting auto insurance premiums or discounts, classifying customers into segments to better serve their needs, and to support decision-making in claims adjudication.
Deep learning
Deep learning, a popular subfield of machine learning, simulates decision-making using (artificial) neural networks. Modeled after the human brain, deep neural networks consist of hundreds of layers of interconnected nodes called neurons. These neurons work together to process data and make decisions.
Capable of quickly processing enormous amounts of unstructured data like images, text, and sound, deep neural networks can automatically learn patterns and complex relationships within a dataset without requiring human intervention. This “learning” or adjusting to improve performance occurs using experience or data.
Usage-based insurance and autonomous vehicles represent applications of deep neural networks under development, where information about experience updates the performance of each, bringing both closer to more widespread acceptance by regulators, legislatures, and consumers.
Generative artificial intelligence
Within deep learning, a specialized subfield called generative artificial intelligence (Gen AI) accounts for much of the recent innovation and attention given to AI. Generative AI tools can produce text, images, sound, video, synthetic data, or computer code in response to a written prompt. If you have used Google’s Gemini, OpenAI’s ChatGPT, or the Risk & Insurance Education Alliance’s Allibot, you have used a product powered by generative AI.
Unlike many AI tools, Gen AI creates, but not out of thin air. Large data sets and user input make such creations possible.
Most Gen AI tools specialize in one or two of these capabilities: content generation, coding generation, design ideation, or audio and video creation.
Of these, content generation may be most relevant to Rough Notes readers.
Gen AI content creation tools can create marketing materials like social media content, blog posts, and email campaigns tailored to different customer segments, saving time and boosting productivity. With credible data, they can serve as knowledge resources and generators of study materials for professional agents and adjusters licensing exams, helping to level up employee knowledge.
They can extract text from documents (e.g., a medical diagnosis from a record), sort customer complaints, and compare documents (e.g., endorsements), reducing time allocated to these tasks and costs. They can also summarize customer needs, claims, policies, and reports, all to facilitate faster decision-making.
Chatbots created using Gen AI technologies can generate instant, automated responses to basic customer questions. Available 24/7, chatbots improve communications by providing consistent, direct answers to questions, helping improve customer experience, reducing claims processing times, and making operations more efficient.
Questions to ponder
Machine learning, deep learning, and generative AI tools can help make the workplace more productive and reduce costs. When implemented well, they can also help free up time to interact with clients, an essential consideration for the relationship-centered insurance business.
However, decision-makers cannot solve most of the challenges in the insurance business by deciding to implement another AI tool. With this caveat in mind, think about these four consequential questions when considering implementing more AI technologies:
- What do you want to accomplish using AI? Be clear about your goals and do your research to know whether AI technologies exist that can help you achieve them. Are your goals to reduce costs, make operations more efficient, or increase your workforce’s insurance knowledge? If so, AI technologies exist that, when implemented and supported effectively, may help you achieve these goals. However, AI technologies cannot replicate or replace the caring, empathy, and support that a customer experiences from agency personnel after a loss.
- Will you buy or develop the AI tool that you want? Developing an AI tool includes risks. Development costs that overrun your budget, an unexpectedly long development period, and the technology becoming obsolete before it is available can diminish the value of building your AI tool. However, the tool may better suit your organization’s needs than off-the-shelf tools.
Buying a tool also entails risks. Artificial intelligence results depend heavily on the data used to create them. Irrelevant or biased data contributes to irrelevant or biased results. Knowing data sources can help determine whether results seem biased.
Not all companies fully disclose data sources that power AI technologies, but some do. For example, the Risk & Insurance Education Alliance discloses data sources, all insurance-specific (e.g., CIC, CISR, and CRM course content, ISO forms, state insurance laws and regulations) that feed its AI-powered chatbot, Allibot 3.0.
- How will you train users? Some AI technologies (e.g., machine learning tools) require many hours of training for users to use them well, while others require less time (e.g., some Gen AI tools).
Because many of these technologies require or undergo regular updates, managers must schedule training to help employees remain current. Such training empowers employees to do their jobs well and helps ensure they use AI tools as intended.
Failing to create an implementation plan that includes training would be a mistake. Even tools requiring the user to type in a question to generate an answer require training. Remember, what makes sense to you as a question may not make sense to computer software.
With any technology, user training and experience help make implementation successful. If your organization cannot build a training program, some AI vendors can help.
- How will you know if you have achieved your hoped-for outcomes? As part of your plan to implement AI, incorporate ways of measuring success so that you know if you have achieved your goal.
Some goals, like improvements in sales conversions, customer satisfaction, or customer retention rates, are easier to measure. Others require more effort, like time saved in performing specific tasks, error reduction, and the degree of AI tool adoption in your workforce.
Regardless of the effort required, having reliable metrics can make AI implementation more successful.
Summary
Thanks to researchers, computers can mimic certain human brain functions. Artificial intelligence applications can recognize objects, warn of dangerous conditions, and direct motor vehicles to move along specific paths.
They can identify consumer spending patterns, predict fraud, and correctly sort individuals based on attributes. They can also generate images, sound, code, and text, make comparisons, and comprehend written and spoken language. Despite these advancements, AI abilities, including in insurance, remain limited compared to the human brain’s remarkable capabilities.
Insurance enables many everyday aspects of living, from owning or operating a motor vehicle to accessing medical care, operating businesses, and planning for the unthinkable loss of a family member. When losses do occur, they can be devastating.
Whether you work at an agency or as an adjuster, helping insureds navigate what may be some of the worst days of their lives remains a distinctly human task, involving empathy, caring, kindness, and honesty at its best. No technology replaces the value of human connection in our lives.
The author
Dr. Lisa Gardner is the associate director, Content and Research, at the Risk & Insurance Education Alliance. Her experience includes more than three decades as a professor in higher education where she taught courses about risk management and insurance, statistics, and finance, and provided programmatic leadership.
