The evolving role of AI in Insurance and Insurtech (Part V)
AI is probably the most hyped word in the last couple of years, and the use cases seem to be intriguing and bold; its role in insurance is at the forefront of the industry. As insurers grapple with complexities ranging from risk assessment to claims processing as partially outlined in Part II and Part IV, AI emerges as a promising technology capable of enhancing efficiency, accuracy, and customer experience. By leveraging machine learning algorithms, natural language processing, and predictive analytics, AI has the opportunity to enable insurers to streamline underwriting processes, personalize policies, detect fraudulent activities, and optimize pricing models.
Evidently, the insurance sector is currently undergoing a significant turning point. Factors such as evolving customer demands, outdated systems, and rising operational expenses necessitate substantial transformation efforts. Adding to this challenge is the widespread accessibility of AI-powered tools and the advent of Generative AI, which are boosting work methods and business operations.
Despite its potential, the implementation of AI and specifically Generative AI demands thorough deliberation on various fronts. Alongside discussing the advantages it brings, it is important to understand the hurdles of adopting AI and Genrative AI and stress the indispensable role of experts in ensuring a successful AI transition. Moreover, it is important to understand that Generative AI will not solve all of the industry's problems, but has the potential to assist in bridging the industry forward.
The history of AI
The history of artificial intelligence (AI) spans several distinct phases since the 1950s. It began with Alan Turing’s proposal of the Turing Test in 1950 and the coining of the term “artificial intelligence” at the Dartmouth Conference in 1956.
Early programs like the Logic Theorist and General Problem Solver marked the birth of AI.
The 1960s and 1970s, known as the Golden Years, saw great optimism and developments in natural language processing and computer vision. However, the 1970s to 1980s experienced the First AI Winter due to unmet expectations, though expert systems emerged.
The 1980s to 1990s witnessed an AI Boom with the resurgence of neural networks and the expansion of machine learning. This was followed by the Second AI Winter in the 1990s to 2000s, characterized by reduced funding and a focus on specific problems.
The AI Renaissance from the 2000s to the present has seen significant breakthroughs due to advances in computing power and big data, with deep learning and neural networks driving major achievements in areas like speech recognition, computer vision, and autonomous vehicles.
AI has undoubtedly been in a state of constant evolution over the past few decades, with significant contributions from various stakeholders, including incumbents and new players like OpenAI.
Over time, AI has evolved from simple rule-based systems to complex machine learning models capable of addressing a wide range of problems.
In fact, a first glimpse into what we see in ChatGPT can be traced back to the 1960s with Eliza, an early chatbot created by MIT professor Joseph Weizenbaum. Eliza enabled basic yet plausible conversations between humans and machines, revealing the potential of AI. This pioneering work by Weizenbaum and the efforts of many who followed in the next decades demonstrate the vast possibilities AI holds for humanity. It also highlighted how human interaction with machines changes when these interactions become more emotionally nuanced and human-like.
With the advent of ChatGPT humanity had an “AHA” moment, as AI became a mainstream phenomenon. Other developments such as language translation and sentiment analysis, once challenging problems in Natural Language Processing (NLP), have now become basic expectations. Recommendation systems (like those from Netflix, Amazon, eBay) and route optimization (used by Google Maps and Waze) were in development and use long before AI became a buzzword. Similarly, computer vision and image recognition, once considered very difficult, are now standard features we rely on when calling for an Uber or a Waymo.
However, for the first time we have achieved a comprehensive understanding of how to construct large language model (LLMs) architectures and possess the global processing power to scale them. While LLMs are currently accurate on average ~90% of the time, there is still a lot work required to get to 99%. With essential resources at our disposal, there is way to evolve this technology through a more traditional, iterative approach.
Most importantly, one question that has been debated is if the world is ready for it.
The current social and economic climates are fostering support for the widespread use of these products, while their development hinges on substantial financial investment.
The segmentation within AI
The segmentation within AI highlights the diverse and specialized branches that have emerged to address distinct challenges and applications. From Analytical AI, which excels at processing and interpreting data, to Generative AI, which creates new and original content, each segment serves unique purposes across various industries.
Moreover, it’s essential to distinguish AI from Machine Learning (ML), Deep Learning (DL), Natural Language Processing (NLP) and Large Language Models (LLMs), recognizing their interconnectedness within the broader AI landscape.
ML, a subset of AI, involves software or machines learning and improving through adjustments to its evaluation based on positive or negative feedback. It paves the way for advanced technologies like Generative AI, which encompasses models such as the Generative Pre-trained Transformer (GPT).
DL, a subset of machine learning, takes AI further by enabling machines to operate with greater complexity by allowing models to quantify the relationships between different inputs. It relies on neural networks to provide deep knowledge to machines.
NLP, a crucial aspect of AI, finds applications in diverse fields, from optimizing search engines to analyzing social media content.
Generative AI, leveraging deep learning techniques, can generate new data based on the patterns it learns from vast existing datasets such as Wikipedia.
OpenAI’s Generative Pre-trained Transformer (GPT) is a type of machine learning that specializes in pattern recognition and prediction and has been further trained using Reinforcement Learning from Human Feedback (RLHF) so that ChatGPT responses would be indistinguishable from human responses.
While AI-driven innovations like Tesla’s Autopilot or Waymos Self-Driving cars showcase remarkable advancements, caution is warranted when relying solely on LLMs for critical decision-making due to the potential for hallucinations. LLM hallucinations refer to instances where LLMs generate information that is incorrect, nonsensically structured, or unrelated to the input provided. You can find more information about hallucinations here.
Foundational Model vs. LLM: Understanding the Differences
Foundation models and large language models (LLMs) are both significant advancements in artificial intelligence, each serving unique purposes with distinct characteristics.
Foundation models are designed to be highly versatile, adaptable for a wide array of tasks including image recognition, language translation, and recommendation systems.
They are trained on diverse datasets comprising text, images, audio, and other types of data, which makes them applicable across multiple domains. Prominent examples of foundation models include BERT, GPT-3/4, and PaLM, which serve as bases for various downstream applications. These models are continuously evolving, with ongoing research aimed at enhancing their accuracy and expanding their capabilities.
Foundation models are large-scale neural networks trained on vast amounts of data and serve as a base for various applications.
Foundation models are characterized by their ability to encapsulate extensive knowledge across multiple domains, enabling them to adapt to a wide range of tasks and applications. Huggingface hosts literally thousands of foundation models available for download. Pre-trained foundation models can be taught an entirely new task with a limited set of hand-labeled examples, showcasing the potential for rapid adaptation and versatility in various applications. Fine-tuning of foundation models can then be applied for greater accuracy.
In contrast, LLMs are a specialized subset of foundation models tailored specifically for processing and generating human language.
They excel in tasks such as text generation, translation, and summarization due to their training on large amount of text data, which enables them to generate grammatically correct and contextually relevant text. Notable examples of LLMs include OpenAI’s GPT-3 and Google’s BERT, designed to handle language-specific tasks with high accuracy. LLMs are more mature and widely used compared to general foundation models, being considered stable and reliable for language-related tasks.
The key differences between foundation models and LLMs lie in their scope of application, training focus, and development stages.
Foundation models are more versatile and can be adapted for a broader range of tasks beyond language processing, whereas LLMs are specialized for language-related tasks. Foundation models are still under active development, whereas LLMs are more established and widely implemented in real-world applications. In summary, while LLMs focus on language processing, foundation models offer a general-purpose framework adaptable to a wider range of AI tasks. The choice between using a foundation model or an LLM depends on the specific requirements of the task at hand.
The current overall state of AI
AI-enhanced technologies have become increasingly accessible across industries — Among these advancements, voice-based assistants stand out as frontrunners, driving AI adoption in diverse sectors like IT, automotive, and retail.
In addition to major players, smaller-scale AI solutions like chatbots have made waves for over a decade by enabling brands to enhance customer satisfaction while saving resources. Moreover, the rise of software-as-a-service models have democratized access to AI tools, further broadening their reach. The emergence of foundation models is rising at an unprecedented speed, with a total of 149 foundation models that were released in 2023, more than double the amount released in 2022. With that the costs for training and deploying these models is getting more and more expensive, e.g. OpenAI’s GPT-4 required an estimated $78 million in compute costs to train, while Google’s Gemini Ultra came with a compute cost of $191 million.
While AI has outperformed humans on several benchmarks, such as image classification, visual reasoning, and language comprehension, it still lags behind in more complex tasks like competition-level mathematics, visual commonsense reasoning, or even strategic planning. Nevertheless, an exciting recent development is that OpenAI’s GPT-4 has largely passed the Turing Test, meaning most people can no longer distinguish when they’re interacting with a machine.
Deep learning models excel in complex tasks, such as virtual assistants or fraud detection, by discerning intricate patterns in data, thereby enhancing accuracy over time.
Mobile devices have emerged as a convenient platform for deploying AI technologies, facilitating a wide range of applications including voice assistants, smart monitoring, personalized shopping experiences, and warehouse management. Generative AI tools have significantly advanced, with companies like Runway producing high-quality generative video models that major studios like Paramount and Disney are exploring for various applications, including lip-syncing and special effects​.
Most of what is labeled AI today is actually machine learning, finding patterns in data and making well-founded predictions.
Emerging technologies like augmented intelligence and edge AI are pushing the boundaries, aiming to amplify human intelligence and enable local processing of algorithms without internet connectivity. AI’s influence also extends into robotics, where multimodal models enable robots to perform a broader range of tasks. This shift from specialized models to more general-purpose ones facilitates more versatile and capable robots​. For instance, DeepMind released Robocat (an update on 2023’s Gato), which can learn how to control many different robot arms by generating its own data from trial and error.
AI Agents or “agentic” systems are more and more emerging, enabling workers in various industries, particularly for workflows involving time-consuming tasks or requiring various specialized types of qualitative and quantitative analysis. AI Agents enable business and workers to move from automating structured tasks that are predictable and repetitive, to performing tasks on behalf of users by mimicking human behavior and intelligence. An especially appealing area for deploying AI agents is where there is a “natural human in the loop” e.g. customer service, compliance, underwriting etc.
Notably, the Artificial Intelligence Index Report underscores the rapid pace of progress, revealing that AI computing power doubles approximately every 3.4 months.
This exponential growth highlights the dynamic nature of AI and its potential to reshape industries at an unprecedented rate.
Evidently, the adoption of generative AI (GEN AI) has surged, with industries increasingly integrating these technologies into their operations. Companies report seeing cost reductions and revenue increases, particularly in areas like human resources, supply chain management, and marketing. For instance, human resources have benefitted from cost decreases, while supply chain management has seen revenue boosts​.
According to Goldman Sachs senior economist, Generative AI will ultimately automate 25% of all work tasks and raise US productivity by 9% and GDP growth by 6.1% cumulatively over the next decade.
However as of mid 2024, adoption rates of AI are much higher among technology industries and other digitally-enabled fields, while insurance carriers and related businesses are only expected to grow modestly. It is also unclear if the surge of AI usage will be just based on Generative AI.
Another fundamental question that emerges is around key supply constraints: Even if AI could potentially generate significant benefits for economies and returns for companies, could shortages of key inputs — semiconductors, chips and specifically power supply — keep the technology from delivering on this promise?
AI within the Insurance Value Chain
Humans excel at analyzing data, but machines have taken this ability to the next level. In the insurance industry, machines can sift through vast amounts of information to uncover patterns for various applications, from detecting fraud and assessing claims to forecasting risk and recommending personalized policies. In the realm of AI this can be categorized as traditional AI, and machines are continually improving in this area.
However, human creativity extends beyond analysis. We are capable of developing innovative insurance products, design engaging marketing campaigns, and craft comprehensive policy documentation. Until recently, machines were not equipped to compete with human creativity, being limited to analysis and routine tasks. Now, with the advent of “Generative AI,” machines are beginning to generate new, meaningful, and effective solutions. Unlike Analytical AI, which focuses on existing data, Generative AI in insurance for instance is about producing original content and ideas, such as personalized policy offerings, predictive risk models, and dynamic customer engagement strategies.
However, in the insurance sector, having transparency and confidence in the decision-making of predictive models is crucial. In fact, insurance regulations play a key role in how AI can pave a way forward. In insurance explainability is an important matter as consumers as well as regulators need to understand how for example pricing is set. Historically, insurers relied on straightforward models that were easily understandable for regulators.
Today, a wide array of machine learning models and techniques — such as linear regression, decision trees, Generalized Linear Models (GLMs), and Gradient Boosting Machines (GBMs) — are employed across the insurance industry to meet diverse business requirements.
These traditional models have been used and worked for decades within the insurance industry as they bring mathematical familiarity and relatability. The logical and understandable structures make deployment easier, while newer AI technologies are more complex and require more time, effort and understanding.
Thinking about the broader applicability of AI and how AI can further enhance insurance practices, it’s essential to delve deeper into visualizing the detailed insurance value chain.
Product Development
Generative AI offers a novel approach to product development within the insurance industry. For instance, in the realm of auto insurance and autonomous vehicles, synthetic data and simulation software can be used to comprehensively test the effectiveness of safety features, handle edge cases, and detect anomalies without real-world risks. By creating complex models that simulate various driving scenarios and conditions for autonomous vehicles, insurers can more accurately assess potential risks. Swiss and Waymo recently published a paper around “Safety Performance of Autonomous- and Human Drivers”, indicating that autonomous vehicles are significantly safer than human-driven ones.
However, a significant question remains: How will regulators adapt to AI-based insurance products, and will they facilitate the integration of AI into insurance product development without further slowing down the approval process?
Sales & Distribution
A predictive scoring model within the distribution channel, designed to forecast the likelihood of a lead purchasing a policy, is an intriguing concept. One of my portfolio companies, Olé, has recently launched a new AI initiative to increase conversion rates by enhancing predictability and efficiency for agents during the quoting process; a first to market in Latin America. With this system, insurance agents can prioritize leads with high scores first, while those with lower scores can be deferred or skipped entirely if resources are limited.
Underwriting & Risk
From improve risk scoring, processing unstructured data from applications, to advanced risk modeling and generating scenarios for risk assessment, the opportunity for advanced AI technologies within the underwriting and risk segment seems to be braod. Evidently, this is also one of the critical segments within the insurance value chain, that has significant implications on the overall value chain, and any adjustments have to be evaluated with utmost scrutiny.
Cytora’s platform for example enhances insurance underwriting by automating risk submission evaluations, converting various document formats into “decision-ready” risks. This process streamlines quotations, reducing response times from days to minutes. It uses large language models (LLMs) and Google Cloud’s PaLM 2 model to extract pertinent information from documents, customizing models for each insurer. Additionally, Cytora employs a detailed risk taxonomy, with predefined schemas for different insurance lines like cyber and commercial property, improving risk assessment accuracy.
Claims & Fraud
Claims processing is one of the most time-consuming activities in the insurance sector. Traditionally, this process involves manual document review, verification of various details, and coordination with multiple stakeholders. However, AI presents an opportunity to significantly reduce this workload. Specifically the use of AI Agents or “agentic” systems can be very intriguing concept in the claims and fraud realm. According to the Coalition Against Insurance Fraud, insurance fraud costs the industry more than $308.6 billion per year.
Moreover, synthetic data enables testing diverse fraudulent scenarios, enhancing machine learning models’ ability to identify various forms of fraud, including rare and complex patterns. By balancing the dataset, these models can detect fraud more efficiently. Additionally, synthetic data enables rigorous testing of models against extreme and evolving fraud tactics, promoting early detection and providing cost-effective alternatives to gathering extensive real-world data.
One of my portfolio companies, Inshur, highlights the transformative impact of AI in claims handling by automating data extraction from various sources and leveraging connected car data for real-time pricing adjustments. This approach makes insurance more accessible and inclusive, particularly benefiting young drivers.
A company applying “agentic” systems to the claims process is Five Sigma, which recently introduced Clive, an AI claims adjuster and extended team member. As an AI-powered insurance adjustment agent, Clive excels at automating and managing tasks traditionally handled by human adjusters across various lines of business.
Some fundamental questions to answer and risks to understand are:
- Data Privacy & Governance: Ensuring the correct and compliant use of data is crucial, especially when retraining LLMs. Challenges persist, particularly in handling sensitive data, copyrighted or patented information and ensuring unbiased predictions. It is an open question whether the data that is currently used to train Generative AI is subject to fair use or must be licensed.
- Regulatory Compliance: Insurance companies must navigate a complex web of compliance regulations that govern data privacy and consumer protection. Challenges persist with the use of LLMs to understand how data is used and decisions are made.
- Bias and Discrimination: There is a risk that LLMs can perpetuate biases present in training data, leading to discriminatory outcomes.
- Ethical Consideration: Heightened awareness of AI risks, especially in generative models, emphasizes the importance of cautious implementation and ethical considerations to ensure the safe and responsible integration of AI in insurance practices.
- Misinformation — Hallucinations: The potential for LLMs to generate persuasive but inaccurate information poses risks for both organizations and consumers.
The AI race between Incumbents and Insurtech’s
Historically, innovation started outside of big organizations, where processes, policies, and speed were significant hurdles to creating new things. While technically this still holds true, in the dynamic realm of AI, collaboration between incumbents and emerging players continues to drive meaningful change, signaling that no one is “asleep at the wheel” in this era of technological innovation.
As highlighted in Parts I and II, the insurance industry has undergone a substantial transformation over recent decades. Initially, traditional insurers were hesitant to embrace newer technologies. However, as detailed in Part II, Insurtech v1.0 companies successfully managed the complex task of persuading established insurers, service providers, consumers, regulators, and the broader financial markets that digitizing insurance was inevitable. Their mission was to underscore the imperative inclusion of advanced technologies in the future of insurance while outlining a viable path for venture backable companies. With the arrival of AI, incumbents have become more proactive in experimenting with and deploying these new technologies. Yet, it remains too early to predict the ultimate outcome.
From a technology perspective the breakthrough in AI is undoubtedly important and will eventually have enormous impact on our lives and society in general. Some view the current wave of AI as a transformational, generational technology, comparable to the advent of the internet, which certainly has a lot of merit. However, it’s important to look beyond the hype as the influx of startups and venture funding in the AI space can create a slippery slope.
From a venture perspective, the sentiment around AI and related technologies is widespread and can maybe be labeled “bubbly and crazed”. In fact, according to Pitchbook, currently AI is artificially inflating valuation numbers across the venture capital industry, with the sector capturing nearly 50% of total deal value just in Q2/2024. On another note, 70% of Y-Combinator’s Winter 2024 batch is made up of AI startups. By contrast, just 57% of the Summer 2023 companies and about 32% of the Winter 2023 batch were AI-focused. In the insurance sector, for instance, many startups are targeting similar issues — sometimes problems that are too niche — while incumbents are already implementing in-house solutions.
Additionally, a significant factor to consider is that many startups focused exclusively on software enabled AI solutions for insurance, face a narrow path to significant growth and eventual exit. In fact some startups in the claims and fraud category have been around for over a decade, with lofty valuations and significant venture funding. However, what is unclear is if those valuations can be directly linked to significant growth and revenue figures. Looking ahead, with many incumbents already deploying their own solutions or adopting other solutions, the key question is whether these incumbents will engage with multiple startups or if the market will lean towards a “winner takes most” scenario.
In Part VI of Insurtech and the captive insurance market, I take a deep into the captive world and how Insurtech’s can play a role. If you have an Insurtech prediction or want to be included, please reach out at amir@overlook.vc or on twitter @AmirKabir99
Amir is a has been one of the earliest investors in the Risk Management & Risk Transfer space.
This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future. (I have copied the disclaimer used by Jamin Ball in his great newsletter Clouded Judgment)
