MTech Capital

Recent technological breakthroughs in Generative AIhave the potential to transform the insurance industry. At itscore, insurance is the ultimate data-driven business. Every day, insurers and brokers receive and generate massive quantities of data via email, call centers, PDFsand spreadsheets. This data, often granular and proprietary to the insurer, has never been comprehensively analyzed. But the emergence of large-language-models (LLMs) presents an opportunity for the industry to finally change that.

Widespread commercial access to foundational LLMs like OpenAI’s GPT-4 is still very new and innovation is accelerating, with breakthroughs appearing on an almost weekly basis. Up to now, training machine learning algorithms required enormous amounts of data to perform a specific task. With the recent breakthroughs in foundational LLMs,the additional training or ’fine-tuning’of LLMs to perform the same tasks can be accomplished with much less data, speeding up deployment.

When considering how this might play out in the insurance industry, there are still more questions than answers:What are the killer use cases?What is the best operating model for insurers to take advantage of this technology? Will startups build verticalizedLLM applications? Will established software companies like Guidewire succeed in embedding LLMs into existing products?Will carriers go direct to the source and build their own products on top of foundational LLMs?

To form a view on these questions, we’ve been having conversations with leaders across the industry, including large carriers, established tech vendors, multi-line insurance brokers, and insurtech startups of all shapes and sizes. Here are somepreliminary findings:

Most promising use cases

Self-servicetools–Conversational AI hasbeen available for several years.However, deployment of virtual assistants (i.e., ‘chat bots’) to-date has been limited to the simplest of servicing use cases. This has often caused significant frustration among consumers who find they give all the information to a chat bot and then must repeat it again to a human agent in a call center when the chatbot fails to provide a solution.

Modern self-service tools harnessing the power of LLMs such as Parloa and Replicant can engage in more nuanced, empathic conversations with customers, as well asread and write to core systems and complete more complex servicing tasks without human supervision. For example,today’s models can change policy effective dates, named insureds, and other minor coverages/endorsements.For low severity claims, customers can now complete the entire FNOL process by engaging exclusively with text and voice-based conversational AI assistants.

Document understanding – Given the insurance industry’s reliance on PDFs and spreadsheets, document automationhas long been viewed as a ‘killer use case’ for AI. Butprevious generations of this technology (e.g., OCR, NLP, and RPA) have primarily been useful in extracting unstructured data from documents that arrive in consistent formats. These solutions often struggled with documents where the formatting is inconsistent or the data is nested in complex tables.

The latest product releases from companies like Instabase and Indico use cutting edge deep learning techniques and the latest LLMsto help insurers tackle long standing, complex problems - with more limited training data sets - such as commercial insurance submission intake and bodily injury/medical claims triage.Beyond extraction, these tools give users the ability to converse with documents (e.g., ‘what is the policy number on this page’?) andgenerate summaries (e.g., ‘summarize all of the notes related to this claim file'),enabling more complex workflow automation.

Co-pilots – The holy grailof AI in insurance would be applications thatcreate a sustainable competitive advantage in the core functions of the insurance value chain, namely underwriting and claims.Carriers have long adopted rules-based rating engines to automate underwriting and pricing in personal lines and small commercial business. Generative AI ‘co-pilots’ like Sixfold and Capitola could enable carriers and brokers to bring this level of automation and discipline into more complex business lines such as mid/large commercial risks, specialty products, and high-value life insurance. While these lines of business will certainly still require human input and judgement, LLMs combined with third-party data can help dramatically improve the analytics processes that inform these decisions, enablingdynamic optimization of a carrier’s book of business — based on, for example,target loss ratio, retention, and growth goals.

Buy vs. build?

Our initial hypothesis was thatmost carriers would need the help of third parties to scale production-grade LLMs across their businesses. That said,we have been pleasantly surprised to learn aboutseveral insurers in our network already experimenting directly with foundational LLMs – and indeed several carriers are mentioning LLMs in earnings calls. Some are achieving positive early resultson use cases such as call center and claims note summarization/ triage.

"Widespread commercial access to foundational LLMs like OpenAI’s GPT-4 is still very new and innovation is accelerating, with breakthroughs appearing on an almost weekly basis."

We anticipate there will be stratification in the market betweeninsurance companies that have the resources to ‘build’ LLM use cases internally, namely the largest and most sophisticated carriers with hundreds of data scientists and engineers in house, and those that don’t.However, we believe the playing field for the application of Generative AI across the insurance industry will be relatively flat given how readily available these tools are - even if SMEs need third-party help to deploy them.

Therefore, we believe there will be a significant opportunity for all players in the insurance technology ecosystem to create and capture value from LLMs –from existing software vendors and system integrators/IT consultancies to insurtech startups with LLMs embedded into their products from day one. Based on what we have seen so far, we view LLMs as tools or features that can help create exceptional products rather than being ‘the product’ itself.

Healthy skepticism

The excitement about the potential impact of Generative AI in insurance should be balanced with a healthy dose of skepticism and practicality. There are several considerations carriers (and investors) should take into account when exploring LLMs in insurance:

• Safety and data security – especially with personally identifiable information (PII) and claims data,

• misplaced confidence - LLMs are very capable of providing authoritative sounding but inaccurate answers. Thesehallucinations can be especially dangerous when answering policy and coverage related questions that have a definitive answer (and expose carrier/broker to liability if answered incorrectly),

• the pace of change- do organizations have the ability to continuously maintain and upgrade to the latest tech,

• compute cost – cost and availability of GPUs will remain a gating factor for adoption across industries, and

• data interoperability (or lack thereof) across legacy systems and between different stakeholders in the value chain will ultimately limit the most grandiose use cases from coming to fruition until more modern underlying infrastructure is in place.To paraphrase Hemingway, we see the impact of Generative AI on the insurance industry to be slow at first, then all of a sudden.