AI in insurance: Calculating the risks

The AI build-out continues at full throttle, with it being heralded as both a potential job creator as it multiplies knowledge but also a threat to employment as it reduces the need for humans to do mundane knowledge-based tasks. While the potential "use cases" feel infinite, specific uses are often harder to identify. AI, as it relates to the insurance industry, carries with it the prospect of many opportunities: increasing underwriting precision, better identification of fraudulent claims and narrowing of the gap between buyers and sellers of insurance. Most commonly, investors latch onto the first use listed above: underwriting precision. AI-aided underwriting will be smarter than human underwriters (or so some argue). Our view is that this case is overstated. Ideally, insurance underwriting is aiming toward precision at pricing risks, and underwriters and actuaries have been running massive regressions on big data sets for decades. Underwriting ability has indeed gotten more precise, but each incremental layer of underwriting complexity drives model improvements that tend toward increasingly diminishing returns. This is not to say that AI cannot improve underwriting sophistication but rather to say that the refinements will be modest as opposed to transformative leaps. That said, we believe AI-powered underwriting tools will have the ability to improve by a step function the skill of poorer and newer underwriters, narrowing the gap with underwriters currently reaping the best results human intelligence has to offer. 

We also see in the market plenty of examples of AI-assisted customer interactions to augment and speed up the efficiencies of human interactions. Examples include AI-powered claims-handling bots advising claims-department customer-service representatives on how to handle customer requests and comments. AI-aided claims adjusting will be able to compare and source construction, repair and replacement costs in real time and perhaps even negotiate an improved level of claims satisfaction with the customer. Data mining, drone deployment and geospatial mapping can cut down on in-person personnel as well as assessing damages in claim events. We are certainly confident that there are efficiencies to be exploited among insurance carriers and distributors, but we are not so confident that this theme is particularly unique to the insurance industry.

Currently, buyers and sellers of insurance find a broker or an agent in the middle who ties the transaction together. Agent commissions tend to equal a 15% commission charge, through in practice commission rates err toward the high teens as carriers induce agents to place business not necessarily with the carrier whose price represents the best value but often with a preferred carrier offering a more attractive commission arrangement. The cost of the commission payment, ultimately borne by the policyholder, is not transparently disclosed.  

The role of the insurance agent has been durable, and, if anything, the pendulum of value in the transaction has moved more toward the distribution and away from the capital over the past two decades despite the carriers bearing all of the risk. The agent tends to "own" the relationship. It seems that "agentic AI" powered by large language models may be able to provide the support formerly the exclusive domain of the local agent. Such a "digital agent" may even be powered by the comforting voice of Flo or Dennis "Good Hands" Haysbert. Current iterations of AI are already quite good at comparing two (or ten or a hundred) insurance policies side by side and nearly instantly explaining the differences among them. Regarding those who believe that they would never deign to outsource insurance advice to a "talking computer," these must be the same people who have never asked ChatGPT or Gemini a complicated question or currently would refuse to take a ride in a driverless Waymo. By comparison, the needs of a small-ticket insurance buyer — a property policy for a 20,000-sq ft warehouse in Peoria or a business operator’s policy for a florist in a strip mall — are simply not that complicated. 

This is not to say that the globe’s largest companies are going to be buying their insurance from chatbots. We suspect that the market most at risk concerns policies with “ticket” sizes below $75,000. Often the agents who distribute these products know their customers through family, friends, the church, the health club, the civic organization, etc., and they do a yearly recurring cursory commission-paid renewal for that customer. By comparison, an individual's personal insurance bot may constantly be searching the insurance markets "in the background," notifying their "owner" after spotting a carrier willing to duplicate the terms of their current policy at a far more attractive price. We expect "the bot," who won't be charging 15–20% commissions in perpetuity atop the policy, will likely be able to deliver much better value in the long run. 

Is there early evidence that the markets are moving in this direction? There is, effectively, none. Currently, with the exception of a substantial minority of the U.S. personal auto market, almost all U.S. property & casualty insurance transactions are intermediated by agents and brokers. Many carriers have made significant investments into digital sales capability, but the volume of transactions have so far failed to meet many expectations for where commercial direct-to-consumer and digital insurance sales for small commercial and homeowners' insurance would be today. 

At this point, it remains merely speculation, but it may be the case that D2C capability conjoined with agentic LLM-based AI might be the tipping-point event that provides the tool to catalyze the change. When a hypothetical news story hits that a major personal lines insurer has licensed the gecko’s voice for its AI offering, it might be the signal that it is time to listen.