The insurance industry continues to invest heavily in analytics and artificial intelligence. But in specialty lines, where risks are complex, infrequent and often lightly documented, certainty remains elusive.
For Gary Head (pictured), chief underwriting officer at Optio Group, that limitation is not a flaw in modelling. It is a structural reality of underwriting itself.
“The data is the data, and it is the truth,” Head said. “The data is only 100% true if you’ve got an infinite data set.” He questions how often that boundary is acknowledged in industry debate.
“Every day now we hear about data and AI replacing pretty much everything. But my question would be: how many infinite data sets do we have? The answer, as far as I’m aware, is zero.”
Head is clear that modelling plays a critical role, particularly in high-volume classes such as motor, where scale reduces volatility. In those environments, machines are “very adept at both selecting and pricing risk, learning all of the time and edging towards that infinite data set – but they will never get there.”
Large portfolios of homogeneous risks allow patterns to stabilise over time. Specialty underwriting, he argues, operates in a different context.
“My view has always been that underwriting is 50 percent science - by this I mean data and modelling – and 50% luck. What I mean by that is, if you take a tornado risk as an example: two houses right next to each other, one can be untouched and the other one can be completely destroyed. Same risk, really.”
He offers a similar illustration in liability. “The very same risk profile of person can walk across a wet floor in a shopping centre and not fall over, and the person right behind them can slip and have a nasty injury.”
Models may be directionally informative, he said, but they cannot account for every variable within a given risk.
For Head, the dividing line between acceptable and unacceptable uncertainty is straightforward. “If you don’t think you can price and understand the uncertainty, then you don’t write the risk.”
He points to space insurance to illustrate the commercial discipline involved. Underwriters assess historical failure rates, launch location, technical specifications and risk management procedures. Those variables narrow the exposure. The commercial test then becomes one of rate adequacy.
“If I take broker commission off my gross premium, I’m left with, let’s say, £40,000 net risk premium for a £1m indemnity limit. I’m roughly betting that fewer than one in 25 launches - which is 4% - are not successful.”
At the individual risk level, that may be a rational position. At portfolio level, insurers aggregate exposures and model one-in-100-year scenarios to determine capital allocation and reinsurance purchasing.
That modelling, Head explains, allows insurers to allocate capital appropriately and purchase sufficient reinsurance to withstand extreme loss scenarios. Science defines the parameters. Judgment determines whether the return justifies the exposure.
“The more specialist, large, volatile and unique a risk is, the more judgment, experience and risk management come to the fore, and the less data is going to give you the answer,” Head said.
That includes nuclear property and liability risks. “I would back an experienced, gut-feel underwriter every day of the week against a data scientist in predicting the loss patterns for this customer base.”
In such classes, broker engagement becomes materially important. “If the broker can talk knowledgeably about the customer’s positive attitude towards risk and give some clear examples of their commitment to risk management and prevention up front … then that’s a great start.”
Where brokers surface exposures not already reflected in pricing models, “that can be worth a discount in the hands of an experienced and knowledgeable underwriter,” he adds.
Head does not dismiss artificial intelligence. He expects it to improve efficiency and analytical capability. But he remains sceptical that it can replace long experience in volatile classes.
“AI is not going to predict whether or not a rocket ship is going to take off successfully. Because if it did, we wouldn’t need NASA and they wouldn’t have any failures.”
Even minor variables can influence outcomes. “If the grip on your shoe is not recorded and the data is not in the system, then the underwriter can underwrite around that and can give a reduction, for example.”
He notes that carriers’ financial security ratings ultimately provide assurance that claims will be paid. “Whether the underwriting and pricing came out of a programmed machine – which by its nature must have some flaws, or from consideration by a 40-year veteran underwriter who’s seen it all and has the T-shirt is somewhat irrelevant at that point.”
In specialty insurance, modelling frames the exposure and capital sets the boundary. But when volatility is real and datasets are finite, underwriting still depends on the ability to assess what cannot be perfectly measured – and to decide whether the price is worth the risk.
