Home insurance: Will AI kill risk pooling through hyper-segmentation?

Can artificial intelligence and data fragment insurance to the point of destroying the very principle of risk pooling? This is one of the central questions of our ebook “How can AI and data save the insurance industry? The example of home insurance in France”, written by Addactis experts. This article is its first part: it revisits the technical foundations of risk pooling, examines the excesses of certain pricing practices, and shows how, in the face of rising climate risks, AI can become a tool in service of fair segmentation and preserved pooling.

In the age of data and AI, some fear that technology will push insurers toward hyper-segmentation, undermining the principle of risk pooling that underpins insurance.
By enabling a detailed understanding of risk, data and AI would provide insurers with a new opportunity to segment their portfolios, adapt and “granularize” their pricing policies based on individual risks, and thus undermine insurability through underwriting rules or individualized pricing.

This dystopian vision threatens not only to exclude the most vulnerable or those most at risk, but above all to break both the social contract of insurance and its technical foundations. It would effectively exclude an entire segment of the population from the benefits of insurance (those who need it most) and require others to pay a premium equivalent to the cost of their risk (in a financial approach akin to life insurance with a single pooling over time, where everyone knows they will die, with only the timing remaining uncertain).

It is worth recalling here that all insurance operations are based on the dual principle of risk pooling and segmentation. Insurance is, at its core, a pooling mechanism into which members contribute (a segment) so that those who suffer a loss can be compensated by drawing from the accumulated pool (the pooling of risks previously defined through compensation for incurred losses). Members of this community may agree to pay for management in proportion to the service provided, with any surpluses returning to them after a specified period, just as they are responsible for contributing to the pool if it becomes insufficient.

Risk pooling occurs over time (the hazard and its randomness), across the probability of occurrence (exposure), and the consequences if the risk materializes (vulnerability), and applies to a homogeneous group of policyholders with similar risk levels or who, at a minimum, agree to pool their risks among themselves for a given period. All insurance products are thus first segmented into homogeneous risk groups, if only in terms of coverage and rates, and then seek to broaden their scope to further disperse risks.

Risk pooling and segmentation are therefore complementary and inseparable, with segmentation serving to preserve pooling within a segment defined and accepted by the insured.

Furthermore, over the past thirty years, insurance has become increasingly financialized worldwide, introducing strict profitability constraints in a context of heightened competition. In principle, premiums must not only cover claims and administrative costs but also finance development, marketing, and advertising, and above all, ensure solvency and provide a return on equity. In this fundamental equation of insurance, risk pooling has at times given primacy to issues of profitability and return on investment, with several fundamental technical consequences, notably the implementation of hyper-segmentation and the emergence of pricing optimization systems.

‘Price optimization’, for example, based on customer willingness to pay, consists of determining not the actuarially fairest pure premium for the segment, but the most advantageous premium for the insurer, taking into account what the insured is willing to pay for their policy or competitors’ premium levels. This practice, which runs counter to risk pooling by creating quasi-systematic discrimination against existing policyholders in favor of new customers, is increasingly being restricted or prohibited by regulators.

With data and AI, the insurer must return to its fundamentals as a player in the sharing economy: knowing, understanding, and managing risk in order to underwrite it and organize a sharing arrangement with policyholders, taking preventive or mitigating action.

The real challenge, therefore, is achieving the right segmentation, based on actual risk, that ensures risk-sharing aligns with the insurer’s “policy” and strategic choices, incorporating varying degrees of solidarity-based or equitable rules.

The example of comprehensive home insurance in France demonstrates that AI and data will, on the contrary, help ensure the long-term viability of risk pooling and the very principles of insurance, in a sector and type of insurance that have nevertheless been severely disrupted by the effects of climate change and natural disasters. This is because climate risks, which accounted for 10% of the pure premium 10 years ago, now account for approximately 30%. And they are becoming one of the main sources of vulnerability in the insurance market. When climate change transforms risk from a probability into near-certainty, it calls into question the very principle of insurance; the only possible pooling of risk then being the smoothing of expenses over time. In its late 2023 “UN Global Risk Report,” the UN listed uninsurability among the “six perils threatening humanity and the planet,” warning that as natural disasters increase, insurers could withdraw, rendering certain areas “uninsurable.” The report did not merely refer to isolated disasters, but warned that if insurance systems, essential to societal resilience, gradually ceased to function, this would lead to a collapse or breakdown of the entire system.

And even without considering climate-related risks, home insurance in France is nonetheless not immune to the other numerous and growing challenges facing all insurance models worldwide. The challenge lies in operating a model based on a very precarious balance between, on the one hand, risk costs that are increasingly accurately known thanks to AI and data, and on the other hand, multiple constraints ranging from the social acceptability of risk pooling to the growing demands of regulators, not to mention the necessary profitability of the portfolio.

While models based on historical data remain the gold standard of the technical state of the art, we are witnessing a daily shift toward AI and data, which alone prove capable of modeling the occurrence of events without usable historical data and the vulnerability of buildings with a high degree of precision.

Every insurer today knows that external data can already replace historical data when underwriting new business. In the case of home insurance, historical data alone will be of little use when taken in isolation to assess the risk of a once-in-a-century storm, as the vulnerability of each building will vary greatly when faced with this unprecedented risk for the insurer. Conversely, having vulnerability data without building exposure data would be just as limiting.

Thus, in Brittany, a century-old house well-oriented relative to the winds and built to withstand the elements could prove more resilient in the face of a storm than a newer house with a more original design but not inherently built to withstand exceptional storms. Conversely, a new house could benefit from significant advances in construction if it fully complies with the latest standards in the field. Yet these three houses are pooled together in the insurer’s portfolio, and the difference in vulnerability is not necessarily reflected in the premium.

Risk cost models today are based on:

• Hazard models with event scenarios based on location, intensity, and frequency of occurrence;
• Calculations of the exposure of insured properties in the portfolio, insured values, policies in force, and risks covered;
• Vulnerability assumptions to quantify damages based on the hazard;
• A comprehensive financial model to translate the previously calculated damages into potential claims costs according to insurance terms (and the impacts of reinsurance).

While understanding which areas are exposed to various climate-related hazards and how these events are evolving as a result of climate change (particularly shifts in the return periods of different intensity levels) is certainly essential, managing the specific risks associated with each property in a portfolio in the face of these hazards becomes even more critical.

In a recent white paper titled “Is the Geographic Exposure Approach Still Sufficient for Insurance?”, Addactis demonstrated that mapping risks and territorial exposure levels must necessarily incorporate detailed, localized knowledge of the intrinsic vulnerability of each insured building.

The predictions of Addactis’ AI models have been confirmed by the results observed across a large number of portfolios. They reveal significant disparities between different cost estimates, as discrepancies frequently range from single to double the amount, underscoring the need for an in-depth understanding of each property. The building-level approach (and not just the address, since the discrepancy between the address and the building can have significant impacts on certain risks, such as flood risk) fundamentally improves the accurate understanding of exposures and the actual risks incurred during a climate event.

It is precisely thanks to AI and external data on hazards, exposure, and vulnerability that today’s insurers will be able to both preserve the insurability of properties across the country and safeguard the sustainability of their insurance models based on varying levels of solidarity and risk pooling. AI thus becomes a powerful tool for managing and controlling risk pooling in home insurance.

The opposition between risk pooling and segmentation is a false debate: the two have been inseparable since the origins of insurance. What is at stake today is the quality of segmentation. Insurers who integrate granular building-level risk knowledge, through AI and data, will be able to manage their portfolios with fairness, precision, and resilience in the face of climate hazards. Those who forgo it expose themselves to silent, progressive adverse selection.

To go further, discover our ebook“How can AI and data save the insurance industry? The example of home insurance in France”, written by Addactis experts.