Introduction

Big Data and Artificial Intelligence (BDAI) have become widely discussed buzzwords that are associated in the public mind with a strong force for disruption and a high potential for changing existing business models. In particular, they are increasing the opportunities for individualising products and prices, for contacting customers and for automating every possible business process.

These trends are also affecting the insurance industry. Among other things, they are resulting in potential applications in risk selection, (individualised) premium calculation and rate setting, raising questions above and beyond the areas of data security and data protection. At the same time, aspects relating to the reliability and admissibility of Big Data applications and possible discrimination against policyholders with higher risks due to their personal life circumstances are being discussed. Last but not least, many people are wondering whether and to what extent the balancing of risks in the community of policyholders, and hence the principle underlying insurance, still works when premiums are calculated and rates set individually.

To come straight to the point: individualised insurance premiums that can be calculated with far greater precision based on Big Data and Artificial Intelligence do not generally harm the community of policyholders and the balancing of risks – in fact the opposite is the case. The precise impact of these digital and technological trends on the balancing of risks in the community of policyholders is analysed in the following.

Definition of insurance

“Insurance provides cover for funds that are needed and whose specific amount is uncertain but are estimated in the aggregate on the basis of the balancing of risks in the community of policyholders and over time”

Dieter Farny¹

This definition encompasses both the key characteristics and the actual business model of an insurance undertaking. An adequate understanding of the concept of private-sector insurance cover requires a more detailed explanation of the individual components of this definition.

“Cover for funds that (...) are estimated in the aggregate.”

Stochasticity is a key feature of the insurance business. In simple terms, stochastics deals with probability distributions of events and their outcomes.² Based on past experience and the law of large numbers³, probability distributions with expected values and scattered outcomes of random events are projected into the future. The more individual values that are used for the calculation, the more reliable the probability distributions will be. One example is the outcome when dice are thrown. Although the individual outcome is determined by chance, the expected value for the number of pips is always 3.5 and the distribution of the outcomes ranges from 1 to 6. The probability of reaching 3.5 pips with only a single throw is zero. However, the more often the random experiment is repeated with an average then being calculated, the more reliable the estimate will be of obtaining an outcome close to the expected value of 3.5.

Most insurance business covers the economic consequences of the occurrence of undesirable real-world events, i.e. perils. Comprehensive contents insurance⁴, for example, covers the perils of burglary, water damage, storm and fire. The economic consequences are termed “risks”⁵ and are losses measured in cash or funds needed to finance the events that have occurred.⁶ The individual policyholder thus receives a compensation payment specified in the insurance contract if one of the insured risks and hence an economic risk materialises. However, because this depends on chance, the benefit to be paid is not known in advance. Based on statistical historical values and taking risk factors (e.g. residential address, design) into account, expected values are calculated for the frequency and amount of claims in a portfolio of insured risks, and the total funds required to be covered by the insurer are estimated. As in the example of the dice, the more random experiments a community of policyholders consists of, i.e. the more risks (= individual insurance relationships) that are included in the portfolio, the more accurate the estimate will be.

A reliable estimate, however, requires a substantial degree of homogeneity of the insured events or risks. That is because an estimate of average losses from heterogeneous risks at Insurer A would lead to an inflated calculation of the funds required for the risks with a below-average exposure to losses and thus also to inflated insurance premiums. Conversely, risks exposed to above-average losses would be calculated too low and underpriced. Under competitive conditions with a competitor B who calculates correctly, and with corresponding market transparency, the consequence would be an anti-selection of excessive risks at average premiums at Insurer A.

“whose specific amount is uncertain”

A functioning insurance business requires the occurrence of each individual insured event to be uncertain; if not, a fundamental condition for the insurability⁷ of risks will not be met. It does not make sense to insure an event that is certain because the risk premium would have to match the amount of the (certain) loss; administrative expenses would additionally be incurred. And an event that will definitely not happen does not require any commitment to provide cover. There must therefore be a random factor affecting the occurrence of a loss. This means that both the policyholder and the insurance undertaking do not know in advance whether, in what amount and/or when the individual loss will occur. To avoid information asymmetry and hence both sides trying to outsmart each other when insurance decisions are made, the degree of uncertainty on both sides should be roughly the same. In reality, though, the insurance business is strongly characterised by information asymmetries.⁸ As a rule, the policyholder knows their personal risk much better than the insurer and can even influence it. Insurance undertakings can reduce such information asymmetries with the help of application questions, inspections and (in the personal insurance business) medical examinations and obligations.⁹ On the other hand, their extensive statistics also give insurers an information advantage about the probability distributions of losses, which they could play off against policyholders in the form of inflated risk premiums. However, since price transparency increases in a time when comparison portals and online brokers have multiplied, this sort of behaviour would tend to be punished by the mechanisms of the competitive market.

Balancing risks in the community of policyholders

The primary focus of an insurance contract is the transfer of a probability distribution¹⁰ of losses from the policyholder to the insurance undertaking. Since an insurance undertaking covers a range of loss events affecting a large number of policyholders, it also has a large number of risks in its insurance portfolio that balance each other out.

In this case too, the example of the dice can be used to explain how risks are balanced in the community of policyholders. With a fair dice, the probability of all possible dice outcomes, i.e. throwing the different numbers of pips 1, 2, ... 6, is the same. Since a range of random events (in this case individual dice throws) result in outcomes that are independent of each other, higher and lower dice outcomes will tend to balance each other out in relation to the expected value, and the average outcome will settle at 3.5 in accordance with the original estimate – and the more frequently the dice is thrown, the more reliable this will be (law of large numbers, see above). The same applies in real life to insured risks. However, another condition is that the risks must be independent of each other. If there is no independence, for example in the case of storm damage within a defined region, risks impacted by claims and claim-free risks can no longer compensate each other and risk balancing fails.

Whereas the homogeneity of the individual risks is therefore of great importance for the estimation community (see above), this does not apply to the balancing of risks in the community of policyholders. This is where there is often a misunderstanding about the principle underlying insurance. Heterogeneous risks, measured in terms of expected claims, can also balance each other out provided they are backed in each case by the right insurance premium. For example, insureds of different ages and hence different mortality probabilities can also be included in a community of term life insurance policies if, under otherwise identical circumstances, older insureds pay a higher premium – commensurate with the risk – than younger insureds. The balancing of risks will not then be disturbed. Before this can happen, however, estimating the risk and thus calculating the premium has to be differentiated in line with the age of the insureds.

Balancing risks over time

Balancing risks in the community of policyholders only ever has a limited observation period and typically never functions perfectly. Within a period, there will regularly be above- or below-average claims, although these will also tend to offset each other over time. In other words, periods with below-average claims balance out periods with above-average claims. This explains how risks are balanced over time; at the same time, it requires and shapes the long-term nature of the insurance business. The ideal conditions for balancing risks over time are constant characteristics of all individual risks in the insurance portfolio as well as no change in the portfolio composition. This means that, as far as possible, the expected claims and the loss distribution are not subject to any deviation risk, or that the deviation risks will tend to balance each other out within the overall community of policyholders. In practice, however, the overall community of policyholders is rarely constant over time, as the insurance portfolio is characterised by inflows and outflows as well as a changing risk environment. Because of the limitations on risk balancing in the community of policyholders and over time, there is always a residual underwriting risk for the insurance undertaking.

Underwriting risk

Underwriting risk describes the possibility that the effective claims in the community of policyholders will exceed the expected claims in the community, thus exposing the insurer to the risk of loss or even ruin, as the risk premiums collected may not be sufficient to pay the losses incurred.¹¹ In terms of cause, underwriting risk can be broken down into the risk of random fluctuation, change and error.¹²

The risk of random fluctuation occurs when effective claims differ from expected claims because an above-average number of losses and/or a high level of losses have randomly occurred.¹³

The risk of change is based on the possible case that the risk conditions will undergo an adverse change over time compared with the assumptions originally used to calculate the premiums, and that the effective claims will therefore ultimately exceed the expected claims.¹⁴

What risk of error means is already apparent from its name. Risk of error materialises in the case of incorrect estimates or assumptions.¹⁵

How insurance works: the actuarial principle of equivalence

The conceptual basis for calculating risk premiums¹⁶ is the actuarial principle of equivalence. Under the actuarial principle of equivalence, the risk premium corresponds to the expected claims relating to the covered risk.

A further distinction can be made between community and individual actuarial principles of equivalence. In the case of premiums calculated according to the individual actuarial principle of equivalence, the amount of the risk premium payable by the policyholder is equal to the amount of their individual expected claims. Under the community actuarial principle of equivalence, the aggregate risk premiums from an insurance portfolio correspond to the aggregate expected claims for the community. The individual actuarial principle of equivalence includes the community actuarial principle of equivalence, as it were, because if each policyholder pays the risk premium for the share of their expected claims in the overall community, the sum of the individual risk premiums will also correspond to the aggregate expected claims of a community of policyholders.¹⁷ Conversely, the risk premium for the community divided by the number of policyholders does not have to equal the individual expected claims. In the case of heterogeneous risks, there is then an average premium for different risks, i.e. some risks with an above-average loss exposure and some risks with a below-average loss exposure for which insurance premiums offering insufficient coverage or excess coverage would be required – with the consequences of anti-selection, as described in the following:

Let us assume that several insurance undertakings operate in a transparent insurance market under competitive conditions. In this market, Insurer 1 charges risk premiums that are aligned with the individual expected claims, and a second insurer, Insurer 2, charges average risk premiums for its insurance portfolio. In this case, utility-maximising rational policyholders with below-average risks (i.e. policyholders with lower expected claims than the average) take out insurance with Insurer 1 and pay correspondingly lower risk premiums. Policyholders with above-average risks (i.e. policyholders with higher expected claims than the average) will opt for Insurer 2 and its average risk premium, as this is lower than their actual expected claims. If Insurer 2 does not align its risk premium quickly enough with the new composition of its community of policyholders, it runs the risk of ruin.

As things stand today, a risk premium can at best be calculated approximately based on the individual actuarial principle of equivalence. A greater degree of approximation is not possible at present because there is no reliable data pool, and nor are there adequate IT processes. The data pool basically comprises information about risk characteristics that must essentially meet three requirements:

1. There must be a statistically significant correlation between the values of the risk characteristics on the one hand and the shape of the loss distribution¹⁸ (with the expected claims and the loss distribution) on the other.
2. The relationship must be plausible so as to rule out spurious correlations.
3. The insurance undertaking must be in a position to operationally capture the values of the risk characteristics.

Capturing them is made easier today because of the preponderance of “objective risk characteristics”¹⁹ whose values tend to be easily and reliably obtained from external sources. Examples in motor vehicle insurance are the type class of the automobile, the residence of the owner (regional class), how long the driver has held a licence, etc. With over 100 risk characteristics used throughout the market, motor insurers are already in a position today to break down the actuarial principle of equivalence into the smallest rate cells containing only a few – in some cases only individual – risks. The number of rate cells that can be formed with the given risk characteristics according to the rules of combinatorics is probably significantly higher than the market-wide number of insured risks. The sum of the individual premiums calculated in this way, when extrapolated for the community of policyholders, is likely to result in a total premium that is highly correlated with the actuarial principle of equivalence. However, the individual actuarial principle of equivalence still remains imperfectly satisfied. What is missing above all is the collection and processing of subjective risk characteristics that represent in particular the attitudes, abilities and behaviour of drivers and that are particularly relevant for the likelihood of losses occurring – but that could not yet be captured, or only marginally. In light of this, the objective risk characteristics in many cases serve as imperfect substitute characteristics for the subjective risk characteristics.²⁰ To be more specific: the type class of the vehicle, the residence of the owner and how long the driver has held a licence are not responsible for whether the driver causes an accident while driving. However, they are statistically significant, plausible and easily recordable indicators of the attitudes, abilities and behaviour of the policyholder on the road. On average, these correlations may be correct; but in individual cases, there are frequently likely to be deviations from these plausible assumptions and statistical values.²¹

The digital revolution

Fundamentals of digitalisation

The speed and momentum with which technology, and above all IT, is advancing is both impressive and challenging. The development of new sources of data, technologies and potential processes is often faster than the willingness and ability of companies and entire industries to apply them in a legally still partly uncertain environment. The enthusiasm, interest and sometimes mere normality with which the new technologies are being used by the general public and by consumers encourage further developments all the more. The sense of entitlement to and expectations of digital applications can even be observed in public administrations and schools.²²

In the first instance, the term “digitalisation” itself refers only to the transformation of numbers, letters, texts, images, videos and other types of data into a digital format, as well as their storage and processing using different computer technologies.²³ The German insurance industry also needs to rapidly and successfully develop corresponding digital skills if it wishes to remain competitive compared with other industries and internationally. The factors of Big Data and Artificial Intelligence are presented in the following as important drivers for insurance undertakings, and their potential impact on the community of policyholders is examined in detail.

Big Data

Big Data refers to the availability of large volumes of digital data and the technical means for exploiting them. The trend towards Big Data is being driven and accelerated by the expanding storage capacities and growing processing speed of new computer technologies.²⁴

Four dimensions of criteria can be used to characterise big data: volume, velocity, variety and value.²⁵

“Volume” describes Big Data using the available volume of data and the byte size as the unit of measure. The volume of digital data generated annually will grow sharply in the next few years. Whereas 16.1 zetabytes²⁶ of digital data were generated worldwide in 2016, this figure is expected to increase tenfold by 2025.²⁷ The “velocity” criterion stands for the speed of data generation and processing. Data diversity is the subject of the “variety” criterion. This does not merely mean different file formats (e.g. images, emails, Word and PDF files, videos), but also the degree to which they are structured. Data is unstructured if it does not correspond to a formalised system. This is normally the case with images and emails. Semi-structured data either does not have any fixed type of structure, but only a hidden structure, or it is structured differently overall. In addition to the file formats, which must be compatible with the company database, structure also refers to the structures of the field types (for example Title_Source_Date) that describe a file in a database in greater detail.²⁸

Investments in information technologies for storing and processing Big Data should pay off, of course, and increase enterprise value. If not, collecting, storing and evaluating Big Data will not be a worthwhile business exercise. This objective is included in the “value” criterion.²⁹

When externally sourced data is used, its quality and informative value must be validated. Another factor that has to be considered is that where personal data is involved, the data subject has a right of access to any information about the source of the data pursuant to Article 15(1)(g) of the European General Data Protection Regulation (GDPR).³⁰

The transition from conventional databases and the way the data they contain is evaluated to Big Data is fluid. Traditionally, insurance undertakings have large data inventories with long histories, resulting for example from data collected from applications for coverage and claims experience. In many cases, however, this data is stored in distributed database systems and it has not been/will not be merged. Data quality – in the sense of continuous updating and harmonisation – also mostly needs improving. Data about subjective risk characteristics can now also be made available using smart gadgets. And there are growing opportunities for obtaining data from external sources. These include special data providers, websites and the social web.³¹ However, there is a large gap in particular to the US big techs (Google, Amazon, Facebook, Apple: collectively Gafa, etc.) and other industries because very little open data and real-time data is captured. “Open data” is “unfiltered, machine-readable electronic data that is available to everyone publicly, for no particular purpose and without obligation”.³² It usually comes from the federal administration, with examples including information about traffic, tourism or the weather.³³ In the same way as all other companies in all industries, insurance undertakings are also free to additionally integrate open data to optimise their products, services and business processes, for example as a basis for rate setting.

In addition, real-time data can help to decipher the activities and behaviour of policyholders and insureds in order to draw conclusions about risk situations, among other things. Real-time data is already being collected today about users in large quantities, for example from smartphones and wearables, and evaluated by the manufacturers. Automobile manufacturers are able to measure the driving behaviour of drivers in real time. The question will be: who will be legitimised in the future by the sovereign owners of this data (they should always be the citizens who are themselves being measured) to use it for what purposes? Using personalised real-time data, insurers in turn would also be able, for example, to come much closer to the subjective risk characteristics.

However, the insurance industry is also making progress in the Big Data playing field. In motor vehicle insurance, for instance, pay-as-you-drive and pay-how-you-drive premium rates are increasingly being developed and offered. With pay-as-you-drive rates, the insurance premium is calculated precisely per kilometre actually driven, whereas with pay-how-you-drive rates, the driver’s personal driving behaviour is analysed and forms the basis of the individual premium calculation.³⁴

Pay-as-you-live rates represent further lines of development in other classes of insurance. In this case, data on personal everyday behaviour is read using wearables such as smart watches or transmitted to the insurance undertaking through smartphones or smart home technologies³⁵. However, insurers are still quite cautious about these concepts as a whole.³⁶

Managing Big Data is tied to technical, human resources and intellectual requirements. The trick is to generate Smart Data that meets other quality criteria, such as data protection requirements and social acceptance.
To be able to use Big Data efficiently and rationally in risk analysis and insurance rate setting, developments in the field of Artificial Intelligence offer new approaches.

Artificial Intelligence

It is very difficult to find a generally accepted definition of “intelligence”. Characteristics of human intelligence, for example, are understood to include practical problem-solving skills, verbal skills and social skills.³⁷

“Artificial Intelligence” refers to the problem-solving ability of computer technologies that otherwise only humans possess because of their intellectual processing abilities.³⁸ A distinction can be made by intelligence levels between weak Artificial Intelligence and strong Artificial Intelligence although the boundaries cannot be said to be well-defined.³⁹

Using weak Artificial Intelligence, computer-driven machines are able to develop solutions that are limited to certain tasks they have been taught to carry out.

Examples include navigation systems and the correction functions in electronic writing media. Strong Artificial Intelligence is the term used when machines emulate human intelligence and achieve broader cognitive performance.⁴⁰ This is the case when machines can draw logical conclusions, continually learn new things or make clever decisions when faced by uncertainty.

To develop Artificial Intelligence, neural networks, i.e. networks of nerve cells, in the human brain have been and continue to be researched, modelled using state-of-the-art technologies and thus imitated piece by piece. In the case of strong Artificial Intelligence, computer-operated machines are taught how to learn (machine learning⁴¹) and hence to make judgements and solve problems.⁴² Deep learning is an aspect of machine learning in which machines develop prediction abilities and the skills to make their own decisions.⁴³

How Big Data and Artificial Intelligence affect calculations and rate setting

For the first time, Artificial Intelligence allows Big Data to be used rationally and efficiently. Innovative processes that draw on the power of Artificial Intelligence, such as developments in the field of smart data analytics, can harness large, unstructured masses of data by filtering, selecting and sorting them, and transforming them into a single format. Using Artificial Intelligence instead of human intelligence saves expensive human resources capacities and allows them to be deployed elsewhere. In addition, there are no symptoms of fatigue when Artificial Intelligence is used to process large volumes of data, making processing faster and more error-free.

Using pattern analyses, among other things the relationships between real-time data on the one hand and behavioural characteristics and loss potential on the other can be explored reliably and efficiently. This also significantly advances the analysis and evaluation of subjective risk characteristics.

These insights and “black box” processing of Big Data enable real-time rate setting, among other things. Incoming behaviour-based data from the sensors of a motor vehicle, from a smart watch or other fitness trackers, as well as from a smart home, can be rapidly analysed at the insurer using intelligent, automated processes and processed in a personalised insurance rate with an individual premium. Bonus-malus rules could also be automated and implemented rapidly in the case of certain behaviour patterns defined in advance. Ultimately, the substantial gain in processing capacity, speed and flexibility can lead to almost entirely individualised insurance premiums. Moreover, through the meticulous analysis of Big Data, artificially intelligent processes can reduce existing information asymmetries that disadvantage the insurer.

As things stand today, however, some of the prerequisites needed for using Artificial Intelligence in insurance undertakings still have to be established. This starts with explaining to policyholders the benefits of transferring data. Data that is already distributed across many older systems in the insurance undertaking must be centralised in compliance with data protection regulations and made available for analysis. It should also be noted that, under Article 13(2)(f) of the GDPR, the insurance undertaking must inform policyholders about the logic involved in automated decision-making. This will in itself be problematic because in most cases, strong Artificial Intelligence is no longer based on an easily understandable algorithm that could be disclosed and communicated.

What this means is that Artificial Intelligence will not be able to completely replace human intelligence in rate setting under the current rules. It is certainly the case that the requirements of statistical significance and (together with the digital possibilities for data mining) the problems associated with recording behaviour patterns and subjective risk characteristics can be overcome using Artificial Intelligence, as described above. However, Artificial Intelligence makes the problem of verifying plausibility all the more critical. In BaFin’s view, to the extent expressed, insurers must at all times also be in a position to explain the logic of the algorithms in accordance with Article 13(2)(f) of the GDPR and, where appropriate, to verify adherence to certain ethical and legal principles (for example, compliance with the prohibition on premium rates based on gender and nationality).⁴⁴ A sceptical approach still appears justified as to whether these requirements are in any way compatible with the use of Artificial Intelligence. In light of this, if the use of Artificial Intelligence is restricted by (supervisory) law, the question arises of whether German insurance undertakings (and the German economy in general) will suffer critical disadvantages and be left behind in the international and cross-sectoral competitive environment. It is already the case today that the development centres for Artificial Intelligence are situated in the USA, China and India, rather than Germany.

Impact on the community of policyholders

The objective of using Big Data and Artificial Intelligence in rate setting by insurance undertakings is therefore to calculate an insurance premium that is closer to the individual expected claims associated with the individual policyholder or risk. A premium that approaches the individual actuarial principle of equivalence can also be perceived as fairer or more performance-based.

An individually actuarially equivalent premium will, of course, prove to be advantageous for policyholders with below-average risks and disadvantageous for policyholders with above-average risks – for the latter in particular if their adverse risk characteristics cannot be positively influenced by some other behaviour (e.g. genetic risk factors in personal insurance). In the case of individual and behaviour-based insurance premiums, a question that will be increasingly asked is which factors, characteristics and behaviour patterns may or may not be included in rate setting.

In addition, whether the relationships between concrete behaviour patterns and their impact on individually expected losses have already been sufficiently researched still remains to be seen. For example, there is a need for a more detailed examination of how exactly speed, acceleration and braking behaviour in a motor vehicle actually affect the potential risks. In the case of pay-as-you-live premium rates, for example, it is necessary to investigate which foods are really health-enhancing or harmful for which groups of people, or which intensity and duration of which sports activities promote or damage physical and mental well-being in the long term – and all this with regard to both statistical significance and plausibility. Misleading conclusions, or at least conclusions that are not accurate to the greatest extent possible, can lead to unwarranted discrimination against certain groups of policyholders.

However, to the extent that rate setting on the basis of Big Data and Artificial Intelligence leads to individual premiums, the balancing of risks in the community of policyholders will not generally be impaired. In light of the fact that the community actuarial principle of equivalence is also necessarily satisfied when a premium is calculated in accordance with the individual actuarial principle of equivalence, the collective insurance premium of the community will be sufficient to cover the overall expected losses even if rates are set individually. This unequivocally counters the widespread misunderstanding that the community concept and hence the principle of insurance are no longer satisfied by an individualised premium. The opposite – see above – holds true: “Insurance provides cover for funds that are needed and whose specific amount is uncertain but are estimated in the aggregate on the basis of the balancing of risks in the community of policyholders and over time”. Even – and especially – in the case of an individual insurance premium that has been actuarially calculated correctly, risks in the community of policyholders are balanced and anti-selection is avoided. From a risk theory perspective, this also results in a fair premium. Whether this premium will also be viewed as fair by the general public is another question that will not be discussed here.⁴⁵

Reduced operating and administrative expenses can be a significant advantage of using Artificial Intelligence and the associated black box processing in rate setting for insurance contracts. As with all innovations, however, a certain payback period must first be taken into account here. In turn, the entire community of the policyholders benefits from falling costs, as the cost advantages should be reflected in lower premiums.

With regard to compliance with data protection requirements, the use of Big Data and Artificial Intelligence currently still faces significant challenges. In particular the GDPR, with its requirements for data minimisation⁴⁶ under Article 5(1)(c) of the GDPR, storage limitation⁴⁷ under Article 5(1)(e) of the GDPR and purpose limitation⁴⁸ under Article 5(1)(b) of the GDPR, constrain the effective opportunities for deploying Big Data and Artificial Intelligence. In the absence of clearly defined rules so far, the real challenge in practice today is to explore the legally permissible scope for using and analysing available data.

Assessment

In principle, using Big Data and deploying Artificial Intelligence can improve risk balancing in the community of policyholders through even sounder individual risk premium calculations. In addition, there are many opportunities for cost optimisation that will benefit the community of policyholders. It is important for the degree of individualisation in rate setting to go hand in hand with the pace at which social acceptance of these technologies and the related business models increases.

The extent to which the growing degree of individualisation will jeopardise the concept of solidarity and thus disturb the sense of justice among those responsible for consumer protection, policymakers and the general public is explicitly not an essential question in the private insurance industry business model. Private insurance does not follow the solidarity principle based on the criteria of “healthy for the sick, rich for the poor, strong for the weak” that lies at the heart of the pillars of the social insurance system, but rather the principle of risk balancing in the community of policyholders.⁴⁹ Nevertheless, private insurance companies also have to meet expectations, and they will face criticism if they do not take solidarity sufficiently into account in their premium rate calculations. A prominent example is the debate about insurance premiums in liability insurance for midwives, which are actuarially correct but unaffordably high.

However, while there is the problem of high premium charges for certain above-average risks in the community of policyholders, there is also the possibility that positive behavioural effects may emerge. This is because, to the extent that undesirable behaviour patterns in policyholders or insureds are the reason for the excessive risk, the associated insurance premium might prompt them to correct or stop that behaviour. For example, if Big Data, Artificial Intelligence and real-time processing are used to show a significant premium supplement to a car driver on their dashboard display when there are indications that the driver is speeding, this could lead to an immediate adjustment to the driving behaviour and, in the long term, to desirable behavioural adaptions as well. Such consequences probably also do not disrupt the concept of solidarity; on the contrary, they strengthen responsibility for personal behaviour in the community of policyholders and hence also actuarial solidarity. However, the issue of demarcation here is difficult and lies in the grey area between the behavioural and fate-driven characteristics of policyholders and their underwriting risks.

Summary

Overall, integrating Big Data and Artificial Intelligence can bring benefits – for both insurers and policyholders. In the race for potential applications and deployments, however, insurance undertakings are facing competition from technology giants such as Amazon, Apple, Facebook and Google, among others, which are not only far more skilled at collecting and processing data, but also enjoy a higher level of social acceptance when they do this. Moreover, insurance undertakings’ databases (still mostly distributed), their data quality (in need of improvement) and their IT systems (frequently outdated) are proving to be drawbacks to implementing promising, high-potential technologies. Insurance undertakings will be forced, and should do their utmost, to welcome innovation with open arms and exploit it to their advantage. It will not be easy to keep pace with the legal environment, especially in the field of data protection, as well as the economic opportunities (investment costs!) and, last but not least, social acceptance.

Another challenge will be to create a more profound level of trust at their customers and to prove their sovereignty in handling the data. There will continue to be a growing focus on generating genuine added value that is tangible for customers so as to persuade them that it is worthwhile revealing data on subjective characteristics.

In line with its title, this article has focused on and largely confined itself to the impact of Big Data and Artificial Intelligence on the community of policyholders. It has ignored other promising applications in the insurance industry, such as detecting fraud, claims management in general, end-to-end process optimisation as well as generating customer-centric value-added services above and beyond pure insurance cover. These aspects open up additional and probably even considerably greater and decisive potential future opportunities – and what’s more: they are likely to be essential for the survival of individual insurance companies.

Footnotes: