Broken loops,open futures: The building and breaking of behavioural insurance

Abstract

Digital looping effects between algorithmic technologies and users are promoted as reshaping various industries by optimizing operations, improving predictions and creating new market opportunities. Insurers are exploring these promises by collecting customer-generated data and testing its use in risk calculations and behavioural interventions. However, these novel insurance technologies have been criticized for enabling totalizing forms of surveillance, control and discrimination, potentially leading to the foreclosure of future actions. This study tests the argument that emerging insurance technologies ‘narrow the future’ by analysing Finnish life insurers’ efforts to build a digital feedback loop into their behavioural policies. It centralizes breakages in these new data practices as the locus of analysis, showing how the feedback loop dissolves at various points or is never established due to shortcomings in the new technologies, regulatory barriers and aspects inherent to insurance logic itself, thereby undermining the policies’ aims; yet, the insurers’ experimentations cannot be simply framed as failures because they produce knowledge that can inform future practices. The present study illustrates the importance of examining how versions of dominant technological visions are reworked in local, field-specific practices. It shows that instead of producing ‘guaranteed outcomes’, these novel insurance operations raise new questions and uncertainties that can have harmful effects beyond the totalizing scenarios. Focusing on breakages challenges techno-deterministic perspectives, keeping the future open while enabling a more precise critique of the harms associated with these visions and their often-imperfect implementation.

Introduction

‘This virtuous data-driven relationship loop – a flywheel – is the foundation of the BigTech companies’ success.… It enables them to respond agilely to new insights and continually improve their products and services, which generates more consumer trust and higher organic growth.… Insurers have not yet built the digital infrastructure necessary for this type of business model.’ (Raverkar et al., 2020).

The technological visions of the past decade suggested that digital data and algorithmic technologies would reshape all kinds of industries using similar principles and logics: by creating automated systems that gather constant streams of detailed information, analyse them and translate them into actionable insights (Ruckenstein and Pantzar, 2017; Van Dijck, 2014). This mechanism, often described as a digital feedback loop, is integral to technologies such as recommendation algorithms, targeted advertising and health tracking devices. Combined sometimes (but not necessarily) with addictive designs to ensure user engagement (Schüll, 2016) and a digital choice architecture of ‘hypernudges’ to steer user behaviour in a certain direction (Yeung, 2017), it strives to affect users’ actions and adjust itself based on those actions. ¹ Thus, the digital feedback loop creates an interactive and mutually influential relationship between an algorithmic technology and a user, where actions and interventions on one side impact the other (Fourcade and Johns, 2020; Mathieu and Pruulmann Vengerfeldt, 2020). These looping effects can, among other things, enable detailed profiling of users and the interactive optimization of services, improve prediction accuracy and create new market opportunities (Burrell and Fourcade, 2021).

Insurance is one field experimenting with the promises of digital looping effects. New data types and algorithmic techniques are compelling for insurers because they hold the promise of more precise ways of calculating and pricing risk and tools for proactive risk and customer relationship management. Inspired by the benchmark practices of Big Tech, insurers across sectors, including motor vehicle (Cevolini and Esposito, 2022; Francois and Voldoire, 2022; Meyers and Van Hoyweghen, 2020), health (Jeanningros and McFall, 2020; Presset and Ohl, 2023) and life policies (Tanninen et al., 2022a), have been experimenting with the use of customer-generated data collected via accelerometers and self-tracking technologies.

Alongside the promises of innovation, behavioural insurance policies have drawn the attention of critical scholars who have raised concerns over how the new policies could lead to an all-encompassing system of digital surveillance and discipline, capturing people's lives in a responsive data-driven infrastructure (Sadowski, 2023). Such a system could allow anticipatory control of policyholders’ actions (Sadowski, 2023) and ‘real-time adjustments’ to their individualized risk scores (Zuboff, 2019: 214), developments that shift responsibility to individuals, threatening the solidaristic structures of insurance and constraining possibilities for action (Cevolini and Esposito, 2020, 2022). Traditionally, insurance has been a technique for liberating action by enabling policyholders to engage in riskier activities because the financial burden is shared with the insurance collective (Ewald, 1991: 201–205). However, with new techniques aimed at enhancing risk assessment precision and enabling intensified surveillance of policyholders, insurance might have precisely the opposite effect – foreclosing future opportunities through its totalizing control (Cevolini and Esposito, 2022: 133; Sadowski, 2023). A similar phenomenon dubbed by Nowotny (2021) the ‘paradox of prediction’, depicts how the performative power of predictive algorithms to make what they forecast happen may inadvertently restrict the openness of the future and contribute to a move towards a more ‘deterministic world’.

In the present study, I test the notion that emerging insurance technologies ‘narrow the future’ by analysing Finnish insurers’ practices of developing behavioural life insurance policies. Specifically, I discuss how breakages are at stake in insurance professionals’ efforts to build a functioning feedback loop, combined with nudging elements, such as gamified wellness apps, into the policies. Although insurance is based on statistical and calculative techniques, there is empirical evidence that implementing new data-driven insights in insurance is difficult due to technological, regulatory, market and cultural barriers (McFall, 2019). While some industries already have algorithmic technologies at their core, insurers are only beginning to explore them. As this article's epigraph from a Swiss Re publication states, the insurance industry lacks the infrastructure necessary to foster digital looping effects and to harvest their promised benefits. With longstanding traditions, strict regulation and legacy systems, insurance is a conservative business (Francois and Voldoire, 2022), making the implementation of new technologies and ‘livelier’ digital data difficult (Lupton, 2015).

To study what building new capabilities in insurance looks like when shifting the focus away from the dominant visions of the digital transformation and the potential effects of behavioural insurance, I draw on Stephen Jackson's ‘broken world thinking’ which forefronts ‘breakdown, dissolution and change rather than innovation, development or design as conventionally practised’ as the focus of analysis (2014: 222). Following Jackson, I see that ruptures and the responses to them, such as repair and repurposing, are vital for innovation because in the moments of breakdown, the relations, values and orders embedded in sociotechnical systems become visible and open to negotiation. Furthermore, my study aligns with previous research underscoring the role that pilot policies of interactive car insurance have in scoping and forging future markets, illustrating how these policies have broken down – or how they have been designed to fail from the outset, as their primary objective is to produce information (Francois and Voldoire, 2022; Meyers and Van Hoyweghen, 2020). I develop these perspectives by examining a case of life insurance – a product entailing issues specific to its case as it has a long time window and is both more abstract and affective than car insurance (Lehtonen, 2014; McFall, 2014).

By focusing on breakdown and rupture, I illustrate how versions of the dominant technological vision are reworked within the local, field-specific practices of developing behavioural life insurance policies. I show how, despite the insurer's efforts, the digital feedback loop dissolves at every stage or is never fixed in the first place. The shortcomings in the new algorithmic technologies, Finnish regulatory requirements and features of the insurance technology itself all create gaps, cracks and misalignments, some of which fundamentally compromise the looping effects. Yet even though it is obvious that behavioural life insurance policies were not an immediate success, Finnish insurers persist in their efforts five to seven years after the original fieldwork, repurposing parts of the original policies for other uses such as customer wellness benefits or keeping them alive as dormant competences. Hence, the breakdowns and uncertainties they create are not straightforward failures either, as they can open space for ideas and new operations.

The results illustrate that implementing a technological vision in a well-established field is a precarious process. While in some other (regulatory) contexts, digital feedback loops may operate successfully to monitor and steer people's behaviour, ‘totalizing control’ and ‘a deterministic world’ remain distant realities for Finnish life insurers. However, it should be acknowledged that imperfect algorithmic operations can also cause harm and restrict possibilities for action, including in the insurance context. Although they do not follow the totalizing scenario, the operations of Finnish insurers can still be harmful by, for instance, introducing new forms of distrust and anxieties into a relationship whose aim should be to generate security (Tanninen et al., 2022b). Evaluating scenarios of algorithmic harm in a situated manner and considering the contextual, legal and technological issues at stake allow us to identify harms beyond these scenarios and to carry out more precise critiques. In doing so, centralizing breakages in algorithmic practices becomes a powerful tool for interrogating recent techno-deterministic narratives and the expansion of (big) tech into new fields (Stevens et al., 2024) in ways that preserve an open future while also recognizing the (potential) harms of the visions in practice, thereby fostering alternative ways of thinking.

Looping effects in insurance

The idea of a feedback loop is not new to insurance. However, instead of the well-aligned positive feedback loops that data-driven technologies are supposed to engender, insurers have concentrated on negative looping effects such as moral hazard, a concept from economic theory that crucially steers insurance operations. Its meaning has evolved over time ² , but the most common account holds that when people opt into insurance and shift the economic responsibility for their actions to the insurance collective, they are more likely to engage in risky behaviours (Pauly, 1968). From the insurer's perspective, this can lead to a negative spiral of increased indemnities, diminished profits and, in the worst-case scenario, issues with solvency. Another classic concept is adverse selection, which refers to a situation of asymmetric information, where policyholders ostensibly know more about their level of risk than the insurers (Baker, 2003). In this case, insurers cannot properly calculate risk because they lack relevant information, which can lead to high-risk individuals comprising most of the pool and the indemnity rate increasing. Consequently, insurers must raise premiums, which then drives low-risk policyholders to opt out of coverage, hence leaving only higher risks in the pool.

The economic understandings of moral hazard and adverse selection have been criticized for disregarding the fact that insurers typically possess much more information about risk exposure and the overall insurance arrangement than policyholders (Van Hoyweghen, 2007) and can act in a risk-inducing way (Ericson et al., 2000). Nevertheless, insurers see moral hazard and adverse selection as real vicious cycles that inform their practices. They aim at minimizing their effects through, for instance, risk calculations, underwriting and insurance terms and conditions. The experimentation with data-driven practices is partly motivated by the goal of combating these issues. Behavioural data are envisioned as helping mitigate moral hazard and adverse selection by allowing insurers to access detailed information about a policyholder's actions over the course of the insurance contract, not just at the moment of underwriting. As the use of demographic parameters for risk categorization has become more regulated by the European Union in recent years, insurers are motivated to use new attributes over which people supposedly have control (Meyers and Van Hoyweghen, 2017).

Insurers are not just envisioning using data-driven technologies to mitigate negative feedback loops and improve risk calculations; they hope these digital tools will create ‘more timely feedback loops’ (Avramakis et al., 2020). A key promise of behavioural insurance is that it could shift the relationship to risk from reactive to proactive (Cevolini and Esposito, 2022: 126), allowing insures to affect the probability of risk. This would ostensibly be accomplished by establishing a positive feedback loop between insurance technologies and policyholders: the data, combined with different incentivizing models, would be fed back to customers with the aim to nudge (Thaler and Sunstein, 2008) their behaviour towards habits deemed safer or healthier. However, the purpose of these operations is not merely to impact policyholders’ risk behaviour but also to foster positive customer interactions and relationships. The data-driven tools promise a continuous presence in people's lives: this ‘relationship loop’ (Raverkar et al., 2020), the insurers believe, helps transform the typically distant customer relationship into a long-lasting and more intimate bond (Jeanningros and McFall, 2020; Tanninen et al., 2021).

What the insurance industry often presents as a win-win situation for both insurers and policyholders appears less rosy when considering the impact that policies could have on people who fail to comply with the predetermined ideals of prudent conduct embedded in them. Behavioural policies could discriminate against those deemed high-risk and, in the worst case, exclude them from coverage. Insurers’ tendencies to control policyholders, refine risk categorizations and narrow risk pools have long trajectories (Ericson and Doyle, 2004; Ericson et al., 2003), and data-driven tools could intensify and accelerate those efforts (Barry and Charpentier, 2020; Sadowski, 2023). Insurers have a financial incentive to minimize the occurrence of insurable events and want to know the risk that they are insuring precisely; hence, the purpose of these governing efforts is to make the operational environment more controllable and predictable (Sadowski, 2023; Zuboff, 2019), aligning with what scholars have described as the data-enabled narrowing of the future (Cevolini and Esposito, 2022; see also Nowotny, 2021).

Still, even though insurers undeniably aim to control the risks that they cover, insurance requires a collective structure and some degree of uncertainty to function. Risk is always calculated in a pool, where the rate of its occurrence is known, but the specific individuals or entities affected by that risk remain unknown (Ewald, 2021). Depending on the particular insurance arrangement and political decisions regarding how premiums should be determined, pools can be narrower, with each member paying for their own risk, or more subsidizing, as in social insurance (Lehtonen and Liukko, 2011); nevertheless, some form of collective is needed to spread risk. Critiques of the totalizing control of the new insurance policies often overlook this need for a baseline collective structure, employing a rather narrow ‘insurantial imaginary’ (Ewald, 1991) that perceives insurance as driven solely by an individualizing logic. However, if in an extreme scenario, digital data provided insurers with an omniscient perspective, they would not be able to offer insurance as they would already know who is going to seek compensation; consequently, their business model would shift towards risk prevention (Cevolini and Esposito, 2022) or some other activity unlike insurance as we know it. This scenario is, of course, quite unrealistic, and the integration of behavioural data within insurance remains ‘more aspirational than actual’ (Sadowski et al., 2024: 233) – and, in its extreme form, it might not even be desirable because it could leave insurers without a business (Metcalf and Sadowski, 2024: 6–7). It is important to regard this tension between insurers’ goals of governing the future and their need for some degree of uncertainty because it informs the experimentation around behavioural products.

Broken loops

This study examines a side of technology familiar to most developers and users: breakages, frictions and mending. The concept of ‘breakage’ can be understood in different ways in the context of innovation and technology development. First, the breakdown of a technology can be viewed as a failure. Recent studies show that despite heavy investments, most digital transformation efforts in both the private (Oludapo et al., 2024) and public sectors (Kempeneer and Heylen, 2023) have failed to deliver on their promises. Hence, the outcomes of such projects are often disappointing, highlighting the discrepancy between grand technological visions and often lacklustre results.

However, if we view failure subjectively rather than objectively by treating it as it is perceived by relevant actors (Mica et al., 2023), labelling all breakdowns in technological development as straightforward failures becomes complicated. For instance, studies on interactive car insurance argue that even though the policies failed to engage customers and perfect technologies, leading to their quickly exiting the market, they were successful experiments (cf. Gross, 2023) because they generated new knowledge for insurers (Francois and Voldoire, 2022; Meyers and Van Hoyweghen, 2020). Furthermore, the idea of a breakage is connected to the Silicon Valley innovation culture of ‘moving fast and breaking things’, where disruptiveness is cast as an essential part of the innovation process, echoing the Schumpeterian idea of creative destruction; at the same time, this stance of purposeful rupturing while ignoring practices of care and repair can lead to ethical issues (Grieser et al., 2023).

The breakdown of the technology itself can be seen as a site for innovation as, following Jackson (2014), technological development works through practices of repair and restoration. In other words, moments of breakdown are when technologies are adjusted, repurposed and become open to negotiation. Because of this opening, they can also act as entry points for researchers. Algorithmic technologies are socio-technical systems that are intimately intertwined with human practices, interactions and qualities and that participate in and reproduce society (Airoldi, 2022). Yet, the values integrated within these systems can be difficult to see when they are intact and functioning as intended. When there are ruptures, the relations, values and orders embedded in technologies become more visible (Bowker and Star, 1999; Jackson, 2014).

I draw from these different senses of breakdown and failure when analysing the practices of developing a digital feedback loop in behavioural life insurance. By focusing on the difficulties that insurers encounter as the loop falters at every step, I gain insights into how visions and values are woven within the new data-driven technologies. Moreover, the emphasis on breakages opens avenues for diverging from techno-deterministic narratives, fostering alternative perspectives.

Empirical materials

I analyse two Finnish behavioural life insurance policies introduced in the latter half of the 2010s by firms I refer to as Company X and Company Z. Apart from a brief experiment by a third insurer, no other company has developed similar policies in the Finnish life insurance market; therefore, my focus is on these two policies. At the time of the fieldwork, they both operated on similar principles. Life insurance in Finland is a voluntary policy supplementing state benefits and social insurance. The behavioural policies in particular were presented as ‘additional services’ or ‘experimental products’ that were still under development but available for customers. They invited policyholders to engage in self-tracking practices and share their activity data in exchange for access to e-health services and/or financial incentives. Both companies had a ‘bring-your-own-device’ approach ³ , allowing tracking via different technologies, including smartphones, activity wristbands and smartwatches.

The self-tracking practices were linked to core applications developed by partnering Finnish data analytics companies specializing in workplace wellness programmes and insurance services. These applications allowed customers to monitor their activity levels, interact with their policies and access various online wellness and coaching tools produced by different start-ups, including services for stress management, smoking cessation and healthy eating. These tools employed both educational features, such as short texts about healthy living, and features designed to nudge people through push notifications, haptic cues and gamified structures.

Importantly, the core applications facilitated the collection and handling of customer-generated activity data, managed by the data analytics companies. By outsourcing these services, the insurers quickly accessed expertise that was not readily available to them in-house. Furthermore, the collaborations ensured compliance with Finnish insurance regulations ⁴ , which restrict data use to demonstrable and legitimate purposes. To avoid legal issues, the data analytics companies provided insurers only with aggregated activity data or specific filtered metrics; the insurers did not have access to the ‘raw data’. For example, Company Z's data analytics provider assessed whether a customer earned enough ‘activity points’ for a financial reward – a bonus on insurance coverage ⁵ – with a 15% increase for ‘moderate activity’ and 25% for being ‘very active’. However, neither company integrated these new data types into actuarial risk calculations that form the core of the insurance policy, nor did they use them to adjust premiums. Instead, Company Z's three-tier reward system was crafted as a bonus structure on top of the traditional risk categorizations. Thus, while Finnish life insurers have considered integrating these data types into risk calculations, they are still far from implementing ‘real-time rate adjustments’ (Zuboff, 2019) or other advanced uses of behavioural data.

This article examines insurers’ attempts to integrate data-driven capabilities into their policies and provide actionable insights to policyholders. The analysis is based on collaborative fieldwork that I conducted at the companies between 2017 and 2019. During this period, I interviewed insurance professionals, observed their meetings and organized focus groups with policyholders. I provided the insurers with practical insights from these focus groups to aid in product development. The collaboration required me to balance my roles as a collaborator and an independent scholar, but it was essential to gain access to the often impenetrable field of private insurance. Due to the companies’ experimental approaches to their new products, they were receptive to various sources of knowledge, including the critical perspectives I presented.

The main empirical materials for this article consist of 16 interviews with insurance professionals involved in developing behavioural policies. These 45- to 75-min interviews were recorded and transcribed. The policies were developed in project groups involving employees from various departments. Thus, with the help of key collaborators, I recruited interviewees with diverse educational and professional backgrounds, including service designers, marketing managers, sales managers, a lawyer, an actuary and some of the companies’ upper management. This diversity provided a comprehensive picture of the developing work, ranging from the high-flying visions of the project leaders to the more realistic perspectives of the actuary and the legal expert. To maintain confidentiality, I have pseudonymized all identities and omitted specific company details, focusing on the broader implementation of data-driven services in life insurance rather than comparing practices between the two companies. In 2023, I attempted a follow-up study to assess the current status of the policies, but that effort was unsuccessful due to key informants changing roles and the policies no longer being a main focus for the companies. However, in the discussion, I do reflect on the policies’ current state using publicly available marketing materials and policy terms and conditions.

I conducted the analysis in several steps, starting with identifying general themes in the interviews, such as ‘data relations’ and ‘developing work’, through a careful reading of the transcriptions. I used ATLAS.ti for thematic coding and its automated features to find extracts related to terms such as ‘data’, ‘tracking’ and ‘developing’. After reviewing and discarding irrelevant extracts, I re-examined the transcriptions to ensure that no important data had been missed. I focused on practical discussions of developing work, behaviour measuring and data integration in risk calculations, aiming to move beyond hype speech to more tangible developing practices. In the second phase, I conducted detailed manual coding, applying up to three codes per extract. I developed the final analysis through rounds of writing, revisiting both the original material and the theoretical concepts. The analysis is structured around three feedback loop stages: behaviour tracking and data collection, experimentation with data analysis and insurance integration and looping behavioural insights back to customers.

Measuring wellness

In the Finnish insurers’ visions and developing practices, the mechanisms of a digital feedback loop and hypernudging were intimately connected. The insurers did not use the term ‘feedback loop’ themselves because its Finnish equivalent is not widely used. However, they described aims that were very akin to a feedback loop; they wanted to be able to gather data about policyholders’ lifestyles, use those data for analysis and provide policyholders with relevant and actionable input or feedback based on their data – information that could help them reflect on their lifestyles and change their behaviour. Consequently, insurers wanted to create a ‘seamless’ and even pleasurable data-driven experience that would accurately recognize and respond to customers’ needs. These aims were combined with normative ideas about how to guide policyholders’ actions to create a win-win situation for both customers and insurers. Aligning with insights from behavioural economics, policyholders were viewed as needing support to lead healthier, lower-risk lives, which financial incentives and behavioural nudges could provide.

Because of this dual focus on both the interactive customer relationship and lifestyle modification, instead of straightforwardly focusing on measuring certain parameters and improving the calculations of policyholders’ health and/or mortality risks, both insurers framed their operations as assessing and promoting ‘holistic wellness’. Consequently, the insurers intended policies to consider all aspects of a ‘good life’, including activity, sleep, nutrition, stress management and financial wellness. The focus on policyholders’ wellness rather than health or mortality was informed by the need to manoeuvre in the Finnish legislative landscape. To make health claims, insurers are supposed to have robust proof that behavioural data reflect people's health status and that health interventions have measurable outcomes. Furthermore, the wellness rhetoric is a way to connect with policyholders. Wellness is a high focus in current market societies with an influx of data-driven technologies aimed at improving it (Schüll, 2016; Specker Sullivan and Reiner, 2019). Insurance companies try to tap into this market and appeal to people's curiosity and concern about their lifestyles, thus attaching the product to consumers and providing reasons for them to engage in data tracking.

Although behavioural policies’ functioning logic aligns, at least to some extent, with the traditional practices of the industry (Sadowski, 2023), the Finnish insurers confronted difficult questions regarding the connections between behaviour, wellness and data. How can behaviour be measured reliably? What can be done with the data? How are behavioural data related to the more abstract notion of wellness? In the interviews, the insurance professionals themselves raised these unresolved issues. For instance, an actuary expressed doubts about the new data practices:

I was constantly uncertain whether this was something that we can do; it felt almost too simplistic and had this atmosphere of uncertainty. We have some recommendations for exercise, an activity wristband that might or might not measure the right thing, and a mediating application that turns the tracker data into points which are then transferred to us. It feels like we are dealing with experimental insurance business. But, then again, I thought, ‘Ok, we are trying this out, let's see what results we can get.’

The actuary discussed the numerous unsettled elements in the policies and the lack of clarity regarding their coordination. What is at stake is a chain of translations. First, the self-tracking devices translate customers’ behaviour into activity data. Second, opaque algorithms produced by the partnering data analytics vendor render the data into activity points that ostensibly indicate whether the user has exercised enough. In this process, general recommendations for exercise act as a device to turn those behavioural data and activity points into a measurement of wellness. These transformations require a great deal of faith from insurers because it is not clear how the end result compares to what they are trying to measure. Actuaries, as the primary professionals at insurance companies with access to the new data types and a deep understanding of determining and quantifying risk, seemed to play a key role in bringing a sense of realism to the developing work. At the time of the fieldwork, they were just beginning to review and figure out what could be done with the aggregated and filtered data provided by the analytics firms. Yet, despite the uncertainties the actuary raised, he remained open-minded about the operations. Following behavioural economics models and principles of service design, even with an unknown result, the experimentation itself was seen as a worthwhile effort in making future markets (cf. Meyers and Van Hoyweghen, 2020).

A related trouble the insurers faced was the contradiction between the focus on holistic wellness and their ability to measure and use only activity data. The firms were seeking suitable digital services to target and measure the different building blocks of overall wellness. However, it proved to be complicated to find satisfying solutions and to combine the different data with the insurance policy, as was noted by a sales manager:

I’m not sure if [sleep data] are something that we can connect to the insurance policy the same way that [the service that handles activity data] is connected. Even though you can, of course, cheat trackers.… I think that, in principle, the level of activity is a no-brainer. But when it comes to measuring the length of sleep and sleep patterns.… I think the connection is fuzzier; how can we prove that when you sleep eight hours a night during a certain period, your life expectancy will rise this much? I think this is much more difficult to understand than [recommendations for] activity; a normal person can comprehend that ‘Ok, I need to exercise this much per week’.

From the sales manager's perspective, the connection between risk and customers’ sleeping habits was more difficult to establish than the ostensibly straightforward relation between physical activity and risk. In this sense, insurers saw sleeping data as less reliable than activity data. Insurance professionals thought that the objectivity of automatically generated data is superior to forms of data that rely on customers’ manual recordings, even if tracking devices could be manipulated. For insurers, objectivity is important in terms of insurance fraud and moral hazard; they believed that customers would exaggerate their scores if the services and rewards were based on their manually given accounts. Furthermore, in the case of sleep, the uncertainty expanded beyond the issue of manual recording. Sleep was regarded as inherently less controllable than physical activity; hence, using sleep data in risk scoring and providing coaching for better sleep was regarded as difficult.

Even though manually produced data were not deeply integrated with the insurance policies, insurers still wanted to use those data for personalizing health services and for encouraging people to take all aspects of wellness into account. Hence, they were piloting a variety of services produced by different start-ups. The interviewees had, however, noted that it was difficult to motivate people to engage in manual forms of self-tracking. Reflecting on a service used to track eating habits, a service developer estimated that ‘nobody has the energy for doing it for months, perhaps not even for weeks’; hence, the lack of automatization was perceived as a hindrance to taking customers’ holistic wellness into account.

Consequently, despite their explicit claim of promoting holistic wellness, the insurers were able to collect and use only activity data. However, they also had to deal with issues measuring this supposedly ‘no-brainer’ factor. First, the interviewees acknowledged that self-tracking devices had trouble measuring different kinds of activities and sports, thus producing skewed data on policyholders’ activity levels. Depending on the monitoring device, activities such as swimming, yoga and cycling would not even appear up in the records. Second, insurance professionals wondered whether even a policyholder who reached a high activity level would necessarily be a low risk. This was discussed by a marketing manager: ‘Is it really healthy to run marathons? Even though you are getting a high [activity] score, is it good for you or for the insurance company?’ Once again, the seemingly straightforward relation between people's behaviour, wellness and data breaks down: how do activity data represent the policyholders’ wellness? This issue remained unresolved.

Experimenting with behavioural risks

The difficulties with establishing a clear connection between behavioural data and people's wellness continued in the next phase of the feedback loop: analysing data and using it for calculating risk. As discussed above, the extent to which behavioural data accurately reflected a policyholder's wellness and risk level remained uncertain; therefore, insurers were uncertain about what they could claim based on the data. During data analysis, these issues were exacerbated by the diversity of devices generating data, each with its own distinctive operating logic; as different trackers operate on distinct algorithms, the data they generate are not easily comparable. This does not necessarily matter when aggregated data are used for evaluating risk at a general level. However, issues arise when data are intended for service personalization, as noted by an actuary:

If two people are walking or running one kilometre in exactly the same manner, but they have different trackers, the outcomes can be different. We then need to ask if we are putting policyholders in an unequal position, especially if we are somehow including activity [data] in the insurance policy.

Again, the vague connection between behaviour and data is at stake. Here, however, this discrepancy not only exists between individuals and their data but also among different policyholders, as policies may favour those with access to more advanced self-tracking devices. Consequently, the actual correspondence between people's health-related habits, their behavioural data and their risk status might be poor. This could both impede attempts to create more individualized risk categorizations and raise concerns about the fairness of the operations.

Beyond the issues related to data reliability and usability, a perhaps more fundamental problem was the lack of sufficient data; there was simply not enough information to draw conclusions. This was due to two reasons. First, policyholders did not regularly engage in self-tracking, so insurers struggled with sparse data. An actuary was ‘disappointed with the data that had been generated thus far’ as the dataset showed that some customers had stopped tracking altogether and that some had not started in the first place. The lack of data could hinder insurers from achieving the policies’ aims as data analysis does not deal well with it. Even if the data were not ‘broken’ (Pink et al., 2018) in the sense of, for instance, corrupted information files, sparse and inconsistent data break the functioning of the digital feedback loop. Consequently, the insurance professionals were not sure what could be achieved with this relatively small and irregular pool of data.

Second, data were lacking because behavioural life insurance policies had been on the market for only a short time. Typically, life insurance contracts are long, and the occurrence of insured events over a short period like a few years is relatively sparse. Therefore, the experimentation must be a long-term trial; as a project leader said, data can be used in actuarial calculations ‘only when we can prove the connections’ between activity and risk. Rushing to conclusions and making erroneous risk calculations could become expensive; if pools were constructed and priced incorrectly, issues with adverse selection and solvency might ensue. This uncertainty is not unique to behavioural policies, as an element of precariousness is always present when calculating and pricing risk (Van Hoyweghen, 2014). Still, the new data types are an uncharted territory when compared to the usual data used in insurance.

Finally, insurers were uncertain about what they were allowed to claim based on the data. As a service designer pondered, ‘Can we as an insurance company analyse customers’ behavioural data and state that “you are an inactive person”? Is that considered health counselling, or are we allowed to do that?’ Similarly, an actuary warned that insurance companies should be ‘cautious when tracking and using health data’, as they are not ‘medical professionals’, although he also acknowledged that ‘some correlations could be found’. These statements show that behavioural insurance operates within regulatory and professional boundaries that are still fuzzy. A former project leader explained that ‘we are getting new sources of data which we can use, for instance, in data modelling.… We are looking at groups of customers, not going to the individual level. That was the recommendation of the data protection ombudsman’. Hence, it is not clear whether insurers in the Finnish context are allowed to use the data for defining the wellness and/or risk status of a given person. Because of this uncertainty around regulatory and technical issues, insurers remained conservative in their operations, resulting in data being kept separate from actuarial risk calculations and used only as an experimental addition to traditional life insurance policies.

Manipulating behaviour

The final step of the feedback loop is to transmit data to users in the form of actionable insights and – in the case of behavioural insurance specifically – by creating hypernudges and financial incentives to encourage and push policyholders towards certain behaviours. Whereas using data for calculating risk was still mostly aspirational rather than actual, financial incentives in the form of bonuses to insurance coverage and nudging through gamified wellness services were already happening. The insurance professionals put significant effort into developing and testing ways to measure and influence customers’ wellness-related behaviours. This shows that although a larger feedback loop in the sense of integrating customer-generated behavioural data into risk calculations and using them to determine premiums was still unrealistic, there was trust in the effectiveness of new types of behavioural interventions and the smaller feedback loops between these wellness services and policyholders. The idea was that with behavioural change interventions, the burden of risk would decrease even if the connections were uncertain: the effect might not be enormous on an individual level, but the general health of the pool might improve.

However, the insurers acknowledged that facilitating behavioural change is difficult; as discussed above, they struggled to keep customers ‘in the loop’ and make the wellness services and data tracking features work. Furthermore, the insurers had little evidence of the effectiveness of their behaviour change interventions. A marketing manager reflected on the difficulty of measuring the impact of behavioural nudges:

We conduct a customer satisfaction survey twice a year. It is sent to everyone; we measure NPS [net promoter score] and ask at the same time if customers have changed some aspects of their lifestyle and if these changes have been permanent.… Perhaps later we could get some other measures as well … [lifestyle] is very difficult to measure. Of course, in the future our actuaries might see some results [in the behavioural data].

The questionnaires insurance professionals used to probe the impact of their interventions did not of course give an exact measure of the effects of the behavioural interventions but information about how people experienced the services. Although asking people's opinions contradicts the idea of an automated digital feedback loop providing real-time insights into the intervention impacts, for the Finnish insurance companies it was an important measure to assess the new data practices’ effects on customer relationships: whether or not customers found the services useful and enjoyable. Still, insurers hoped that they would obtain ‘harder’ quantitative evidence on the interventions’ effectiveness in a later stage of experimentation.

In addition to the uncertain behaviour change interventions, insurance professionals resorted to another strategy to shape the risk level of the pool: marketing behavioural policies to already health-conscious groups of people. Insurance professionals from both companies defined ‘younger women who are interested in holistic wellness’ as their most important target segment. As others have argued, the main goal of behavioural policies might be marketing (Jeanningros and McFall, 2020; McFall, 2019). Thus, measuring and manipulating customers’ behaviour is not the only way to enhance predictability; instead, targeting the already healthy segments might be a more effective strategy.

Discussion

This inquiry into Finnish life insurers’ attempts to build a positive digital feedback loop between their behavioural policies and customers reveals numerous, ongoing difficulties in implementing data-driven systems in insurance. The analysis shows how the feedback loop breaks and dissolves at every step: that is, when collecting data, analysing and using that data in risk calculations and looping it back to policyholders in the form of actionable insights and behavioural nudges. Although the insurers think it is likely that a physically active lifestyle does decrease the risk of mortality, they struggle to measure, document and prove this connection within their population. It remains uncertain how customers’ self-tracked data relate to their wellness and risk levels and how those data should be effectively used for interventions. Many of these issues stem from the shortcomings of the new algorithmic technologies. However, equally importantly, Finnish insurance regulation and the features of the insurance logic itself prevent the loop from taking place, as insurers’ concerns about compliance, moral hazard and their own solvency inform and limit their operations. Consequently, the larger feedback loop where customers’ behavioural data would be deeply integrated within the insurance machinery of risk making and management does not materialize. Smaller looping effects do occur when policyholders use the integrated wellness services, but these tend to be short-lived due to limited engagement and technological constraints. Consequently, at this stage, setting up behavioural policies that would function following the technological visions of continuous surveillance and automated rate adjustments is unrealistic, as insurers lack the necessary skills, supporting operational environment and technologies.

This gap between industry visions and real-life practices was not a surprise to the Finnish insurance companies. While the experimentation could be framed as a failure, insurers viewed it as valuable, representing the first step towards new types of insurance operations. Creating a functioning digital feedback loop that enhances predictability might be a long-term objective for insurers, but they also have other goals. In the Finnish case, insurers were exploring how they could implement new technologies to support the marketability of life insurance through the value of holistic wellness and establish more durable connections with consumers.

Despite the difficulties and limitations, both policies still exist in some form five to seven years after the initial fieldwork. When observing the marketing materials and websites of the companies instead of the grand openings they were given at the time of the original fieldwork, the policies have been shunted off to the side. At Company X, the behavioural policy is only discoverable through a search engine or by downloading the health insurance app into which it is now integrated. While many of the original online coaching tools are available in the app, there is no mention of data collection or monitoring. At Company Z, the behavioural policy is more visible and has been integrated into a new app within the general company brand. Although it is no longer the main selling point of life insurance, it is cited as an additional service that helps policyholders gain more insurance coverage through exercise. Hence, the original offering remains as it was – Company Z is still collecting people's activity data and categorizing them into three groups reflecting overall activity levels. At both companies, the partnerships with the original data analytics companies appear to have dissolved and been replaced by other service providers. In future research, it would be valuable to examine the role of these partnering companies in setting up behavioural insurance policies and the markets related to them.

The persistence of behavioural life insurance policies separates them from interactive car insurance policies where, after an experiment, the products exited the market (Francois and Voldoire, 2022; Meyers and Van Hoyweghen, 2020). One possible explanation for this is the long time window of life insurance – the experimentation in this case needs to have an extended duration to show meaningful results in terms of mortality. However, although the practice of self-tracking has become popular during the last decade, creating relevant and robust enough market attachments (McFall and Deville, 2017) between policyholders and insurance-related self-monitoring practices is difficult, especially in the European market space where mandatory tracking is forbidden by law (McFall, 2019). Given the evidence on the episodic (Gorm and Shklovski, 2019) and evolving (Kristensen and Ruckenstein, 2018) nature of typical self-tracking practices, a vision that life insurance policyholders would engage in decades of self-monitoring is unrealistic. Furthermore, tracking these ‘vital signs’ can make the new data relations more visible and thus increase consumers’ worries about the consequences of these practices (Tanninen et al., 2022b). Although it is uncertain whether such operations and technologies are salvageable, insurers do appear to perceive continuing the experiment as valuable. The policies remain in the companies’ reserve, perhaps ready to be mobilized when the time is right – for instance, when there are enough data, the technological capabilities have improved, or there is more regulatory and/or market acceptance of insurance companies engaging in such operations.

Conclusion

The present study illustrates how the vision of a universal data logic often operates in specific contexts and situations through the processes of breakage, repair and repurposing. Instead of managing to create a seamlessly functioning predictive loop that steers action to become more uniform – and thereby narrows future possibilities – the Finnish insurers encountered numerous technical, regulatory and deeper epistemological issues concerning the relationships between people's behaviour, data and wellbeing. Their struggle with new data practices demonstrates that building digital capabilities in a well-established field is precarious, many times generating new uncertainties instead of leading towards optimized services. Even though the fieldwork is somewhat dated, these dynamics remain unchanged: new data-driven practices do not simply emerge automatically with the introduction of digital technologies but must be actively developed within existing organizational structures through practices that are often partial, failing and characterized by continuous repair.

This is not to say that digital feedback loops are unsuccessful everywhere or that critiques of them are unfounded. A body of empirical research, along with numerous publicly discussed cases, details the devastating consequences and lock-ins that algorithmic technologies can produce, as in predictive policing. However, the case of behavioural insurance highlights the importance of examining these developments situationally, considering variations among technologies and (regulatory) contexts. Although it is frequently presented as the ‘worst-case scenario’ of datafication (Tanninen, 2020), insurance is an immensely complex field where legal, technical and affective aspects intertwine and complicate the introduction of innovation. Critiques that concentrate on the totalizing idea of behavioural insurance tend to overlook this complexity, framing insurance as driven solely by the aim to increase individual responsibility and yield maximum profit and disregarding, for instance, the collective aspects necessary for its function. This leaves us with only a partial view of emerging insurance practices, reinforcing a rather deterministic idea of the new technologies’ inevitable impact and overlooking the context-dependent, diverse and imperfect practices of datafication.

However, it should be noted that algorithmic operations do not need to be perfect to cause harm and limit possibilities for action – and this also applies to the insurance context. For instance, beyond the potential damages of digital feedback loops, the operations of Finnish insurers can already cause harm. The unsettled digital elements and actors introduced into the insurance arrangements engender uncertainties, anxieties and distrust (Tanninen et al., 2022b), disturbing the insurer-policyholder relationship, the purpose of which should be to generate security, thus raising questions about the meaningfulness of implementing such technologies. Examining algorithmic practices in a situated manner with close attention to contextual, legal and technological issues can help us identify these additional harms that may stem from the very flaws in the operations, allowing for more precise criticism.

In doing this, centralizing breakages as the focus of the analysis can be a powerful tool. Furthermore, it can help us critique future-narrowing effects that arise from the rampant techno-solutionism that is pushing various fields to view all problems and solutions in the same way, prompting both public and private actors to invest heavily in digital transformation projects. The entrance of these data logics into new fields can have – and has already had – serious consequences by increasing the dominance of new technological expertise, potentially reconfiguring the internal values and practices of independent domains (Stevens et al., 2024). However, as most of these projects are unsuccessful in producing what they promise (Kempeneer and Heylen, 2023; Oludapo et al., 2024), focusing on the situated data practices, and particularly on the various ways breakages and responses to them are involved in them, provides a sobering moment to consider the suitability of data-driven technologies in diverse contexts, raising a set of questions and concerns about their usefulness, effects and ethics. Therefore, concentrating on the fragilities of technology development helps poke holes in dominant technological visions while acknowledging their harms can open avenues to uncover alternative perspectives on the dominant narratives and thus assist in keeping the future open.