Addressing vulnerability in customer experience with AI-agents Free

With increasing global focus on improving well-being of people and the planet and attention turning to the United Nation's 17 goals (https://sdgs.un.org/goals), it is unsurprising that there are increasing calls for service research on improving well-being (Anderson and Ostrom, 2015; Field et al., 2021; Chandy et al., 2021; Ostrom et al., 2021; Rosenbaum et al., 2017). Forty-nine percent of the UK adult population showed one or more vulnerability characteristics in a recent report, illustrating that consumer vulnerability warrants attention from both service providers and policy makers concerned with well-being ^[1].

Consumers experiencing vulnerability may present at various stages in their life (Mende et al., 2024). We recognize such individuals might not directly purchase as “customers” and so adopt the nomenclature “consumer,” except when referencing service concepts, including customer experience (CX) journey. Research has examined consumers experiencing vulnerability (e.g. Cheung and McColl-Kennedy, 2019; Mende et al., 2024; Rosenbaum et al., 2017) and how those experiences might be improved (Field et al., 2021; McGraw et al., 2024; Mozafaria et al., 2022; Ostrom et al., 2021; Rosenbaum et al., 2017). However, more recent recognition of the contextual nature (Mende et al., 2024), scale and consequences of vulnerability underscores its significance as a managerial and policy concern that requires further theoretical and practical attention.

In parallel, advances in smart systems technology and artificial intelligence (AI) are revolutionizing how businesses and consumers interact (Davenport et al., 2020; Huang and Rust, 2021a, b; Kopalle et al., 2022; Mariani et al., 2022; McColl-Kennedy et al., 2019; van Doorn et al., 2023). AI defined by Shankar (2018, p. vi) is “programs, algorithms, systems or machines that demonstrate intelligence.” Consumers today are engaging with AI across diverse social, professional as well as personal spheres (Hermann et al., 2024). While AI technologies present significant opportunities to customize services according to preferences and needs, a prominent, growing concern is whether consumers in particular contexts are vulnerable to discriminatory bias from the marketplace (Hermann et al., 2024; Manyika et al., 2019). Risks of technology misuse leading to inequity, exacerbated power imbalances, social manipulation through misinformation and mental health harm are increasingly highlighted within information systems literature, alongside calls for its use for social good (Floridi et al., 2018; Panetta, 2019; Taddeo and Floridi, 2018; Tomašev et al., 2020).

Technology adoption impacts CX (Hoyer et al., 2020; Meyer and Schwager, 2007) particularly for consumers experiencing vulnerability. For example, Wünderlich et al. (2020) highlight channel design impacting experiences of vulnerable consumers while Čaić et al. (2018 p. 192) discuss healthcare robots as an “extended self”, augmenting “focal actor's capabilities.” As AI technology advances and systems adopt agentic AI, with its proactive capacity for autonomous operation, situation analysis, adaptation and workflow design (Purdy, 2024; Bornet et al., 2025), its capacity for human-like interaction and effective management of CX is improving (Larivière et al., 2024; McColl-Kennedy et al., 2019; Sidaoui et al., 2020). Aligning with these developments, personal AI-enabled agents (AI-agents) offer a practical, effective solution, via value co-creation between consumers and AI systems, for consumers to rebalance power and reduce their vulnerability in service exchange (Mozafaria et al., 2022).

Yet, despite the growing proliferation of smart systems based on AI technologies for consumers, there is a dearth of studies investigating consumers experiencing vulnerability in the ambit of designing and employing personal AI-agents or in relation to CX (Lo Presti and Maggiore, 2023; Mozafaria et al., 2022). Therefore, we envision our conceptual paper via the explanatory research question (Simsek et al., 2023): “How can personal AI-agents mitigate vulnerability and enhance CX?”

Agency theory offers a compelling approach to study the rebalancing of relationships between consumers experiencing vulnerability with service providers (Anderson et al., 2025). Agency theory is primarily concerned with the relationship between principals and agents, particularly in situations where there is potential conflict of goals, asymmetry of information and lack of control (Eisenhardt, 1989; Shapiro, 2005). These dynamics are relevant within service settings where consumers experiencing vulnerability interact with organizations possessing greater expertise, resources and influence (Abboud et al., 2023). In such contexts, the consumer (the principal) is dependent on the provider (the agent) to act in their best interest (Singh and Sirdeshmukh, 2000), thereby creating an imbalance in control, especially when they are in vulnerable contexts (either situational or due to personal characteristics; Mende et al., 2024). Furthermore, agency theory suggests that attitudes toward the interests of the principal can shift when relationships are outcome-based, meaning that organizational behavior may seek to reshape consumer goals toward a firm's interests (Eisenhardt, 1989). For example, an energy consumer with cognitive impairments may struggle to interpret complex billing information and whether a tariff change is beneficial.

To facilitate consumer agency, by rebalancing the relationship between consumers and firms, we highlight how (agnostic third-party provided or consumer-created) AI systems can act as intelligent “personal agents” for consumers experiencing vulnerability, providing a touchpoint for them to interact with all digital and some physical services (Lindgreen, 2025). We term such agents “personal AI-agents” that must represent consumer interests, that is, have alignment with values and intentions of the individual consumer when given authority to act autonomously and support their self-determination (Ibáñez et al., 2023). We argue that personal AI-agents would interact with a range of firms and their AI systems to generate best outcomes for consumers, thus reducing their vulnerability.

Thus, our theory-adaptation paper (Jaakkola, 2020) contributes in three important ways. First, we provide a conceptual framework characterizing personal AI-agents that empower consumers experiencing vulnerability (CXV), consisting of five design attributes; second, we offer an actionable design framework comprised of four design archetypes (Figure 1) to guide practice; and third, we provide a research agenda to guide future research in this important area.

To achieve these contributions, we adopted agency theory as a theoretical lens. First, we conducted an integrative critical review of vulnerability literature (Snyder, 2019). This provided a foundation for subsequent theory integration and adaptation (Jaakkola, 2020) in which we aimed for systematicity (universalism, transparency, completeness, saturation, connectedness and coherence (Simsek et al., 2023)). We ensured a degree of universalism via reflective evaluative discussions of the interdisciplinary research team (Simsek et al., 2023). For transparency, the explicated substantive boundaries involved a comprehensive in-depth synthesis of vulnerability literature (seeking completeness and saturation) (Simsek et al., 2023), in combination with integrative qualitative reviews of customer experience and agentic AI-systems across services and computer science literature, to “assess, critique, and synthesize the literature” (Snyder, 2019). Such an approach looks beyond the gap-spotting goal of systematic literature reviews and instead aims to theorize, built on consistent, coherent “narrative reasoning” (Jaakkola, 2020), facilitating development of the “integrative research framework” (Simsek et al., 2023). This approach expediated emergence of an updated perspective (Snyder, 2019, p. 335) on vulnerability and CX (which we term CXV) that we then align with AI-agents, via application of agency theory (Eisenhardt, 1989) as a theoretical lens, facilitating identification of connectedness between the concepts (Simsek et al., 2023). As such, we contribute to the limited work on AI technologies and vulnerability (Hermann et al., 2024; Poole et al., 2021). We subsequently detail what a personal AI-agent entails and discuss personal AI-agents forms, attributes and applications, to provide both our conceptual framework and design archetypes for AI-agents for consumers experiencing vulnerability (CXV).

Consumer vulnerability is impacted by the extent to which service providers recognize disadvantages consumers face and assist in their goal attainment (Rayburn, 2015). Vulnerability is characterized as a threat of being attacked and/or harmed, encompassing susceptibility to external stresses and inability to adjust to these stresses (Adger, 2006; Cheung and McColl-Kennedy, 2019). Introduced by Baker et al. (2005), the term “vulnerable consumer(s)” commonly describes individuals encountering a variety of challenging situations in consumption settings (Hill and Sharma, 2020) attributed to a range of phenomena (Table 1).

Wide-ranging recognition of vulnerability in consumer research has led to multiple definitions of consumer vulnerability and a lack of consensus (Halstead et al., 2007; Mansfield and Pinto, 2008; Mende et al., 2024; Spotswood and Nairn, 2016). Early work links vulnerability to consumer characteristics (Andreasen and Manning, 1990; Smith and Cooper-Martin, 1997). Consumers have commonly been labelled as “vulnerable” when exhibiting one or more “stigmatized personal or social characteristics” that can lead to disadvantages and/or discrimination in service encounters (Rosenbaum and Montoya, 2007; Rosenbaum et al., 2017, p. 310) such as children, elderly, low-income, poor education, minorities or obese (Hill and Sharma, 2020). However, Davey et al. (2023) and Raciti et al. (2022) argue that using consumer characteristics to define vulnerabilities is a deficit approach, as it treats consumer vulnerability as an inherent trait, resulting in stigmatization. Such an approach also implicitly suggests that vulnerability is a relatively stable state.

Further research highlights the issues of power and resource control (Hill and Sharma, 2020). Baker et al. (2005) introduced the idea of vulnerability as a contextual state of powerlessness occurring when an individual lacks certain elements of control, leading to market inequalities and subsequent difficulty in achieving consumption goals. Similarly, Shultz and Holbrook (2009) identified economic and cultural aspects of vulnerability, proposing two essential related consumer characteristics: knowledge of how to solve a problem and access to means of solving the problem.

There is some notable alignment now in multiple studies that recognize vulnerability to be dependent upon interaction of three main variables: personal characteristics, personal state and external conditions (Baker et al., 2005; Adkins and Jae, 2010; Rosenbaum et al., 2017; Overton and O'Mahony, 2018; Johns and Davey, 2019; Riedel et al., 2022). Hill and Sharma (2020) and Salisbury et al. (2023) further develop the contextuality aspect to argue that vulnerability is a dynamic state, with recent studies proposing consideration of both the experiential nature and contextual dynamism of consumer vulnerability (Dodds et al., 2023; Hermann et al., 2024). This aligns with the Universalist Theory of Vulnerability, which asserts that vulnerability is an inherent and universal condition; part of the human experience (Fineman, 2008). Vulnerability may affect individuals regardless of identity categories and is a fundamental aspect of all human life managed and shaped by social structures, frameworks and arrangements, including economic, educational and institutional networks (Fineman, 2008).

Mende et al. (2024) considered the consumer's vulnerability state in terms of its breadth and depth, paralleling Luna's (2019) Layered Theory of Vulnerability, which suggests vulnerability is a composite of distinct, contextual and relationally influenced layers of vulnerability making up the whole or total vulnerability of an individual (Luna, 2019). Thus, consumers may move between different vulnerability states; vulnerability may be temporary or permanent, depending on the individual and context (Baker et al., 2005; Mende et al., 2024; Riedel et al., 2022; Robertson et al., 2021).

We align with studies contending that consumer characteristics are just one aspect of vulnerability; more temporally transitive contexts of individual state and/or external conditions play a key role (Baker et al., 2005; Rosenbaum et al., 2017) and provide a more inclusive definition of vulnerability that can potentially apply to any consumer (Hermann et al., 2024). For example, an individual without traditional vulnerable characteristics might become vulnerable because they lack knowledge or power at a particular point in time (Abboud et al., 2023; Anderson et al., 2013; Rosenbaum et al., 2017).

To provide a comprehensive definition of vulnerability as a foundation to identify the scope for a framework for personal AI-agent design, we integrate key aspects of vulnerability from these multiple studies: personal characteristics, personal state, external conditions, power and resource access, dynamism and inclusion of all consumers (Finsterwalder et al., 2024; Hermann et al., 2024; Hill and Sharma, 2020; Johns and Davey, 2019; Mende et al., 2024; Salisbury et al., 2023; Smith and Cooper-Martin, 1997). Thus, we define consumer vulnerability as: “dynamic, subjective experiences of relative lower power and associated limitations to resource access in consumption settings, leading to economic, physical or psychological harm and/or reduced well-being through failure to realise all exchange goals, resulting from embodied characteristics, personal situation or external conditions” (see Table 1).

To provide foundational support for AI-agents capable of practically and effectively providing resources or capabilities to assist consumers in overcoming their vulnerabilities, we extend the breadth-depth vulnerability framework proposed by Mende et al. (2024). In our revised conceptual framework, breadth consists of eight categorical vulnerable consumer context factor types (each containing three to eight sub-indicators) within three dimensions (embodied characteristics and embedded, situation or external conditions), adapting those identified by Finsterwalder et al. (2024) and Mende et al. (2024) and extending the framework based on literature ( Appendix). Angled arrows in  Appendix denote depth; a continuous variable relating to an individual's extent of vulnerability within a particular factor subindicator. We respond to calls for greater conceptual clarity between contextually dependent and state-dependent vulnerability (Wünderlich et al., 2020) articulated in  Appendix. The revised framework provides guidance for AI-agent designers who must consider how to reduce the range (breadth) and depth of consumer vulnerabilities.

Building on Transformative Consumer Research (TCR) (Baker et al., 2005), Transformative Service Research (TSR) scholars examined consumer vulnerability from a service-oriented perspective (Raciti et al., 2022; Riedel et al., 2022). Such research increasingly focuses on experiences of vulnerability and its effects on consumer well-being (Finsterwalder et al., 2024; Rosenbaum et al., 2017). Consumer vulnerability is understood to be highly contextual and thus TSR scholars advocate for broader exploration of contexts, situations and factors contributing to such experiences (Raciti et al., 2022; Riedel et al., 2022). To understand CX for consumers experiencing vulnerability, we therefore integrate our expanded breadth-depth concept of vulnerability with the multi-dimensional construct of the CX (Lemon and Verhoef, 2016; McColl-Kennedy et al., 2019). CX is a holistic journey from the perspective of consumers comprised of various cognitive, emotional, behavioral, sensory and social responses to organizational offerings through interactions across one or more points of contact, known as touchpoints (De Keyser et al., 2015; Gahler et al., 2023; Lemon and Verhoef, 2016).

While much of CX research has focused on contextual factors for both the experiential response and organizational stimuli for vulnerable consumers (e.g. De Keyser et al., 2020), CXV permeates across all value creation elements in CX, including the context (McColl-Kennedy et al., 2019). McColl-Kennedy et al. (2019), highlight that early detection of vulnerability facilitates higher service satisfaction and ultimately value to both consumers and firms. Therefore, we adopt a wider conceptualization of CXV than viewing it as a consumer characteristic dependent concept, broadening it to a consumer agency-specific concept permeating across all value creation elements of CX (Cheung and McColl-Kennedy, 2019). Early and proactive identification and acknowledgement of vulnerability state(s) is crucial to adaptations in the value creation framework. For example, digitally vulnerable individuals need to be proactively identified for digital retailers to provide them with suitable servicescape experiences (Bitner, 1992). This highlights how specific depth and breadth of vulnerabilities can lead to unique experiential responses requiring adaptations in CXV design/delivery. Consequently, designing CXV requires adaptation of each value creation element by considering differentiated responses based on individual consumer circumstances that relate to the depth of vulnerability ( Appendix).

CXV can differ significantly due to perceived control leading to cognitive and emotional responses in the value creation framework (McColl-Kennedy et al., 2012, 2019). Drawing on agency theory as a theoretical lens, we identify the notion of perceived control (Baker et al., 2005), proposing it as a moderator permeating across the value creation elements to extend the CX framework specifically for consumers experiencing vulnerability. Furthermore, we show the opportunity offered by AI-agents as an intermediary to enhance perceived control and ultimately improve CX for consumers experiencing vulnerability.

Service interactions are rarely neutral exchanges. They are embedded in consumer and provider interactions and relationships marked by asymmetries of knowledge, resources and influence (Eisenhardt, 1989). The notion of agency, therefore, has implications for the well-being of all consumers engaged in exchange (Lamberton et al., 2024). Agency theory offers a theoretical lens for understanding these asymmetries and imbalances by analyzing how principals rely on agents to act on their behalf under conditions of asymmetry of information and potential goal misalignment (Eisenhardt, 1989; Shapiro, 2005). This perspective is particularly relevant for consumers experiencing vulnerability, as structural dependence on service providers can heighten risk of exploitation or neglect across the depth/breadth spectrum of vulnerability (Mende et al., 2024) (See  Appendix).

From a CX perspective, agency theory helps explain why consumers in vulnerable contexts or at risk of opportunism may report diminished confidence, autonomy or trust in service providers that may lead to negative emotional responses or poor perception of service quality (Cheung and McColl-Kennedy, 2019; Singh and Sirdeshmukh, 2000). Situational or contextual lived experiences of vulnerability can erode perceptions of control, intensifying experiential responses of dissatisfaction or harm (Finsterwalder et al., 2024). We argue that perceived control represents the experiential manifestation of these agency dynamics, providing a bridge between theoretical constructs of asymmetry (Einsenhardt, 1989) and the practical design of services that empower consumers experiencing vulnerability.

Perceived control represents a subliminal sense of influence or agency over outcomes from the perspective of the consumer (Baker et al., 2005) that aggregates through cumulative interactions with a service provider (Bellos and Kavadias, 2021). This sense of perceived control, or lack thereof, can fundamentally shape experiences and agency for consumers across the depth and breadth of vulnerability across CX value creation elements (McColl-Kennedy et al., 2019), thereby influencing service satisfaction and loyalty. We apply agency theory to highlight how the relationship between consumers and firms can be rebalanced to enhance perceived control, positively impacting CXV.

For consumers experiencing vulnerability, the issue of experiential goal alignment is often magnified. We contend that vulnerability heightens the risk of loss of perceived control due to potential goal divergence between the service provider and consumer. For instance, a consumer with cognitive impairment may struggle to fully understand a financial product, increasing their reliance on the provider's goals to act in their best interest. Similarly, in healthcare, patients with limited mobility may feel a lack of control over their interactions with providers, particularly when their preferences and movement needs may be overlooked in healthcare providers' design goals (McColl-Kennedy et al., 2012). Such scenarios exemplify the importance of perceived control for CXV. Adopting an agency theory perspective, we argue that perceived control should be actively designed into services for consumers experiencing vulnerability. For example, an energy provider might proactively recommend and implement the most appropriate tariff for older citizens or households, including individuals with disabilities, thereby mitigating feelings of helplessness and strengthening consumers' sense of perceived control. In this way, service providers should anticipate the control imbalances and asymmetries inherent in these interactions (Shapiro, 2005) and design interventions to achieve appropriate levels of perceived control that shift the balance toward consumers to reduce their experiences of vulnerability.

Taken together, based on the elements of information asymmetry and goal divergence from agency theory (Shapiro, 2005) and building on and extending McColl-Kennedy et al.’s (2019) CX framework, we propose that perceived control acts as a moderator across all elements of the value creation framework for CX from the perspective of a consumer experiencing vulnerability. Balancing this perceived control toward consumers is crucial to improve the CX of consumers experiencing vulnerability. This implies that touchpoints, value creation elements and responses for CX need to be adapted to the state of vulnerability of a consumer in such a way that it enhances perceived control. Thus, addressing perceived control in CXV requires both mitigating vulnerabilities and reconfiguring the dyadic relationships between consumers and providers. In short, by aligning goals and empowering consumers, service providers can enhance trust, satisfaction and long-term value creation by enhancing perceived control across each of the value creation elements in the service experience (Huang and Rust, 2018; Puccinelli et al., 2009).

The introduction of a personal AI-agent as an intermediary allows for the realization of enhanced perceived control for CXV. This dynamic highlights the possibility of a shift from the provider as an agent to a personal AI-agent that can be empowered to act as an intermediary on behalf of the consumer. By representing the interests of the consumer experiencing vulnerability, such AI-agents can reduce inherent imbalances and information asymmetries that otherwise erode perceived control, thereby mitigating feelings of helplessness or dissatisfaction (Baker et al., 2005). In this way, we propose an altered principal-agent relationship by introducing a personal AI-agent as an intermediary where the balance of control can shift toward consumers experiencing vulnerability. This means designing personal AI-agents that can act as advocates for consumers, offering simplified decision-making tools or designing services that align explicitly with the unique goals and constraints of consumers. Prior to describing design characteristics of such personal AI-agents that can enhance perceived control for CXV, we begin by defining and describing a personal AI-agent.

The notion of personal AI-assistants is not new. Wooldridge (2009) gives an example of a Personal Digital Assistant (PDA) configuring package holidays. Similarly, Saad et al. (2017) consider Virtual Personal Assistants automating daily tasks, and Searls (2012) describes Vendor Relationship Management systems acting for consumers to manage engagement with suppliers. While the concept of AI existed as early as 1955 (McCarthy et al., 2006), the scope and definition of AI is still largely undefined. We view an “AI system” as defined in the EU artificial intelligence act 2024 to be a “machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments” (Table 2, #1). This means that a wide range of systems are in scope for discussion, from non-interpretable “black-box” systems such as generative Large Language Models (LLMs) or Machine Learning based recommendation systems (Castelvecchi, 2016) through to interpretable and predictable rules-based systems which execute fixed instructions explicitly set by humans (Wooldridge, 2009).

An AI-agent is “a computer system that is situated in some environment, and capable of autonomous action in this environment in order to meet its design objectives” (Wooldridge, 2009, p. 16) where autonomous action is the capability of agents “deciding for themselves what they need to do in order to satisfy their design objectives” (Wooldridge, 2009, p. xi). Furthermore, we expect AI-agents to be intelligent, characterized with reactivity (able to understand, and effectively respond to their environment), proactivity (taking initiative to service consumers) and social ability (being able to interact with the humans they represent and other agents) (Wooldridge, 2009, p. 23).

We recognize two important environments. First, there is the set of other agents that an agent will interact with. These agents will typically be “service provider AI-agents” or other “personal AI-agents” representing different individuals. This is a typical multi-agent system (MAS) construction where agents communicate “not simply by exchanging data, but by engaging in analogues of the kind of social activity that we all engage in every day of our lives: cooperation, coordination, negotiation, and the like” (Wooldridge, 2009, p. xi). Searls (2012) suggests personal AI-agents should primarily perform intent casting in this environment, e.g. intent to fly from London to Boston, which service provider agents would bid to serve. Intent can also be broadcast to other personal agents; for instance, to arrange meetings with a specific person.

The second environment we consider is the “real-world” environment in which the agent interacts with the consumer. This environment is the set of inputs provided by the consumer and their auxiliary devices, and the means by which the agent responds or prompts. Auxiliary devices could range from air-quality sensors providing data to the agent at fixed intervals, through to a voice assistant the agent can interact with, or a humanoid robot controlled by the agent. Additional consumer and auxiliary data may be made available to agents with access to personal data stores such as a Solid pod (Sambra et al., 2016), enabling agents to access any digital information collected about consumers, within the bounds of consumer consent. More recently, Alvarez (2024) writing for Gartner defined Agentic AI as “Autonomous AI [that] can plan and take action to achieve goals set by the consumer” identifying this as the top strategic trend for 2025. However, Gartner envisions these agents taking on normative roles within organizations (Purdy, 2024; Zhi-Xuan et al., 2024), such as being integrated into a SaaS platform to replace some of the functions of a customer service representative. This is not in alignment with our vision of personal AI-agents as separate from service providers and strictly representing the “best interests” of the consumer.

Consumers interact with AI systems using various modalities. LLMs such as ChatGPT are often presented via chat-like interfaces. Embodied AI (EAI) (Franklin, 1997) is increasingly adopted. EAIs are AI systems with some form of physical embodiment, be it a webcam and screen providing a visual interface, or a full humanoid system with sensors capturing all five human senses: sight (vision), sound (hearing), smell (olfaction), taste (gustation) and touch (tactile perception). We consider both embodied (e.g. voice assistant) and non-embodied (e.g. clicking notifications to “confirm groceries”) personal AI-agents to be within scope. This multi-modality is a crucial feature when building personal AI-agents for individuals with heterogeneous vulnerability. We expect that consumers will interact by delegating control to agents using rules-based “authorization controls” (South et al., 2025), and within these bounds, a neurosymbolic system is “aligned” to the consumer (Zhi-Xuan et al., 2024). This alignment could be preferentialist – attempting to precisely emulate consumer decision-making behaviors (Baber, 2011); thick aligned – learning consumers or social values and making rational decisions following those values; or fulfilling a normative role such as being the consumer's counsellor or administrative assistant (Bai et al., 2022; Ouyang et al., 2022).

To summarize, a personal AI-agent is an agent operating within a multi-agent system of personal AI-agents and service provider agents. The ideal personal AI-agent must represent consumer interests, that is, have alignment with the values and intentions of the individual consumer when given authority to act autonomously and support their self-determination (Ibáñez et al., 2023). The scope, or granularity, at which agents may act autonomously is to be consumer-defined. If deployed at scale, we anticipate a range of preferences that consumers will have for the degree of autonomy they wish to delegate.

Building on the above notions of vulnerability, personal AI-agents and our integrative review (Snyder, 2019) of the CX literature, we introduce a CXV conceptual framework. We draw on key papers from service research on CX (e.g., McColl-Kennedy et al., 2019), AI in service (e.g., Huang and Rust, 2018) and service robots at the frontline (e.g., De Keyser and Kunz, 2022) to support our framework. We combine the notion of perceived control from agency theory and an information systems perspective of agents (e.g., Wooldridge, 2009) to derive design archetypes for personal AI-agents. The integrative framework subsequently developed (Simsek et al., 2023) comprises five design attributes that realize enhanced agency and perceived control for CXV. Each service design attribute provides potential benefits for CXV linked to consumer agency (Table 3), compensating for aspects of consumer vulnerability identified in our foundational vulnerability conceptual framework (CXV) (Table 3 and  Appendix), supported by corresponding vignettes from practical examples, guidelines and standards (Table 2). Furthermore, we propose a set of design archetypes of personal AI-agent roles based on the notion of control (whether an AI-agent is a provider-controlled personalized agent, or a consumer-controlled personal agent) and goal alignment in service of customer journeys for consumers experiencing vulnerability (Figure 1).

A personal AI-agent can be misconstrued for a personalized AI-agent or a service robot, as defined in the extant literature (Blut et al., 2021; De Keyser and Kunz, 2022). A critical distinction lies in the inversion of control: a design attribute that shifts the locus of control from the service provider to the consumer. In traditional personalization paradigms, service providers typically dictate the personalization process, using consumer data to tailor interactions and experiences (Blümel and Jha, 2023; Mittal and Lassar, 1996). While this can enhance customer satisfaction, it inherently positions the provider as the decision-maker, with the consumer playing a passive role. These risks outcomes are balanced toward providers based on the notion of goal divergence from agency theory (Eisenhardt, 1989). In contrast, a personal AI-agent empowers the consumer by granting them control over their journey and its personalization (Saad et al., 2017; Santos et al., 2016). Such a personal AI-agent acts as an intermediary or advocate, enabling consumers to control preferences, make semi-autonomous decisions and engage with services on their own terms (Searls, 2012). Inversion of control is particularly significant to realize design of a personal AI-agent, as it addresses the key moderator of perceived control for CXV. Thus, consumers could regain control over their interactions, mitigating risks associated with vulnerability and enhancing CXV with an AI-agent acting as an intermediary to balance the outcomes on behalf of a consumer who may be experiencing vulnerability. A practical example is proposed by vendor relationship management, wherein software acts on behalf of a consumer independent of the various vendors (Table 2, #2). Here, the consumer's agent oversees which vendors have access to data vis-à-vis the vendors offering preferences based on consumer data gathered. Inversion of control shifts agency from the provider to the consumer, improving emotional and cognitive responses in CXV and reducing the principal-agent goal divergence through an AI-agent as an intermediary.

A personal AI-agent should be designed to act on behalf of the consumer by receiving delegated authority to make decisions or perform actions. This is important for consumers experiencing vulnerability, as it reduces cognitive and emotional burden of navigating complex service systems. When delegated authority is granted by consumers, personal AI-agents must ensure that the consumer's preferences, needs and intentions are accurately represented and advocated during interactions with service providers (Searls, 2012). Accurate representation is achieved by aligning personal AI-agents with consumers' preferentialist alignment has agents closely emulate a consumer's actions, while thick alignment has agents make rational decisions supporting the consumer's value system (Zhi-Xuan et al., 2024). However, Zhi-Xuan et al. (2024) suggest that agents should instead align by emulating normative societal roles, e.g., acting as a consumer's manager, psychologist or travel planner. Delegated authority and decision-making alignment are key design attributes to realize various archetypes of frontline service technologies (De Keyser et al., 2019), introducing a mechanism for agents to autonomously act on behalf of consumers. By acting as an authorized intermediary, personal AI-agents can bridge gaps in goal alignment and comprehension arising due to vulnerability or lack of expertise. Algorithmic trading (e.g. ig.com) is an example where consumers delegate authority to personal AI-agents (Table 2, #3). Within constraints such as “never risk losing more than $10000” the trading algorithm works to maximize consumers' net worth. With authorization bounds, lower-income consumers prevent risking emergency funds; whilst less financially literate individuals are given more opportunity to participate in financial markets, thereby enhancing perceived control in CXV. Hence, delegated authority as an AI-agent design attribute reduces a consumer's need to make a decision approval for personal AI-agents, making them more accessible to cognitively vulnerable consumers. Similarly, decision-making alignment ensures the best interests of vulnerable consumers are advocated for.

We contend that personal AI-agents must be sensitive to specific vulnerability contexts of consumers (Mende et al., 2024), tailoring their behavior accordingly. For instance, personal AI-agents should prioritize accessibility features for individuals with physical impairments or simplify decision-making processes for consumers with cognitive vulnerabilities. Personal AI-agents should be able to interface with customers through auxiliary devices adapted for physical vulnerabilities; for instance, a prototype being piloted by Neuralink (Table 2, #4) for visually impaired consumers might allow the individual's brain to control their interaction with digital services (Moeed, 2024). An understanding of the breadth and depth of vulnerability (see  Appendix) becomes a pre-requisite for personal AI-agents adaptation. Adapting to vulnerability as a design attribute could help address information asymmetry between consumer and provider (Eisenhardt, 1989). This balances perceived control in favor of consumers experiencing vulnerability, ensuring that asymmetries are translated into tailored responses to the consumer's situational vulnerability constraints, thereby reducing risk of misaligned decisions.

A defining attribute of a personal AI-agent needs to be its active avoidance of exploiting consumers' contextual vulnerabilities. Unlike traditional systems that might opportunistically capitalize on consumer behavioral data to upsell or influence decisions (Hildebrand and Bergner, 2019), personal AI-agents should operate ethically, provide safeguards, act in alignment with consumers (Zhi-Xuan et al., 2024) and actively protect consumers from opportunistic behaviors in service. This addresses goal divergence, identified within agency theory (Singh and Sirdeshmukh, 2000).

Personal AI-agents must engage in transparent decision-making processes and clear communication about potential risks. Complimentary to the “adapting to vulnerability” design attribute the presentation of information to and decisions requested from consumers, must be appropriately managed, so that vulnerabilities are not inadvertently exploited. Finally, personal AI-agents must be designed to prevent side-effects of opportunism, such as the overcollection and misuse of sensitive data. Kwaai.ai (Table 2, #5) proposes that consumers can retain their own personal and behavioral data, stored, for instance, in a consumer-controlled personal data store such as a Solid pod (Sambra et al., 2016). Narrow signals of consumer “intent,” e.g., to buy a couch, would be sent by the Kwaai.ai agent to the market to engender tendering to supply; rather than service providers collecting wider personal data to infer and market to, a (vulnerable) consumer's needs. Thus, the personal AI-agent design attribute “does not exploit vulnerability” establishes an explicit safeguard against opportunism, a core concern in agency theory when one party possesses more information or power (Berners-Lee and Witt, 2025). Ultimately, knowledge of this explicit design characteristic in personal AI-agents interaction could enhance consumers' perceived control and positively impact CXV.

Finally, as a practical design attribute, personal AI-agents should be interoperable to ensure a wide range of services are accessible to enhance perceived control of CXV. Web browsers are instances of universally interoperable agents providing access to billions of websites; transport and display of web pages made possible with shared HTTP and HTML standards respectively. To maximize its utility, a personal AI-agent should adopt common standards, (e.g. Agent2Agent (A2A) for inter-agent communication; the Model Context Protocol (MCP) for AI tool connection) enabling seamless interaction with other AI-agents and service systems (Table 2, #6, #7). This would ensure consumers do not have access to services limited by their choice of personal AI-agent. Furthermore, to avoid vendor lock-in to personal AI-agent providers; consumer data should be stored separately from the personal AI-agent in user-controlled storage, accessed using standards such as the Solid pod specifications (Sambra et al., 2016). If consumer data is not stored separately, standards to port user-data between agents must be implemented (e.g., Open Banking (Table 2, #8)). This enables consumers to switch personal AI-agents, combine functionalities and maintain continuity across diverse touchpoints whilst privately retaining their history of interactions with service providers. Interoperability as an AI-agent design attribute reduces lock-in and over-dependence on a single agent or provider, mitigating informational asymmetries that limit consumer choice and perceived control.

Based on design attributes of personal AI-agents, we propose design archetypes of AI-agents providing service to consumers experiencing vulnerability, structured around two key dimensions: (CX) goal alignment and the perceived control between service provider and consumer. These dimensions are drawn from agency theory (Shapiro, 2005) to capture how well service provider goals align with delivering CXV and whether perceived control of personal AI-agent is predominantly consumer-led or provider-led. The design archetypes of AI-agents are structured along two conceptual axes by applying the preceding framework. The first axis, degree of control, stems from the attribute of inversion of control. Here, the balance of control can be contiguous between personalized services (balanced toward service providers) and personal AI-agents (in favor of consumers). The second axis reflects on the role of the AI-agent intermediary by considering the level of goal alignment between service providers and consumers. The four roles thus emerge from combinations of these axes, representing distinct designs of how AI-agents can advocate on behalf of consumers experiencing vulnerability.

Depending on the balance of control, an AI-agent can be designed to play one or more of these roles based on the service context. Figure 1 illustrates the four archetypes we propose: Service Orchestrator, Protective Sentinel, Autonomous Ally and Reliable Intermediary. Each role reflects a combination of goal alignment and control dynamics adhering to design elements for CXV for addressing specific needs and vulnerabilities of consumers. At every touchpoint, AI-agents can determine the level of goal alignment: consumer vis-à-vis service provider. Accordingly, AI-agents can be designed to perform specific role types (with provider control on the left and greater consumer control on the right) or it could be configured to adapt its role and persona to one or all four roles, depending on the service context and associated relative significance of inversion of control.

The “Service Orchestrator” role highlights the AI-agent's capacity to act as a coordinator, orchestrating personalized services in alignment with consumer goals while maintaining provider control. It serves as a facilitator, ensuring customer-centric delivery by seamlessly managing resources and touchpoints. The design attribute of interoperability is key to the success of an AI-agent in the service orchestrator role. For example, an AI-powered healthcare navigator could assist vulnerable patients with complex medical needs. Such a system, managed by a hospital, could streamline appointment scheduling, coordinate with specialists and ensure follow-ups. This orchestration ensures that patients with limited mobility or cognitive challenges receive consistent and comprehensive care, even though the provider retains control over the system's broader functioning. A tool like Apple Siri too acts as a co-ordinator across multiple services with a natural language interface (Table 2, #9).

The “Protective Sentinel” takes on a defensive stance, safeguarding consumers experiencing vulnerability from harm by monitoring and restricting potentially exploitative interactions. This role assumes a gatekeeping function, focusing on protecting consumers' well-being, given lower levels of goal alignment and control skewed toward service providers. In this role, there is greater emphasis on the attribute to not exploit vulnerability in the design of such AI-agents. For example, a personal AI-agent interfacing with online banking for older adults could monitor for signs of phishing or fraudulent transactions. By detecting unusual activities or requests, agentic services offered by email providers such as Proton or Privado can block or flag suspicious transactions, providing an additional layer of protection for consumers who might otherwise be susceptible to exploitation (Table 2, #10).

The “Autonomous Ally” role emphasizes a personal AI-agent fully under consumer control, fostering trust by adapting to the unique needs of consumers experiencing vulnerability. Unlike AI systems driven by provider interests, the Autonomous Ally empowers consumers by granting them agency through inversion of control in interactions. The design attribute of delegated authority needs greater emphasis in this role, so that autonomous functions of the AI-agent are authorized by the consumer. Thus, an AI assistant designed for individuals with neurodiverse needs could adapt its communication style and sensory stimuli to match the consumer's preferences, enabling better engagement. A self-created personal AI-agent that has greater autonomy could allow for control over specific cognitive and visual sensory functions in ways that can align directly with consumer needs and would be an example of an autonomous ally. A practical illustration of the Autonomous Ally role can be seen in low-code platforms such as n8n, which enable individuals to create their own AI-enabled agents: e.g., an openly available workflow integrates an LLM with Google Sheets to build a personal finance tracker in just minutes (Table 2, #11). Beyond its technical functionality, this example demonstrates inversion of control that allows consumers to design AI-agents that align with their unique goals and contexts, rather than relying on provider-controlled solutions.

The “Reliable Intermediary” serves as a neutral mediator between the consumer and the service provider, balancing potentially conflicting goals while ensuring fairness and control. This role is particularly valuable in contexts where consumers experiencing vulnerability need equitable access to services but may lack the means or confidence to negotiate terms themselves. In this role, a greater emphasis is on the “adapts to vulnerability” design attribute, e.g. a personal AI-agent might mediate accommodation options for students with disabilities. By working with both the student and the institution, the agent ensures that the student's needs are met, such as providing lecture transcripts or extended test times, while aligning with institutional policies. Similarly, an AI mediator for individuals with limited literacy could help bridge communication gaps with service providers, such as healthcare or financial institutions, ensuring that goal alignment is achieved without bias. For example, the flow neuroscience headset (Table 2, #12) acts as a non-invasive neutral intermediary in the form of a physical headset to support patients suffering from depression.

Taken together, our proposed design archetypes highlight how AI-agents can be designed with inversion of control built to adapt their roles to address the diverse needs of consumers experiencing vulnerability across different service contexts. Each design archetype offers a distinct combination of goal alignment and control dynamics, tailored to enhance perceived control to deliver improved experiences. By leveraging these archetypes, AI-agents can be designed and adapted to enhance CXV. The design archetypes not only advance the conceptual understanding of AI-agents in CXV but also provide a practical framework for designing AI systems that prioritize consumer well-being while balancing objectives of service providers. With this dual emphasis on practical design and goal alignment, our proposed framework provides a critical step forward in shaping future personal AI-agents in service settings for consumers experiencing vulnerability.

In conclusion, our conceptual paper makes a supporting foundational contribution relating to vulnerability, sets out two main contributions to theory, identifies implications for practitioners and presents a future research agenda. We begin by discussing the theoretical and practical implications and then consider future research.

Extant research in experiential marketing centers on examining experiences of mainstream consumers with AI technologies (Castelo et al., 2023; Longoni and Cian, 2022). However, it largely overlooks consumers experiencing vulnerability in their interactions with AI technologies, as well as how AI-agents can be effectively designed and integrated to represent them, act on their behalf and advocate in their interests (Lo Presti and Maggiore, 2023; Mozafaria et al., 2022). The foundational contribution lies in the synthesis of the extant vulnerability literature to propose an extended, comprehensive definition. This advances theoretical clarity and enhances the applicability of vulnerability as a construct for all consumers, across diverse contexts, providing scholars, service providers, personal AI-agent developers and policy makers with a broader and more nuanced lens for recognizing and addressing consumer vulnerability.

Our core theoretical contribution is the development of a novel conceptual framework providing structured guidance for designing personal AI-agents to cater to CXV (with CXV drawing upon the extended definition of vulnerability). Our conceptual framework, which includes five design attributes, acts as a call to the academic community to further theorize and conduct empirical investigations into personal AI-agent design, aimed at ensuring greater control by consumers over technologically mediated experiences. Our second theoretical contribution is the formulation of design archetypes for AI-agents; offering direction to agnostic third-party designers in developing personal AI-agents capable of representing the needs of consumers experiencing vulnerability across diverse service contexts. Finally, we offer practical implications for a diverse set of stakeholders and outline a future research agenda.

Our work articulates practical implications for stakeholder groups that include service designers, service providers and policy makers. First, we respond to calls for contemporary experiential marketing research to take a multi-stakeholder perspective (Hillebrand et al., 2015) of the experience of consumers experiencing vulnerability. The study addresses the implementation and deployment of personal AI-agents for consumers and presents implications for service designers, managers and policy makers. The soaring number of laws and regulations to protect vulnerable consumers amplify the challenge faced by both service providers and policy makers. Our five design attributes and four archetypes in our design framework assist in designing personal AI-agents to effectively tailor to the needs of consumers experiencing vulnerability and offer guidance toward challenges that can be encountered by service providers.

Second, from a service design perspective, designers must develop services that interact not only with human users but also with AI-agents functioning as consumer intermediaries. As AI-agents increasingly serve as interfaces between consumers and broader service ecosystems, service designers need to build AI-agent-facing interfaces capable of communicating, negotiating and adapting dynamically (Zheng et al., 2023). This shift reflects the inversion of control attribute, wherein the AI-agent rather than the human becomes the principal point of contact and service design must shift from consumer-facing experience to agent-facing service infrastructure (Gibbons and Vallejo, 2025). This implies designers will need to design for both human consumers and their delegated personal AI-agents.

Third, AI-agent designers must ensure interoperability and portability, especially when AI-agents function as Service Orchestrators operating across different service ecosystems. In this role, AI-agents coordinate multiple service relationships while safeguarding consumer freedom of choice (Gibbons and Vallejo, 2025). For example, a personal AI-agent managing energy plans should be capable of seamlessly transferring the consumer preferences and usage history to a new provider with minimal friction. To achieve this, designers should adopt open technical standards, ensure data portability and implement modular agent interfaces that enable smooth integration across platforms. These measures should prevent digital exclusion and mitigate the risk of ecosystem lock-in, ensuring consumers maintain autonomy and mobility.

For service managers, we highlight three considerations. First, service managers will need to reorient their goals and adapt systems to work with personal AI-agents designed with inversion of control rendering greater agency to consumers. For instance, the concept of the intention economy (Searls, 2012) takes a radical perspective, proposing that in addition to reduced market interactions, consumer facing advertising becomes redundant. Instead, personal AI-agents operating on behalf of consumers, learn preferences and “signal” to markets what solutions the consumer intends to purchase, freeing up consumer time and reducing cognitive load. Managers will need to adapt service systems that can bid to fulfil consumers' needs at the best value.

Second, personal AI-agents may align with one or more of the four archetypical roles described in Figure 1, depending on the consumer's context. For example, in contexts that require safeguarding of consumers from potentially exploitative interactions, such as upselling attempts exceeding CXV risk thresholds or cognitive capacities, AI-agents should be designed following the Protective Sentinel archetype, embodying the “does not exploit vulnerability” principle. In response, service managers must cooperatively align engagement strategies with ethics and agent-compatible communication, acknowledging that autonomy and protection are now embedded in the consumer interface (Reinhardt, 2023).

Third, service managers must prepare to operate in agent-mediated markets, where AI-agents act with delegated authority, engaging directly with service platforms on behalf of consumers. In such markets, Autonomous Allies and Reliable Intermediaries emerge as active participants in service exchange rather than passive technological tools. For instance, a subscription-based platform may receive purchase requests from multiple consumer agents, each representing households with diverse vulnerability profiles. The platform must discern which actions can be executed autonomously and which should be escalated for human verification to ensure ethical and safe outcomes. Establishing such operational capability may require implementation of governance guardrails, use of simulation-based training to model AI–human interactions and creation of ethical assurance roles within CX management teams to safeguard consumer interests.

Policymakers must establish clear accountability boundaries across agent developers, deployers, end users and service-provider platforms as delegated authority significantly complicates liability attribution, e.g. if a consumer's AI-agent purchases an unsuitable product, policy must determine whether fault lies with the agent's designer, the provider's offer, a regulatory body, or the user's delegation. Such scenarios require regulatory frameworks (Table 2, #1, #13, #14, #15, #16) accommodating layered responsibilities, especially in sensitive areas such as finance, health, or utilities (Díaz-Rodríguez et al., 2023). Moreover, the four archetypes (Figure 1) provide a foundation for risk-based regulation. Autonomous Allies decentralize control enabling consumer autonomy but raise concerns about consent and transparency. Similarly, a Protective Sentinel restricting access to high-risk credit products may necessitate regulatory scrutiny to ensure fairness/non-discrimination. The archetypes support differentiated obligations aligned with the EU AI Act 2024 and similar risk-tiered regimes (Table 2, #1).

Policy makers must also develop regulatory standards requiring service providers to develop open technical standards enabling all personal AI-agents to interact with their services. Consumers must be able to transfer personal data and preferences between personal AI-agent providers without loss of functionality, autonomy, or service quality, to promote open markets and prevent consumers from becoming “locked-in” to specific personal AI-agents. Furthermore, service providers should be required to interact with all certified third-party personal AI-agent providers to promote open markets. Embedding OECD principles of fairness, transparency and contestability within regulatory design would ensure AI-agent systems remain accountable, inclusive and aligned with human-centric governance norms (Table 2, #13).

Governments could consider deploying such personal AI-agents as a digital public infrastructure (DPI) for public good (Table 2, #14). This would ensure access for consumers experiencing financial vulnerability, limiting the threat of social disparity where only wealthy have agentic access. Furthermore, personal AI-agents’ capabilities should be regulated to ensure more exclusive AI-agent offerings do not cause detriment to consumers using other personal AI-agents: e.g. preventing exclusive personal AI-agents from having first choice of high-demand concert tickets. Additionally, AI-agent providers need to be “agnostic”: regulated to ensure personal AI-agents do not possess “competing interests,” i.e. do not offer services which agents interact with, not have any commercial deals, shareholding or other similar conflict of interest with service provider companies.

To implement market regulation, policymakers could establish a certification framework similar to the UK Digital Identity and Attributes Trust Framework (Table 2, #13), established to provide digital standards of certification, identification and verification. Personal AI-agent service providers could be accredited, showing they adhere to accepted design attributes (such as those in Table 3). Policymakers should also continue to demand strong data protection via accountability measures for “oversharing” of data from personal AI-agents to services, in addition to accountability for more traditional data breaches.

In this conceptual, theory-adaptation article (Jaakkola, 2020) we use agency theory as a theoretical lens to consider CX for vulnerability (CXV) to propose (1) an expanded breadth–depth vulnerability structure for CXV, (2) five personal AI-agent design attributes (inversion of control; delegated authority and decision alignment; adapts to vulnerability; does not exploit vulnerability; interoperability/portability) and (3) four agent archetypes differentiated by locus of control and goal alignment. While the paper applies an overarching method, theory and reviews of related bodies of literature, the CXV conceptual framework's five design attributes and the design framework's four archetypes (in particular inversion of control), require empirical testing of their validity, completeness and practical effectiveness. This should enable service researchers to influence the design of AI services, and we outline future research opportunities and potential research approaches in Table 4.

Shifting to an AI-agent-dominated service market requires a radical shift in the operation of consumer markets. This presents key challenges, and we identify three overarching research propositions for researchers to explore ethical tensions relating to the deployment of personal AI-agents. Furthermore, we identify specific CXV design and AI-agent role-related research questions to combine and extend our insights. First, stakeholders face a shift in value-creation mechanisms within services with the introduction of personal and service provider AI-agents in the ecosystem. Therefore, researchers should consider:

1. “How should consumers, providers, regulators and service firms determine appropriate levels of sacrifice of personal control?”.

Second, the proposed inversion of control will lead to significant operational challenges as all stakeholders learn how systems and processes must evolve and thus:

1. “How should stakeholders manage transparency, trust and accountability?”.

Beyond avoiding AI-agents exploiting vulnerability (our proposed explicit design attribute), we must also consider the risk of externalities impacting wider vulnerabilities. Researchers should consider how to design AI-agents to reduce vulnerability more broadly in society and avoid contributing to greater inequality or exploiting vulnerabilities of individuals beyond the consumer being represented. A network of AI-agents maximizing well-being and CXV of individual consumers could still harm wider society. Unregulated, AI-agents are likely to leverage any existing social, political or financial power a consumer has, against or relative to, other actors. Therefore, researchers, policy makers and agent service providers need to address ethical questions on safeguards and standards to maintain responsible AI-agents. Increasing dependence upon technology could lead to increased personal skill loss through delegation (Heap, 2015), making consumers more dependent and vulnerable. AI introduction could also significantly impact the employment of humans (Huang and Rust, 2018), leading to negative outcomes for society (Wirtz et al., 2018). Similarly, replacement of human carers may increase isolation and loneliness of care-dependent individuals, increasing vulnerability. Finally, utilization of LLM AI systems has significant carbon cost implications (van Wynsberghe, 2021) and unchecked, these could have proportionally more impact on more vulnerable countries and communities (Islam and Winkel, 2017; Markkanen and Anger-Kraavi, 2019). While AI offers significant potential to reduce the frequency and impact of vulnerabilities in service settings, the achievement of this goal will be challenging. Therefore:

1. “How can society prevent failure to reduce overall vulnerability/avoid contributing to greater inequality?”.

Finally, we outline specific questions related to design, role and adoption of personal AI-agents in Table 4 that potentially require diverse and interdisciplinary work across human-computer interaction, service blueprinting, as well as embodied AI (for multiple modalities and AI manifestations). For example, what are the barriers and enablers to the adoption of AI-agents for CXV, and how can AI services be designed and implemented to improve meaningful adoption? We encourage future research along these exciting avenues with participation from policy makers, community leaders, service providers and academic scholars to shape and determine the future for personal AI-agents in service.

We thank Cambridge Service Alliance, University of Cambridge for conducting the forum on the future of service management. The author team acknowledge the contribution of Mohamed Elmasry of Tactful AI at the initial scoping meetings in the forum.