AI in service industries: striking a balance between value co-creation and co-destruction

Artificial Intelligence (AI) has become a central driver of managerial decision-making, reshaping how organizations coordinate activities and create and capture value (Bock et al., 2020; Kumar et al., 2024). Across service industries, AI is increasingly embedded in core decision and coordination processes rather than remaining peripheral to strategy. Recent evidence indicates that over 60% of service organizations have integrated AI into such processes, underscoring its growing relevance for managerial practice (IBM Institute for Business Value, 2025).

Despite this rapid diffusion, AI-related initiatives generate highly heterogeneous value outcomes (Castillo et al., 2021; Kaartemo and Helkkula, 2025; Li and Tuunanen, 2022). While some implementations enhance coordination, efficiency, and learning, others produce unintended consequences, implementation frictions, or value erosion (Singla et al., 2025). Prior research identifies important organizational conditions that distinguish successful from problematic AI adoption, including strategic coherence, sustained investment in skills, and cultural adaptation (Cristofaro and Giardino, 2025). Yet as AI systems increasingly learn in use, operate with growing autonomy, and exert forms of algorithmic agency (Murray et al., 2021; Leonardi, 2025; Raisch and Krakowski, 2021), value formation cannot be reduced to organizational inputs or design choices alone.

Under these conditions, AI acts not merely as a technical tool but as an interactional actor that shapes courses of action in relation to human actors (Wirtz and Stock-Homburg, 2025). Value formation therefore unfolds through ongoing human–AI relations, in which practice alignment (Echeverri and Skålén, 2021; Skålén et al., 2015) and the distribution of agency (Murray et al., 2021; Leonardi, 2025) are continuously reconfigured through relational enactment (Mele and Russo Spena, 2025). Heterogeneous value outcomes are thus not episodic anomalies but manifestations of evolving interactional conditions within human–AI service systems.

Research on AI-mediated services documents diverse value configurations, including value co-creation (VCC) and value co-destruction (VCD) (Castillo et al., 2021; Chandra and Rahman, 2024; Li and Tuunanen, 2022). However, these outcomes are often examined as discrete interactional results rather than as dynamically interconnected configurations that stabilize, overlap, or shift over time (Kaartemo and Helkkula, 2025; Krakowski, 2025). Explaining the persistence and transformation of heterogeneous value configurations therefore requires a processual perspective capable of accounting for how interactional conditions evolve within distributed human–AI systems.

The Interactive Value Formation (IVF) perspective provides a foundational lens in this regard. IVF conceptualizes value as emerging from the alignment or misalignment of practice elements (procedures, understandings, and engagements) across interacting actors (Echeverri and Skålén, 2011, 2021; Schau et al., 2009). It clarifies how distinct value configurations arise through situated interaction. Yet IVF offers more limited theoretical guidance on how these configurations become dynamically interconnected (Plé, 2017) and recursively reconfigured in contexts where agency is redistributed and interaction is structurally exposed to competing demands.

To address this limitation, this study draws on paradox theory, which conceptualizes tensions as persistent and interdependent forces that endure over time and are continually enacted in practice (Smith and Lewis, 2011; Putnam et al., 2016). A paradox perspective shifts attention from isolated outcomes to the recursive negotiation and temporary stabilization of competing demands (Lewis and Smith, 2022). In AI-mediated service systems characterized by distributed agency and socio-technical interdependence, such tensions constitute structural conditions that shape whether value configurations stabilize, destabilize, or transform.

By integrating IVF with paradox theory, this study advances a processual account of value formation in human–AI service systems. While IVF specifies the interactional conditions through which value configurations emerge, paradox theory elucidates how persistent tensions reshape practice alignment and agency distribution over time. Together, these perspectives explain heterogeneous value configurations as patterned effects of evolving interactional dynamics rather than as isolated successes or failures.

Accordingly, this paper addresses the following research question:

To answer this question, the paper develops a conceptual model that positions practice-element alignment and agency distribution as structural and interactional conditions of value formation and conceptualizes paradoxical tensions as generative dynamics shaping recursive reconfiguration. In doing so, the study offers a theoretically grounded and managerially relevant lens for understanding how AI-enabled service systems sustain or disrupt value over time. The paper proceeds as follows: Section 2 reviews the literature; Section 3 presents the conceptual model; Section 4 provides illustrative examples grounded in prior evidence; Section 5 further discusses the model; Section 6 concludes with theoretical and managerial implications; Section 7 outlines future research directions; and Section 8 addresses limitations.

This section develops the theoretical foundations of value formation in human–AI service systems. It first examines the evolving configurations of AI and the resulting redistribution of agency within service systems, outlining their implications for service interaction (Section 2.1). It then examines value formation through IVF, highlighting its explanatory strengths and limits in AI-mediated contexts (Section 2.2). Finally, it introduces paradox theory to account for persistent tensions that shape heterogeneous value configurations over time (Section 2.3).

Contemporary AI research distinguishes between automation-oriented systems and more adaptive systems capable of influencing action formation. Early applications operated largely under human supervision, positioning algorithms as predictive or rule-based instruments that supported efficiency while preserving human-centered agency (Raisch and Krakowski, 2021; Abou Ali et al., 2026). More recent developments involve increasingly autonomous, goal-directed systems capable of contextual perception, iterative learning, and coordinated task execution (Ren et al., 2025; Sapkota et al., 2026), including multi-agent arrangements exhibiting varying degrees of autonomy under governance constraints (Al-Bashrawi et al., 2026).

Scholars interpret this evolution as a redistribution of agency, understood as the capacity for goal-directed action within a given context (Patel, 2025), within human–AI configurations (Murray et al., 2021; Leonardi, 2025; Danatzis et al., 2025; Krakowski, 2025).

As algorithmic systems gain delegated autonomy in decision processes, agency becomes increasingly distributed across human and technological actors. In AI-mediated service settings, this redistribution extends decision influence beyond direct human control and reconfigures service interaction, positioning AI as an active participant in shaping shared understandings and influencing beneficiaries' value interpretation and evaluation (Korzyński et al., 2025).

Such developments challenge the longstanding assumption in service research that agency resides predominantly with human actors (Akaka and Vargo, 2014; Mele et al., 2021). As decision authority and intentionality increasingly extend to algorithmic participants (Murray et al., 2021; Leonardi, 2025), value formation cannot be reduced to human-driven alignment alone. Instead, it requires analytical attention to evolving configurations of human–AI agency and the interactional processes through which value configurations are stabilized, disrupted, or reconfigured over time (Kaartemo and Helkkula, 2025). Table 1 synthesizes the AI orientations and agency levels discussed above and links them to typical service settings and expected value risks across the four configurations. This redistribution of agency has direct implications for how practice elements become aligned or misaligned in interaction, thereby shaping the emergence of different value configurations.

IVF builds on service-dominant logic (Vargo and Lusch, 2004, 2016) by conceptualizing value as emerging through interaction among multiple actors while recognizing that such interaction may generate heterogeneous value outcomes (Vargo and Lusch, 2004, 2016; Skålén et al., 2015). IVF conceptualizes value as arising from the alignment or misalignment of practice elements (procedures, understandings, and engagements) across interacting actors (Echeverri and Skålén, 2021; Skålén et al., 2015; Schau et al., 2009).

Within this framework, VCC and VCD are intertwined dynamics (Plé, 2017). VCC emerges when alignment enables intended outcomes and strengthens value-in-context, whereas VCD occurs when misalignment undermines actors' goals and erodes experienced value for at least one actor within the service system (Echeverri and Skålén, 2021; Plé and Chumpitaz Cáceres, 2010). A third outcome, value no-creation, captures interaction episodes that produce no perceptible change in value-in-context (Makkonen and Olkkonen, 2017). Rather than signaling a decline, value no-creation denotes a dormant configuration in which resource integration remains weak due to coordination gaps or institutional constraints that limit value realization (Makkonen and Olkkonen, 2017).

In human–AI systems (Murray et al., 2021; Leonardi, 2025; Danatzis et al., 2025), however, the conditions of alignment are no longer negotiated exclusively among human actors but are partly structured by algorithmic decision logics. Algorithmic opacity, biased outputs, or limited contestability may reshape how procedures are enacted, how situations are interpreted, and how actors engage, thereby influencing the coherence of interaction within the service system (da Silva Coelho and Farias, 2025; Järvi et al., 2018). Alignment in AI-mediated service systems is thus partially conditioned by socio-technical architectures that shape what can be perceived, evaluated, and acted upon (Li and Tuunanen, 2022; Lumivalo et al., 2023). Although IVF clarifies the relationship between practice-element alignment and distinct value outcomes, it offers limited conceptualization of the structural conditions through which these configurations persist, become interconnected, or are reconfigured over time in AI-mediated contexts. This limitation calls for a complementary theoretical lens capable of accounting for the enduring and recursive forces that shape value formation beyond episodic interaction.

Paradox theory conceptualizes tensions as persistent and interdependent contradictions that simultaneously enable and constrain organizational processes (Smith and Lewis, 2011; Schad et al., 2016). Unlike dilemmas resolved through trade-offs, paradoxes comprise competing yet legitimate demands that endure over time and are continually reproduced in action (Putnam et al., 2016; Lewis and Smith, 2022). Accordingly, analytical attention shifts from discrete outcomes to the sustained coexistence of opposing forces and their recursive implications for organizing (Cunha and Putnam, 2019).

This lens addresses the limitation identified in the IVF perspective. Paradox theory complements IVF by framing enabling and constraining dynamics as simultaneously enacted within the same interaction system (Putnam et al., 2016; Lewis and Smith, 2022), thereby extending IVF with a structural lens on persistent and interdependent tensions. In this framing, IVF specifies the interactional conditions associated with different value outcomes, whereas paradox theory accounts for the persistence of competing demands and their recurrent effects on interactional alignment conditions over time.

Service research has long conceptualized value formation as embedded in persistent paradoxical tensions (Tóth et al., 2022), including efficiency versus relational depth (Mele et al., 2021), control versus empowerment (Frow et al., 2019), and standardization versus personalization (Edvardsson et al., 2011). In AI-mediated contexts, these tensions become inscribed in the socio-technical architectures that coordinate interaction and decision-making (Raisch and Krakowski, 2021; Murray et al., 2021; Leonardi, 2025).

Building on paradox theory (Smith and Lewis, 2011; Lewis and Smith, 2022) and service scholarship, this study focuses on four interrelated paradoxes that are particularly relevant for examining the persistence, oscillation, and transformation of value configurations in AI-mediated service systems.

This paradox captures the tension between substituting human effort to enhance efficiency and scalability (Davenport and Kirby, 2016) and augmenting human judgment through collaborative human–AI learning (Brynjolfsson and Mitchell, 2017; Wilson and Daugherty, 2018). Augmentation can enable VCC through structured human–AI collaboration (Raees et al., 2024). Rather than representing mutually exclusive design choices, these logics co-evolve through human-in-the-loop dynamics that recalibrate task allocation and expertise over time (Rahwan et al., 2019). Automation stabilizes consistency and scalability, while augmentation restores flexibility and contextual judgment (Putnam et al., 2016), rendering substitution and collaboration structurally interdependent within ongoing interaction.

This paradox reflects the tension between explainability and algorithmic complexity (Raees et al., 2024). Transparency underpins accountability and trust, particularly in high-stakes domains (Singh et al., 2024), whereas opacity enables computational depth and non-linear pattern detection (Raees et al., 2024). Opacity sources include system complexity, user illiteracy, and intentional design (Langer and König, 2023) positioning explainability as a negotiated relational practice through which algorithmic outputs become interpretable and actionable for organizational actors (Hamida et al., 2024; Engström et al., 2025). Efforts to increase transparency coexist with functional opacity, reinforcing the persistent character of this epistemic tension.

This paradox concerns the tension between procedural consistency and contextual responsiveness. Standardization ensures reliability and fairness (Rust and Huang, 2021; Kim and Kim, 2025) whereas personalization leverages predictive and generative capabilities to tailor interaction and emulate social presence (Chaturvedi and Verma, 2023). Excessive standardization may reduce engagement and experiential differentiation (Kim and Kim, 2025), while excessive personalization risks intrusion, perceived manipulation, and algorithmic fatigue (Ferraro et al., 2024; Grossetti et al., 2021). These competing logics remain interdependent, because scaling consistent experiences often constrains local responsiveness, while intensifying personalization increases coordination demands and exposure to user backlash.

This paradox governs agency allocation (Krakowski, 2025; Kaartemo and Helkkula, 2025). Unlike the automation–augmentation paradox, which concerns how tasks are executed and capabilities are distributed between human and AI actors, the autonomy–control paradox relates to how decision rights, authority, and accountability are allocated within the service system. AI systems exhibit delegated or apparent autonomy within human-defined constraints (Ren et al., 2025; Sapkota et al., 2026; Ågerfalk, 2020). Increasing algorithmic self-direction may generate perceptions of diminished human control (Faraj et al., 2018) and power displacement (Leonardi, 2025; Adam et al., 2025). Autonomy and control are recursively co-constituted through ongoing processes of delegation and attribution (Ciardo et al., 2020), sustaining a dynamic redistribution of agency within interaction (Leonardi, 2025; Murray et al., 2021).

This section presents a conceptual model of value formation in human–AI interactions (Figure 1). Grounded in IVF (Echeverri and Skålén, 2011, 2021; Makkonen and Olkkonen, 2017), it conceptualizes value as an emergent, context-dependent outcome of situated practices involving humans, technologies, and organizations (Herrmann and Pfeiffer, 2023). Value arises from the interplay of human intentionality and algorithmic performativity (Kaartemo and Helkkula, 2025).

The model is developed as a diagnostic lens that specifies the interactional conditions under which heterogeneous value configurations stabilize or destabilize in human–AI service systems characterized by distributed agency and persistent tensions.

The model integrates two dimensions:

Agency distribution (horizontal) captures how intentionality and decision authority are allocated between human and AI actors, whose agency operates as delegated, operational influence within socio-technical systems rather than autonomous human-like intentionality. Human agency is purpose-driven and context-sensitive, whereas AI agency is perceived, goal-oriented, and adaptive within defined constraints (Leonardi, 2025; Beck et al., 2022; Brandtzaeg et al., 2023; Murray et al., 2021). Agency may be concentrated in a single actor or enacted in a more distributed manner through reciprocal adjustment. For analytical clarity, the AI system, whether architecturally single- or multi-agent, is treated as a unified organizational counterpart within the service encounter.

Practice alignment (vertical) refers to coherence among procedures (rules/routines), understandings (shared knowledge/sensemaking), and engagements (motivation/relational energy) across interacting actors (Schau et al., 2009; Skålén et al., 2015; Caridà et al., 2019; Echeverri and Skålén, 2011, 2021). High alignment enables coordinated enactment; misalignment fragments interaction.

While the framework centers on human-AI interaction, in service organizations employees often act as interpretive and behavioral mediators between algorithmic outputs and service outcomes (Brynjolfsson and Mitchell, 2017). Frontline and professional employees translate AI recommendations into situated action as learning algorithms become incorporated into work practices (Faraj et al., 2018) and decide whether to comply with or override system outputs in the face of algorithmic control (Kellogg et al., 2020). These micro-level responses influence not only understandings (sensemaking and contestability) but also procedures (workarounds, escalation routines) and engagements (trust, perceived fairness, willingness to collaborate). Employee reactions can therefore become a pivotal mechanism behind configuration shifts. For instance, resistance or low uptake can keep the system in a no-creation state even when the technology is formally deployed, whereas constructive appropriation can re-distribute agency and strengthen practice alignment, supporting co-creation. Conversely, superficial compliance with opaque recommendations may temporarily stabilize value compliance while seeding longer-term misalignment and potential co-destruction. In this sense, employees are not only impacted by agency redistribution; they actively co-shape how agency and alignment are enacted over time.

Building on the interplay of these dimensions, the model delineates four value configurations that emerge from human–AI service interactions: co-creation, co-destruction, no-creation, and compliance.

VCC emerges at the intersection of distributed agency and high practice alignment. Human and AI actors assume complementary roles, pooling resources through iterative cycles of interpretation and recalibration (Rahwan et al., 2019; Mosqueira-Rey et al., 2023; Krakowski, 2025).

Human expertise provides contextual judgment and normative orientation, while AI contributes analytical processing and adaptive capabilities. This reflects human-in-the-loop dynamics and a hybrid form of agency (Mosqueira-Rey et al., 2023) in which algorithmic outputs and human interpretation are continuously recalibrated through interaction, sustaining coordinated decision-making and contextual responsiveness (Rahwan et al., 2019; Krakowski, 2025).

VCD materializes when distributed agency intersects with misaligned practices resulting in breakdowns in procedures, distorted interpretations, or unintended consequences despite the presence of collaborative intent (Echeverri and Skålén, 2011, 2021).

In line with IVF, co-destruction is conceptualized as an interactional outcome that presupposes reciprocal resource integration. Accordingly, in this framework VCD is analytically associated with distributed agency, where multiple actors exert goal-directed influence and misalignment emerges through their interdependent enactment. Such disruptions may stem from algorithmic opacity, biased data, automation failures, or shifts in perceived power, undermining shared understanding and accountability and weakening coordination and trust (Burrell, 2016; Langer and König, 2023; Leonardi, 2025; Engström et al., 2025). Although both actors enact goal-directed action, misaligned operational logics generate inefficient, exclusionary, or ethically problematic outcomes (da Silva Coelho and Farias, 2025; Ferraro et al., 2024). VCD thus reflects not the absence of agency or interaction, but the destabilization of alignment under conditions of distributed intentionality.

Value no-creation reflects configurations in which exclusive agency intersects with misaligned practices. Decision authority remains concentrated in a single actor, yet procedures, understandings, and engagements lack coherence across the interaction. Interaction persists without generating substantive resource integration or breakdown, resulting in an inert value condition in which potential remains unrealized (Makkonen and Olkkonen, 2017).

In AI-mediated service systems, this configuration often arises when systems are formally implemented yet remain only partially integrated into decision processes and everyday routines due to limited user engagement or organizational inertia (Makkonen and Olkkonen, 2017). Here, stalling typically reflects an interactional mismatch: users do not trust the outputs, cannot easily contest or interpret recommendations, or perceive that integration adds effort without commensurate benefits. As a result, adoption becomes superficial (AI is consulted sporadically, used mainly for reporting or symbolic and non-substantive use, or bypassed through parallel routines and workarounds) resulting in interaction that remains low-impact and inconsequential for value-in-context. Over time, prolonged no-creation can harden into inertia or trigger redesign efforts that re-align practices and redistribute agency. This configuration differs from value compliance in that interaction persists but remains substantively inert: AI is formally present yet weakly integrated into decision routines, resulting in minimal resource integration or experiential impact.

Value compliance captures interactions characterized by high practice alignment under exclusive agency. In this condition, procedures, understandings, and engagements are coherently enacted, but coordination is structured unilaterally rather than reciprocally. Value is procedurally delivered through rule-based execution and efficient task performance, without shared intentionality or reciprocal adaptation. This configuration typifies routinized service systems and algorithmic management settings, where one actor (human or AI) structures action while the counterpart assumes a supervisory or verification role (Heracleous and Wirtz, 2014; Engström et al., 2025). Although reliable and predictable, value compliance privileges consistency and control over dialogical learning and collaborative reconfiguration. It reflects effective but unilateral integration of practices, whereas no-creation captures situations in which interaction persists without translating into meaningful integration. These configurations represent provisional stabilizations of interactional conditions that may shift as practice alignment and agency distribution evolve within the same service system. Such shifts reflect the activation of paradoxical tensions inherent in human–AI service systems (Smith and Lewis, 2011). Among these, the tensions of automation–augmentation, transparency–opacity, standardization–personalization, and autonomy–control operate as generative mechanisms that reshape practice alignment and the allocation of agency over time.

Table 2 synthesizes the linkages between value formation conditions, shaped by the interplay of practice elements and agency distribution, and the paradoxical tensions that structure their dynamics. These linkages are neither rigid nor exhaustive; rather, they highlight analytically relevant tensions within each practice domain, without precluding overlap or alternative configurations.

Procedures encompass the recurrent doings and sayings that orchestrate coordination, reliability, and adaptation in human–AI interactions (Skålén et al., 2015; Echeverri and Skålén, 2021). They embody the automation–augmentation paradox, which reflects the structural tension between efficiency-oriented substitution and capability-enhancing collaboration within practice (Raisch and Krakowski, 2021). Automation stabilizes routines through reliability, scalability, and control (Davenport and Kirby, 2016; Brynjolfsson and Mitchell, 2017), whereas augmentation introduces adaptability, contextual discernment, and iterative learning (Raisch and Krakowski, 2021; Wilson and Daugherty, 2018).

When automated execution becomes increasingly embedded in human–AI interactions, procedures tend to consolidate around rule-based coordination, often corresponding to value compliance under conditions of high alignment and exclusive agency. When recurring exceptions or contextual variability require discretionary adjustment, routines reopen to augmented enactment, introducing interpretive flexibility that may sustain VCC when alignment is preserved, or contribute to VCD when coordination weakens (Echeverri and Skålén, 2021).

Over successive enactments, these adjustments sediment into procedural arrangements that stabilize coordination while preserving the underlying tension between automation and augmentation. Such stabilization remains provisional: renewed efficiency demands or contextual misfits may reactivate the paradox and prompt further recalibration (Krautzberger and Tuckermann, 2024), enabling shifts across value compliance, co-creation, and co-destruction over time.

Understandings encompass the sensemaking processes through which actors establish shared meaning in interaction (Skålén et al., 2015; Echeverri and Skålén, 2021). They embody the transparency–opacity paradox, reflecting the structural tension between the need for intelligibility and the inherent complexity of algorithmic reasoning (Burrell, 2016). Transparency supports accountability and shared understanding (Langer and König, 2023), whereas opacity enables computational depth, efficiency, and intellectual property protection (Raees et al., 2024).

Sufficient intelligibility allows algorithmic outputs to be meaningfully integrated into human judgment, reinforcing shared understandings and, under distributed agency and aligned practices, supporting VCC.

Greater opacity, by contrast, encourages deference to algorithmic authority without substantive integration. Agency recentralizes, and interaction stabilizes as value compliance. If engagement weakens further without overt breakdown, value no-creation may arise, as outputs are enacted but fail to significantly influence actors' interpretations or behaviors, leaving them experientially inconsequential (Makkonen and Olkkonen, 2017).

Persistent misunderstanding or contested interpretations under distributed agency strain coordination and may culminate in VCD (Echeverri and Skålén, 2021).

Alignment in understanding is therefore inherently contingent. Intensified accountability demands or contextual ambiguity can reactivate the tension between intelligibility and complexity, triggering renewed cycles of explanation and reinterpretation (Lewis and Smith, 2022) and enabling shifts across value configurations over time.

Engagements embody the motivational and relational dimension of practice, balancing the need for scalable and predictable service delivery with the demand for contextual sensitivity and authenticity (Edvardsson et al., 2011; Mele et al., 2021). They are structured by the standardization–personalization paradox, which juxtaposes procedural efficiency and control with emotional connection and context-sensitive responsiveness.

Standardization channels engagement through predefined scripts and automated interaction patterns. Participation becomes procedurally organized and agency tends to concentrate in the algorithmic system, often stabilizing interaction as value compliance under conditions of high alignment. If such standardization preserves procedural coherence but gradually attenuates motivational involvement, emerging misalignments may develop, particularly at the level of engagement, and interaction may drift toward value no-creation. In this configuration, exchange continues and procedural routines remain operational; however, weakened engagement and limited interpretive integration constrain meaningful resource integration, rendering participation increasingly inconsequential for value-in-context (Makkonen and Olkkonen, 2017).

Personalization introduces adaptive responsiveness and redistributes agency. When supported by aligned procedures and shared understandings, it can deepen relational involvement and sustain VCC. However, increased variability heightens coordination demands; if relational expectations outpace procedural or interpretive alignment, engagement may fragment, contributing to VCD (Echeverri and Skålén, 2021).

Across interaction episodes, adjustments prompted by scale pressures, user feedback, or contextual shifts reactivate the tension between consistency and responsiveness. These recalibrations may reinforce existing engagement patterns or destabilize them, enabling shifts across value configurations without resolving the underlying paradox (Smith and Lewis, 2011; Lewis and Smith, 2022).

Agency captures the structural allocation of decision rights and accountability through which action is governed within human–AI service systems. The autonomy–control paradox structures this dimension by expressing the enduring tension between delegated algorithmic autonomy and retained human oversight (Leonardi, 2025; Krakowski, 2025; Kaartemo and Helkkula, 2025).

Concentration of decision authority in a single actor, human or algorithmic, consolidates coordination around predictable execution and unilateral governance. Under high practice alignment, such concentration stabilizes interaction as value compliance. Where interpretive integration and engagement lack coherence with procedural arrangements across actors, interaction may remain formally organized yet strategically inert, contributing to value no-creation (Makkonen and Olkkonen, 2017). By contrast, redistribution of decision influence across human and AI actors introduces shared intentionality and adaptive responsiveness. Under aligned practices, distributed agency supports VCC by enabling reciprocal adjustment and iterative recalibration (Danatzis et al., 2025). However, expanded discretion simultaneously intensifies coordination demands and complicates responsibility attribution. In the absence of aligned practices, this ambiguity destabilizes interactional coherence and may contribute to VCD (Faraj et al., 2018; Leonardi, 2025).

The autonomy–control paradox unfolds through recurrent reallocations of decision rights, oversight thresholds, and accountability attributions within interaction (Leonardi, 2025). Human–AI service systems do not converge toward equilibrium; they adjust the balance between delegation and oversight in response to evolving coordination demands and institutional expectations. These adjustments shape agency distribution and influence the stabilization or disruption of value configurations (Echeverri and Skålén, 2021).

Drawing on IVF and paradox theory, the framework conceptualizes value formation in human–AI service systems as structured by enduring tensions within interaction. As these tensions are enacted, they reconfigure the interactional conditions underpinning value formation (Heracleous and Wirtz, 2014), thereby influencing whether value configurations stabilize or destabilize. In doing so, the model helps explain the emergence and transformation of heterogeneous value configurations over time. Co-creation, compliance, no-creation, and co-destruction (Echeverri and Skålén, 2021; Makkonen and Olkkonen, 2017) are thus understood as temporary configurations within an evolving interaction system.

Importantly, value configurations need not unfold uniformly across the service system. Multiple configurations may coexist across interaction domains and stabilize unevenly across stakeholder groups, reflecting the layered nature of human–AI interactions.

The examples discussed in the following section (Table 3) are introduced as theoretically informed illustrations designed to support sensemaking around the proposed framework. Rather than providing empirical generalization or causal testing, they are mobilized to illustrate how value configurations and their underlying tensions can be analytically interpreted within documented human–AI service interactions.

The examples are derived from peer-reviewed studies, reputable industry reports, and well-documented organizational deployments of AI in service settings. They were identified through targeted searches of documented AI implementations across service industries (e.g. banking, telecommunications, hospitality, human-resource management, and healthcare), complemented by tracing widely cited or extensively discussed cases in both academic and practitioner discourse.

Selection followed three guiding criteria. First, examples were selected to ensure coverage of the four value configurations (co-creation, co-destruction, no-creation, and compliance). Second, they were chosen for their capacity to make visible at least one of the paradoxical tensions shaping interactional dynamics. Third, preference was given to cases illustrating either configuration stabilization or transitions across configurations over time.

This approach privileges analytically rich and well-documented instances and may therefore result in a bias toward highly visible or extensively reported cases, often originating from Western or large organizational contexts. While this limits representativeness, it is consistent with the purpose of the examples as vehicles for theoretical exemplarity and sensemaking rather than empirical inference.

Each example is not presented as a standalone case description but is interpreted through a structured analytical reading grounded in the core dimensions of the framework: agency distribution, practice alignment, and the salient paradoxical tension. Agency distribution was assessed by examining where decision influence and evaluative authority predominantly resided within each interaction, distinguishing between more concentrated (human- or AI-led) and more distributed configurations. Practice alignment was inferred from the degree of coherence among procedures, understandings, and engagements, based on whether interactions enabled coordinated action, exhibited misalignment, or remained only weakly integrated into decision routines.

Value configurations were identified by analytically matching the observed combination of agency distribution and practice alignment to the four theoretically defined configurations. Specifically, interactions characterized by distributed agency and aligned practices were interpreted as co-creation, whereas distributed agency combined with misalignment indicated co-destruction. Configurations in which agency remained concentrated and practices were aligned were interpreted as compliance, while situations characterized by concentrated agency and weak or fragmented integration of practices were interpreted as no-creation. Configurations were identified at the level of specific service interactions rather than at the system level, allowing for the possibility that multiple configurations may coexist or shift over time within the same service system. The key paradoxical tension was identified as the tension that most visibly shaped the interactional dynamics and contributed to the stabilization or disruption of the configuration within each case. Where changes in system functionality, oversight arrangements, or interaction patterns were observable, shifts in agency distribution and practice alignment were used to infer transitions across configurations over time. Accordingly, configurations are interpreted as temporary stabilizations within an evolving interaction system shaped by ongoing paradoxical dynamics.

Bank of America's Erica is a virtual financial assistant embedded within the bank's mobile service system to manage standardized customer requests—such as balance inquiries, transaction searches, and basic account services—through scalable, rule-based execution (Chou and Chou, 2025). In these routinized exchanges, interaction unfolds through predefined prompts and system-generated outputs. Coordination is predominantly procedural, and decision influence remains concentrated within the algorithmic system. From the perspective of the framework, this arrangement reflects value compliance, as aligned procedures are enacted under concentrated algorithmic agency, ensuring reliability and efficient task completion.

Since its launch in 2018, Erica has supported billions of interactions and progressively incorporated proactive and personalized financial insights (Bank of America, 2025). As functionalities expanded to include spending alerts, budgeting guidance, and investment-related support, interaction differentiated across domains. Customers review algorithmic analyses, assess suggested saving targets, and evaluate financial notifications in relation to their own priorities (Fuscaldo, 2019). In these advisory exchanges, algorithmic outputs frame potential courses of action, while evaluative judgment and enactment remain with the customer.

Within the same service infrastructure, routine transactions and advisory engagements instantiate distinct yet coexisting interactional configurations. Standardized exchanges consolidate coordination under concentrated algorithmic execution, whereas advisory use involves more distributed decision influence and greater reliance on customer interpretation. The automation–augmentation tension remains structurally embedded in this differentiation: efficiency-oriented substitution and analytical augmentation operate in parallel without displacing human financial authority.

Amazon's AI résumé-screening tool was introduced to support large-scale candidate evaluation by contributing algorithmic rankings to recruitment decisions. Designed to operate within a distributed evaluative process, the system reproduced gender biases embedded in historical training data (Dastin, 2018). Because its decision logic remained opaque, recruiters were unable to meaningfully interrogate or contextualize the basis of its recommendations. Under these conditions, distributed evaluative influence combined with interpretive misalignment, corresponding to VCD.

Subsequently, algorithmic involvement was restricted to procedurally bounded administrative tasks, and evaluative authority was re-centered in human recruiters. Interaction consolidated around rule-based assistance under concentrated decision authority, reflecting value compliance.

The transparency–opacity tension becomes visible in this reconfiguration. Opaque outputs coupled with distributed evaluative influence strained interpretive alignment and weakened contestability. The restriction of algorithmic functionality to low-interpretive tasks under explicit human oversight narrowed the scope of delegation and consolidated coordination without extending collaborative evaluative integration. The same recruitment infrastructure thus accommodates distinct value configurations as agency and interpretive coherence are variably arranged.

Vodafone's TOBi illustrates how large-scale service architectures are shaped by the standardization–personalization tension. The chatbot's initial design privileged standardized procedures to ensure scalability and cross-market consistency across service interactions. As customer inquiries became more varied, predefined conversational flows struggled to accommodate contextual nuances, generating misinterpretations and repetitive escalation patterns.

Vodafone progressively enhanced TOBi through generative and retrieval-augmented AI, redesigned conversational journeys, and introduced SuperAgent to provide frontline employees with contextualized knowledge support. In Ireland, migration to IBM's watsonx Assistant further restructured the underlying architecture, contributing to fewer misunderstood utterances and faster iteration cycles (IBM, 2024).

From the perspective of the framework, these developments reflect differentiated value configurations within the same service infrastructure. Standardized automation continued to sustain procedural reliability, while generative augmentation and human interpretive checkpoints expanded contextual responsiveness. Rather than displacing one logic with another, the service system incorporated layered coordination mechanisms that recalibrated procedures, shared understandings, and engagement. The standardization–personalization tension thus remains constitutive of the architecture, structuring how consistency and contextual adaptation are variably enacted across interactional domains.

Japan's Henn-na Hotel initially delivered value compliance through robotic execution of check-in, concierge, and in-room services, supported by highly scripted procedures and concentrated algorithmic agency in routine operations (Reis et al., 2020). As robotic responsibilities expanded, limitations in contextual recognition generated misinterpretations—such as ambient noise being registered as commands—along with error cascades and recurrent human interventions (Shead, 2019; Skubis, 2025). These disruptions diffused accountability and weakened engagement patterns, corresponding to VCD.

Subsequently, the operational scope of robots was narrowed and human exception handling reintroduced. Decision authority was re-centered in human staff for non-routine situations, and interaction consolidated around more bounded automated execution. From the perspective of the framework, this reconfiguration reflects a return to value compliance under concentrated oversight.

The autonomy–control tension becomes visible across these arrangements. Expanded delegation of algorithmic autonomy exposed coordination vulnerabilities under constrained interpretive capacity. The subsequent narrowing of robotic scope recalibrated the allocation of agency without eliminating the underlying tension between efficiency-oriented automation and retained human control.

IBM Watson Health's oncology advisor initially provided protocol-based diagnostic support, offering rule-consistent treatment recommendations at scale (Singh et al., 2024). In high-volume case reviews, algorithmic outputs structured decision processes around standardized procedures, consolidating coordination under concentrated AI agency. From the perspective of the framework, this arrangement aligns with value compliance, characterized by procedural consistency and scalability.

In complex or ambiguous cases, opaque predictive outputs derived from deep neural architectures constrained interpretive transparency. Misalignments between algorithmic recommendations and clinical judgment eroded physician trust and strained shared understandings (Irgang et al., 2025). In some instances, clinicians formally reviewed system outputs without substantively integrating them into diagnostic reasoning, reflecting conditions of value no-creation. In others, contested recommendations and accountability ambiguities aligned with VCD as coordination weakened and engagement faltered.

Subsequent adjustments introduced layered explainability mechanisms—including visual rationales and confidence indicators—and formalized clinician veto rights within hybrid workflows (Hamida et al., 2024). These changes redistributed evaluative influence and reconfigured the interaction between algorithmic output and medical judgment. Rather than displacing one mode of coordination with another, the system incorporated structured oversight and interpretive checkpoints, recalibrating the autonomy–control and transparency–opacity tensions within clinical decision-making.

Across these service settings, value configurations do not appear as fixed outcomes but as temporary arrangements shaped by how agency and practice alignment are configured under persistent tensions. They may coexist within the same service system or be re-stabilized over time as interactional conditions shift.

The conceptual model synthesizes IVF (Echeverri and Skålén, 2011, 2021) with paradox theory (Smith and Lewis, 2011; Putnam et al., 2016; Lewis and Smith, 2022; Cunha and Putnam, 2019) to elucidate heterogeneous value outcomes in human–AI service systems. IVF specifies how configurations emerge from the alignment (or misalignment) of situated practice elements—procedures, understandings, and engagements (Schau et al., 2009; Skålén et al., 2015; Caridà et al., 2019)—and how these conditions relate to VCC, VCD, and value no-creation (Plé, 2017; Makkonen and Olkkonen, 2017). Paradox theory adds explanatory leverage by clarifying why competing demands remain persistent and interdependent, shaping interaction beyond episodic episodes and producing recurring shifts rather than stable resolutions (Smith and Lewis, 2011; Schad et al., 2016; Cunha and Putnam, 2019).

In the model, paradoxical tensions activate iterative reconfigurations through a recursive sequence: (1) tensions become salient in practice (e.g. opacity, bias, diminished trust) and disrupt the coherence of procedures, understandings, and/or engagements (da Silva Coelho and Farias, 2025; Järvi et al., 2018; Trincado-Muñoz et al., 2024); (2) actors respond through adjustments that shift practice alignment and/or redistribute decision influence, thereby reconfiguring agency distribution within the socio-technical system (Murray et al., 2021; Danatzis et al., 2025; Krakowski, 2025); (3) these shifts stabilize temporarily as a value configuration (VCC, value compliance, no-creation, or VCD), until renewed pressures reactivate tensions and prompt further recalibration (Smith and Lewis, 2011; Putnam et al., 2016; Lewis and Smith, 2022).

The framework is intended as a diagnostic, processual lens rather than a deterministic causal model. Its explanatory power is strongest where organizations can observe and influence both practice alignment and agency distribution over time; where these levers are structurally constrained (e.g. limited contestability, entrenched opacity, or weak organizational capacity), recursive adjustments may be curtailed and configurations may persist without meaningful reconfiguration.

In services, AI accentuates inherent tensions–efficiency versus relational depth, control versus empowerment (Edvardsson et al., 2011; Mele et al., 2021; Frow et al., 2019; Tóth et al., 2022) - through algorithmic agency (Raisch and Krakowski, 2021). The model delineates four paradoxes in relation to practice elements and highlights that AI can function as a dual-edged force in resourcing and valuing (Kleinaltenkamp et al., 2023), potentially reinforcing asymmetries even when efficiency gains are achieved (Kaartemo and Helkkula, 2025; Li and Tuunanen, 2022; Lumivalo et al., 2023). The model's dimensions of horizontal agency (exclusive to distributed forms; Leonardi, 2025; Beck et al., 2022; Brandtzaeg et al., 2023) and vertical alignment help explain why configurations may remain unstable or shift over time.

Co-creation necessitates synergy, as exemplified by Erica's temporal cycling for augmentation (Rahwan et al., 2019; Mosqueira-Rey et al., 2023). The illustrative example shows both the possibility of movement across configurations and the conditions under which shifts become fragile. First, opacity and bias can destabilize alignment and precipitate co-destruction, even when procedures appear reliable. Watson Health's opacity-induced misdiagnoses undermined understandings despite procedural reliability, only partially mitigated by integration and explainability (Hamida et al., 2024; Singh et al., 2024; Irgang et al., 2025). Amazon's bias escalation highlights how delayed separation exacerbates VCD (Dastin, 2018; Burrell, 2016; Langer and König, 2023). Second, attempts to restore alignment through hybridization or upgraded automation can relieve some tensions while creating others, sometimes producing temporary latency resembling no-creation. TOBi's hybrid upgrades realigned standardization–personalization (IBM, 2024), but Henn-na's autonomy overreach incurred rollback costs, temporarily trapping the system in a no-creation-like latency (Reis et al., 2020; Shead, 2019; Skubis, 2025). Overall, these examples illustrate recursion: tensions can either sustain alignment through iterative adjustments or destabilize it over time. At the same time, they indicate clear boundary conditions, as vendor-driven opacity and limited organizational maturity/skills can constrain contestability and make temporal cycling or other responses difficult to enact in practice (Engström et al., 2025; Singla et al., 2025).

This paper develops a conceptual model to account for the recursive, paradox-driven reconfiguration of value formation in AI-mediated service systems, where agency is distributed across human and algorithmic actors and practice alignment remains dynamically contingent. By integrating IVF (Echeverri and Skålén, 2011, 2021) and paradox theory (Smith and Lewis, 2011; Putnam et al., 2016; Lewis and Smith, 2022; Cunha and Putnam, 2019), the model identifies four paradoxical tensions - automation–augmentation, transparency–opacity, standardization–personalization, and autonomy–control - as generative mechanisms that drive transitions across value configurations (co-creation, co-destruction, no-creation, compliance). Illustrative examples from banking (Chou and Chou, 2025; Fuscaldo, 2019), recruitment (Dastin, 2018), telecommunications (IBM, 2024), hospitality (Reis et al., 2020; Shead, 2019; Skubis, 2025), and healthcare (Singh et al., 2024; Hamida et al., 2024; Irgang et al., 2025) demonstrate how the framework can be used diagnostically and highlight boundary conditions in asymmetric or low-maturity contexts.

The model advances research on value formation in AI-mediated service systems through three interconnected contributions. First, it reframes value formation in AI-mediated service systems as a contingent outcome of interactional configurations, rather than as an inherently beneficial consequence of resource integration (Vargo and Lusch, 2004, 2016). By emphasizing how value emerges from the alignment or misalignment of practices under conditions of distributed agency (Murray et al., 2021; Danatzis et al., 2025), the model explains why similar forms of resource integration may lead to co-creation, co-destruction, or no-creation (Plé, 2017; Makkonen and Olkkonen, 2017). This shifts the analytical focus from the presence of interaction to the conditions under which interaction becomes generative, inert, or detrimental.

Second, the framework introduces agency distribution and practice alignment as two analytically distinct but interdependent dimensions that structure heterogeneous value outcomes in human–AI interactions. In doing so, it extends IVF's focus on procedures, understandings, and engagements (Schau et al., 2009; Echeverri and Skålén, 2021) by explicitly accounting for distributed agency between human intentionality and algorithmic performativity (Sapkota et al., 2026; Ren et al., 2025; Leonardi, 2025). This makes room for organizational actors, especially employees who mediate adoption, contestation, and workarounds, to shape whether AI becomes integrated into everyday routines or remains inert. The result is a more organizationally realistic account of socio-technical interplays than purely anthropocentric formulations (Kleinaltenkamp et al., 2023).

Third, by integrating paradox theory the model helps explain how distinct tensions operating across different dimensions of interaction jointly shape the dynamics of value formation.

Rather than converging toward resolution, these tensions influence task execution, interpretive processes, relational engagement, and the allocation of decision rights, thereby generating ongoing reconfigurations across value configurations.

In this sense, value configurations are not equilibrium outcomes but temporary stabilizations shaped by the recurrent activation of multiple, interdependent tensions.

Importantly, these stabilizations are not neutral in their effects. Different configurations privilege specific actors, forms of knowledge, or evaluative criteria, while marginalizing others. As a result, shifts across configurations also entail a redistribution of benefits and burdens across stakeholders, linking the processual dynamics of value formation to broader questions of responsibility in AI-mediated service systems.

The explanatory power of the framework depends on the extent to which interactional conditions allow actors to meaningfully participate in shaping human–AI relations. Recursive reconfiguration is more likely to unfold when actors can interpret and respond to algorithmic outputs, adjust their practices, and participate in the enactment of agency within interaction. Such conditions typically require minimum levels of organizational maturity, relevant skills, and governance capacity that support contestability, reflexivity, and meaningful engagement with AI systems. Conversely, where interaction is constrained by high levels of opacity, limited capabilities, or rigid governance structures, configurations may stabilize without substantial reconfiguration, remaining confined to compliance-based or inert patterns. The framework is therefore analytically most applicable in contexts where actors retain sufficient capacity to influence how agency is enacted and how practice elements become aligned over time.

Managers can use the model primarily as a diagnostic lens to surface where misalignments in agency and practices are emerging, and to anticipate the trade-offs that may accompany AI-mediated service interactions. Because AI-enabled service systems redistribute agency and reshape practice alignment, value creation and destruction are rarely uniform across stakeholders. The same configuration may improve efficiency and consistency for the organization while reducing employee autonomy and privacy (Khaksar et al., 2024), as well as perceived fairness, particularly in contexts characterized by algorithmic opacity and contested control (Burrell, 2016; Kellogg et al., 2020), and potentially undermining customer trust (Castillo et al., 2021; Bock et al., 2020). This perspective draws attention to the need to assess value outcomes across economic, relational, and societal dimensions and make explicit which stakeholder groups bear costs when tensions are “resolved” in practice. This plural view of value also sharpens accountability: choices about transparency, automation levels, and control thresholds become governance decisions with ethical consequences, not purely technical optimizations.

Building on this diagnosis, the framework surface a range of governance considerations in relation to tensions, including issues related to explainability and contestability in decision processes, as monitoring routines and “paradox audits” (Hamida et al., 2024), escalation and override arrangements in high-stakes settings, and the delineation of accountability boundaries as automation expands. Where vendor solutions are involved, contractual and implementation choices become relevant in shaping reversibility and potential lock-in. More broadly, phased implementation approaches can support organizational learning under constraints (e.g. starting from compliance-oriented uses before moving toward more distributed agency), while cross-functional governance arrangements are relevant for addressing emerging power asymmetries and coordination gaps (IBM Institute for Business Value, 2025; Cristofaro and Giardino, 2025).

As a conceptual framework, these implications are derived from theoretical synthesis and illustrative examples; therefore, targeted empirical testing is recommended to calibrate feasibility across organizational contexts and to specify boundary conditions.

Future research can strengthen and extend the framework by examining its process dynamics, boundary conditions, and stakeholder-specific value outcomes across contexts. Because the model is conceptual and processual, a priority is to study how configurations shift over time and under what organizational conditions such shifts become possible or constrained.

First, longitudinal and process-oriented studies could trace how paradoxical tensions become salient, how actors respond (e.g. through contestation, workarounds, redesign, or governance interventions), and how these responses reconfigure agency distribution and practice alignment over time. Such work can also examine the role of organizational maturity and capability in enabling or curtailing recursive adjustments, especially in complex service ecosystems where multiple AI systems interact (Singla et al., 2025; Cristofaro and Giardino, 2025).

Second, multi-stakeholder research could operationalize value plurality by comparing how customers, employees, managers, and regulators assess value outcomes under the same configuration. This would clarify when efficiency gains coincide with relational or societal costs, how accountability arrangements distribute benefits and burdens, and how these trade-offs shape trust erosion, bias perceptions, and contestability (da Silva Coelho and Farias, 2025; Järvi et al., 2018).

Third, comparative and sector-specific investigations can examine ethical and cultural variations in how paradoxes are experienced and managed, including differences in agency attribution, surveillance tolerance, and expectations of personalization across service contexts. This stream can refine boundary conditions for governance responses and improve the framework's relevance for low-resource settings where skill gaps constrain contestability and learning (Khaksar et al., 2024; Tóth et al., 2022).

Fourth, methodological extensions can complement these approaches by developing measures for key constructs (e.g. contestability, perceived autonomy, practice-element coherence) and testing moderators that condition shifts across configurations. Rather than positioning predictive accuracy as the primary goal, such designs can evaluate which mechanisms and governance responses are more or less feasible and effective under specific constraints, particularly in high-stakes contexts characterized by opacity (Hamida et al., 2024; Irgang et al., 2025) and in settings where bias has been shown to escalate over time (Dastin, 2018).

Advancing this agenda would strengthen the empirical grounding of the framework while preserving its core contribution as a diagnostic, processual lens for understanding AI-mediated value formation.

This paper is conceptual and relies on illustrative examples; accordingly, it does not provide empirical validation of the proposed dynamics. While paradox theory suggests that tensions may be reflexively addressed over time (Lewis and Smith, 2022; Krautzberger and Tuckermann, 2024), organizations may face cognitive pressures, capability constraints, and vendor-driven opacity in service ecosystems that limit contestability and meaningful oversight and reduce the feasibility of reconfiguration, potentially sustaining no-creation or amplifying co-destruction in fast-changing settings.

In addition, ethical spillovers, such as surveillance trade-offs in personalization (Khaksar et al., 2024; Ferraro et al., 2024) and deskilling risks in augmentation (Trincado-Muñoz et al., 2024), are acknowledged but not fully theorized across different stakeholder groups and institutional contexts, thereby limiting our ability to specify how such spillovers translate into stakeholder-specific value trade-offs over time.

Finally, the framework assumes some degree of organizational readiness to observe and influence practice alignment and agency distribution; where readiness is low, the model's diagnostic insights remain useful, but managerial responses may be constrained.

The authors acknowledge the support of the AI-WARE Research Centre (Artificial Intelligence for Wellbeing, Awareness, Responsibility and Equality), “Magna Græcia” University of Catanzaro, Catanzaro, Italy.