Analyze recent research on AI adoption in organizations and identify important theoretical or empirical gaps that remain underexplored. Explain why each gap matters and suggest how it could contribute to the existing literature.

1. From adoption intention to sustained organizational use

Recent research has substantially improved understanding of why organizations adopt AI, but adoption is still frequently operationalized as a binary organizational condition or as employees’ intention to use a system. This obscures the transition from experimentation to routinized, organization-wide use. The distinction is important because AI adoption may produce little value when firms remain in pilot stages, use systems only sporadically, or fail to integrate them into established workflows. Lee et al. show that performance gains in high-tech ventures emerge only after adoption reaches sufficient intensity and is accompanied by complementary investments such as cloud computing, databases, and internal R&D [1]. Similarly, Heimberger et al. identify a gap between technological and organizational readiness and actual implementation in manufacturing firms [2]. Together, these findings suggest that readiness and adoption should not be treated as equivalent outcomes.

The underexplored theoretical issue is therefore the temporal process through which AI moves from exploration to routinization, scaling, renewal, or abandonment. Research could contribute by developing process models that distinguish adoption stages and specify the organizational mechanisms connecting them: complementary assets, process redesign, managerial attention, data maintenance, employee learning, and governance. Longitudinal panel studies, repeated surveys, and comparative case research would be especially valuable. Such work could also test whether the determinants of initial adoption differ from those of sustained use, rather than assuming one set of predictors applies throughout the adoption lifecycle.

2. Explaining heterogeneous outcomes within the organization

The literature increasingly recognizes that AI adoption has both beneficial and harmful consequences, but it remains insufficiently precise about why outcomes differ across employees, teams, functions, and occupations. The multilevel review by Bankins et al. [3] shows that AI research spans individual, group, and organizational levels, including collaboration, perceptions of capability, worker attitudes, algorithmic control, and labor-market effects. Yet these levels are often examined separately. For example, organizational adoption can enhance self-efficacy and thriving when employees interpret AI as helping humans, but it can also increase anxiety when they perceive it as replacing humans [4]. In a different negative pathway, perceived organizational adoption can trigger professional identity threat and knowledge hiding, although a strong learning climate weakens this association [5].

This evidence points to a major gap in cross-level theorizing. Organizational adoption is not experienced uniformly; its effects depend on job design, occupational identity, managerial communication, participation in implementation, perceived control, and opportunities to acquire new capabilities. Future studies should connect organizational-level adoption decisions to team-level work redesign and individual-level appraisals using multilevel and longitudinal designs. This would contribute a more contingent theory of AI adoption, explaining not only whether adoption occurs, but also when it generates empowerment, adaptation, resistance, or concealment.

3. Moving beyond predominantly positive and linear adoption models

Much adoption research still assumes that greater readiness, trust, institutional pressure, or adoption intensity leads progressively to better organizational outcomes. The evidence is more complex. Hong et al. find that AI adoption improves manufacturing performance, but the effect varies with strategic orientation: excessive exploration, exploitation, or even ambidexterity can weaken the performance benefit through instability, rigidity, or coordination costs [6]. Hopf et al. likewise show that implementation is shaped by a tension between “craft” and “mechanical” conceptions of data-science work, helping explain why many projects fail to move beyond the pilot stage [7]. These findings challenge linear models in which adoption itself is treated as the primary source of value.

The theoretical gap concerns the boundary conditions and possible non-linearities of AI adoption. Researchers should examine inverted-U relationships, threshold effects, capability traps, and trade-offs between experimentation and standardization. Adoption intensity may improve outcomes up to a point and then generate coordination burdens, surveillance, skill erosion, or governance overload. Empirically, this requires measures that capture depth, breadth, integration, and quality of use rather than a single adoption indicator. Such work could link technology adoption theory with dynamic capabilities, attention-based views, organizational learning, and implementation theory.

4. Integrating trust, governance, and institutional context

Trust is now recognized as central to AI adoption, but the literature does not yet adequately explain how trust is produced, maintained, or revised through organizational governance. Qualitative evidence indicates that attitudes toward AI can shift from negative or instrumental to positive as users gain experience and observe benefits; however, trust also becomes more calibrated as users learn about technological limitations [8]. Research on generative AI in the workplace similarly identifies trust as a strong predictor of usage and links it primarily to perceptions of reliable data-management practices rather than to abstract explainability [9]. At the organizational level, institutional pressures encourage generative AI adoption, but policy uncertainty and innovative culture moderate these effects [10].

The remaining gap is a more dynamic and politically informed theory of organizational trust. Existing studies often treat trust as an individual attitude or a relatively stable organizational resource, whereas AI governance involves changing regulations, accountability arrangements, data practices, professional norms, and power relations. Governance research emphasizes risks including opacity, privacy violations, bias, information leakage, and control failures [11], but these risks are not always connected empirically to adoption trajectories or everyday use. Future research should examine how trust is recalibrated after errors, breaches, regulatory interventions, or visible successes. Comparative studies across sectors and regulatory environments could show whether formal governance substitutes for interpersonal trust, reinforces it, or becomes a source of resistance. This would connect technology acceptance research with institutional theory, organizational trust, and responsible-innovation scholarship.

5. Accounting for sectoral, geographical, and organizational heterogeneity

The field has identified many general drivers of AI adoption—data capabilities, infrastructure, organizational readiness, leadership, competitive pressure, and institutional support—but often assumes that their effects are broadly transferable. Evidence increasingly contradicts that assumption. Song et al. find that common drivers operate across manufacturing and services, while the relevant infrastructure differs by technological intensity: software infrastructure matters more in high-tech industries, whereas hardware and equipment conditions are more important in low-tech industries [12]. Family-firm research further shows that external network ties facilitate adoption, but passive family ownership and active family management have opposing moderating effects [13]. Studies of SMEs also identify distinct configurations involving compatibility, resources, culture, regulation, ecosystem conditions, and infrastructure [14].

The empirical gap is insufficient comparative testing of these contingencies. Much evidence remains concentrated in particular national settings, industries, firm-size categories, and digitally advanced organizations. Consequently, it is difficult to determine whether a finding represents a general mechanism or a context-specific association. Future research should use harmonized cross-country and cross-sector datasets, multigroup structural models, hierarchical models, and configurational methods that allow multiple adoption pathways. This would contribute to a more context-sensitive TOE framework and prevent policy recommendations designed for large, digitally mature firms from being generalized to SMEs, rural firms, public organizations, or low-technology sectors.

6. Clarifying the role of human capabilities and collective learning

Research increasingly identifies skills, knowledge sharing, networks, and training as conditions of AI adoption, but the mechanisms through which these capabilities create value remain underdeveloped. Studies of scientific research show that AI adoption is pioneered by domain researchers with an exploratory orientation who are embedded in networks containing computer scientists, experienced AI researchers, and early-career researchers [15]. In organizational settings, AI knowledge sharing is associated with adaptability partly through workforce dynamic capabilities and human–AI collaborative tasks, particularly when boundary-spanning leadership is present [16]. Training also predicts adoption and innovative work behavior, with personal innovativeness shaping the relationship [17].

The gap is that “skills” and “training” are often treated as aggregate readiness variables rather than as evolving collective capabilities. This leaves unanswered which forms of expertise matter, how domain and technical knowledge are combined, and how organizations convert individual learning into reusable routines. Future work could distinguish technical, domain, evaluative, and governance-related AI literacies and examine how they interact at individual, team, and organizational levels. Network analysis, longitudinal learning studies, and research on communities of practice could show how knowledge travels across functional boundaries. The contribution would be a stronger capability-based explanation of adoption in which AI implementation depends not merely on possessing skilled employees, but on organizing their knowledge into collaborative and repeatable practices.

7. Establishing stronger causal evidence about adoption outcomes

A final gap concerns causal identification. The literature reports associations between AI adoption and innovation, operational performance, sustainability, and ESG outcomes. For example, AI-adoption intensity is linked to innovation performance through mobility, interactivity, and autonomy affordances, with data quality strengthening the relationship [18]. Firm-level studies also report positive relationships between AI adoption and green investment, ESG performance, and supply-chain environmental outcomes [19][20][21]. However, many studies rely on cross-sectional surveys, self-reported adoption measures, or observational panel data in which adoption may itself be caused by prior performance, managerial quality, or unobserved digital capabilities.

This matters because the field risks overstating the independent effect of AI. Stronger designs should combine administrative performance data with detailed measures of implementation timing, use intensity, complementary investments, and organizational change. Difference-in-differences designs around staggered implementation, instrumental-variable approaches, field experiments, and event studies could improve causal inference. Researchers should also test negative and null outcomes, not only productivity or sustainability gains. This would move the literature from demonstrating that AI adoption correlates with desirable outcomes toward identifying the conditions under which adoption actually causes them.

Overall, recent research has moved beyond simple technology-acceptance explanations by incorporating organizational readiness, trust, governance, networks, employee responses, and complementary capabilities. The most important next step is theoretical integration: AI adoption should be modeled as a dynamic, multilevel, and context-dependent process whose consequences depend on how technologies are implemented, interpreted, governed, and embedded in work.