Artificial Intelligence in Organizations: A Systematic Review of Operationalizing Generative Artificial Intelligence Capabilities for Organizational Performance

You’re listening to “Artificial Intelligence in Organizations: A Systematic Review of Operationalizing Generative Artificial Intelligence Capabilities for Organizational Performance,” by B. MacHkour and A. Abriane. Published in 2026.

Abstract.

Between 2018 and 2025, artificial intelligence expanded alongside major market shifts, while generative artificial intelligence began to reshape organizational strategy and value creation across functions and sectors. This article offers an integrated account of organizational uses of generative models in marketing, operations, human resources, finance and governance and identifies four core capabilities of generative artificial intelligence, namely personalization, decision augmentation, automation and ecosystem orchestration. It links these capabilities to the adoption and, above all, to the intensity and quality of use. Following PRISMA 2020 and PRISMA-S, we conducted a systematic review of Scopus and Web of Science in English, with critical appraisal through the Mixed Methods Appraisal Tool and a SWiM narrative synthesis. From 1,284 records, 83 peer reviewed studies were retained.

Findings converge on a coherent architecture of mediating mechanisms. Dynamic capabilities, absorptive capacity, process and organizational reconfiguration and multi-level innovation form an organizational infrastructure that translates generative artificial intelligence into organizational performance outcomes. Responsible governance of artificial intel-ligence and data, grounded in transparency, human oversight, risk management and environmental vigilance, acts as a moderating buffer that aligns use with strategic objectives, amplifies benefits and contains risks. We also find a transversal role of this governance layer across functions and sectors through standard setting, supervision and accountability arrangements.

The study proposes an operationalized integrative framework with fourteen testable hypotheses and outlines avenues for future longitudinal and quasi-experimental research, together with managerial implications for turning local experimentation into durable strategic advantage.

Introduction.

Since late 2022, generative artificial intelligence (GAI) has diffused rapidly across organizations, generating major expectations in terms of value creation while also raising substantial governance and capability-building requirements. Recent estimates suggest that GAI could generate between

The associate editor coordinating the review of this manuscript and approving it for publication was Turgay Celik.

USD 2.6 and 4.4 trillion in additional annual value world-wide, while organizational adoption has accelerated sharply, with AI increasingly moving from experimentation to core business activities. At the same time, projected spending growth confirms that firms are committing substan-tial resources to data, models, and scaling infrastructures.

However, increased investment and diffusion do not auto-matically translate into innovation or sustained organizational performance. In line with absorptive capacity theory, the conversion of AI-related investments into measurable ben-efits depends on complementary efforts such as experi-mentation, skills development, organizational learning, and knowledge-integration routines. These capabilities shape firms’ ability to recognize, assimilate, transform, and exploit external knowledge embedded in models, vendors, and inno-vation ecosystems, thereby determining whether AI remains an isolated technological deployment or becomes a source of process innovation, product development, and more durable competitive advantage.

In parallel, more and more micro-level studies have emerged on the effects of generative artificial intelligence at the workplace. Real-world experiments show clear produc-tivity gains when employees use AI tools to support their tasks. A study conducted by in the customer contact center of a Fortune 500 software company, covering more than five thousand agents, shows that a generative-AI-based conversational copilot increases productivity by around 14%, with a particularly pronounced effect for less-experienced agents. Other studies conducted with consultants indicate an increase of about 12.2% in the number of tasks completed and an acceleration close to 25% for activities located at the frontier of AI capability, together with a clear improvement in quality when tasks are carefully matched to model capa-bilities.

These robust and replicated findings lend support to the view that GAI augments certain cognitive tasks and reorganizes the division of labor between humans and machines.

However, it is still unclear how the aggregation of these benefits will add up at the macroeconomic level, and they remain highly conditional. References and point out that turning micro-level gains into broader economic benefits requires extra investment, deep process redesign, and tight control of risks linked to data quality, security, bias and the reliability of outputs. In other words, improvements in aggre-gate productivity are far from guaranteed and depend heavily on sector and institutional context. At the same time, the regulatory framework is taking shape: the European AI Act introduces a phased rollout up to 2026–2027, which raises compliance and risk management requirements for organizational uses, especially in so-called high-risk sectors.

In terms of skills and employment, analyses by the IMF (2024) and the World Economic Forum (2025) highlight an acceleration in the demand for AI-related skills and a broad exposure of jobs, with nearly 40% of tasks potentially affected. These dynamics are accompanied by uneven distributional effects that make large-scale support and reskilling policies indispensable.

More broadly, prior digital transformations suggest that technologies create durable value only when embedded in coherent organizational architectures and governance arrangements. Similar patterns have been observed in other technology-driven transformations, such as the deployment of Internet of Things infrastructures in education and training, where data-driven management and automation reconfigure learning environments, pedagogical practices and resource allocation. These parallels reinforce the view that orga-nizational innovation, through new routines, structures, coor-dination mechanisms, and managerial control systems, often becomes the practical vehicle through which AI-enabled experimentation is scaled and stabilized.

Without such orga-nizational innovation, local productivity gains may fail to translate into sustained process improvement, capability renewal, and performance effects.

More broadly, evidence from innovation and strategic management research suggests that organizational innovation often acts as a bridging mechanism through which techno-logical investments are converted into competitive outcomes. It may also support broader innovation trajectories, includ-ing green innovation, by reshaping routines, coordination structures, and resource-allocation practices. This wider per-spective, observed in several European contexts, reinforces the need to analyze GAI not as an isolated tool, but as part of a broader organizational transformation logic.

Although generative artificial intelligence is now at the heart of strategic decision-making, its impact on organi-zational strategies remains heterogeneous and ambivalent. It can simultaneously support value capture while intensify-ing risks of inaccuracy, bias, and technological dependence, reorganize work while weakening certain skills, and generate a competitive advantage that is exposed to rapid erosion as these technologies diffuse.

The current literature, often fragmented by function or sector, also remains insufficiently connected to broader debates in strategic management and international business concerning organizational learning, R&D-based competitive advantage, organizational innovation, and the absorption of external knowledge. As a result, it still does not offer an integrative framework capable of explaining how, through which mechanisms, under what conditions, and with the sup-port of which organizational relays and boundary conditions, generative artificial intelligence reconfigures strategies and performance. We therefore rely on established systematic review standards (PRISMA 2020 and PRISMA-S) to ensure transparency, reproducibility, and comprehensive coverage, and we use MMAT and SWiM to align quality appraisal and narrative synthesis with heterogeneous study designs.

Our objective is to jointly map the mechanisms of action of generative artificial intelligence within organizations, as well as the organizational and institutional conditions that foster or hinder value creation.

Beyond integrating fragmented streams, this review con-tributes in three ways. First, it specifies a mechanism-based causal logic linking distinct GAI capabilities (personaliza-tion, decision augmentation, automation, ecosystem orches-tration) to adoption and intensity/quality of use under task–technology fit. Second, it theorizes four analytically dis-tinct organizational mediators (absorptive capacity, dynamic capabilities, process and organizational reconfiguration, and GAI-related innovation) and clarifies how future empirical work can disentangle their effects. Third, it delineates bound-ary conditions and downside risk channels under which GAI adoption may yield positive, neutral, or negative outcomes, thereby moving the literature from dispersed findings toward a testable, falsifiable framework.

This review also analyses the outcomes and risks asso-ciated with generative artificial intelligence in terms of operational performance and productivity, decision qual-ity, customer experience and competitive advantage, with the aim of developing an integrative conceptual framework articulating antecedents, GAI capabilities, mediations, and moderations, and of formulating testable hypotheses for future multi-sector quantitative validations of organizational performance outcomes. From this perspective, our work is organized around the following main questions:

• What are the mechanisms through which generative AI transforms organizational strategies?

• At which levels does this transformation take place – for instance at the level of functions, processes, or business models?

• Which organizational and institutional conditions effec-tively enable the conversion of generative AI uses into tangible outcomes in terms of organizational performance?

• Which mediators and moderators shed light on the ambivalence between value creation and exposure to risks?

• Which governance and scaling guidelines foster sustain-able and responsible value creation from generative AI technologies?

To this end, we begin by presenting the methodology adopted, detailing the PRISMA protocol, the selection cri-teria, and the coding procedures, after which we develop a thematic literature review. We then present and discuss the results, proposing a mapping of the mechanisms and risks, a synthesis of mediations and moderations, as well as an integrated conceptual model.

II. MATERIALS AND METHODS.

This study implements a systematic review by adopting the general architecture and methodological recommenda-tions proposed by and. The research question is formulated according to the Population–Concept–Context model proposed by. The population covers all types of organizations, including private firms, public adminis-trations, and other organizational forms, without sectoral restriction; the concept refers to generative artificial intelli-gence and to foundation models or large language models mobilized for strategic and managerial purposes, notably for personalization, decision augmentation, automation, and ecosystem orchestration; the context encompasses the main organizational functions, including marketing and customer experience, operations and supply chain, human resources and skills development, finance and control, as well as gov-ernance, across all regions.

Screening involved a title/abstract review, then full-text assessment. Two reviewers screened records independently, and disagreements were resolved through discussion until consensus was reached. Duplicate records were removed prior to screening. Eligibility was then appraised on the 323 full-text records, which led to n = 240 records being excluded because they had a purely technical focus without an identifiable organizational context (n = 132), provided insufficient data (n = 94), or corresponded to an ineligi-ble document type (n = 14). The final sample therefore comprised 83 studies, which were synthesized thematically. The complete selection process is reported in the PRISMA 2020 flow diagram (see FIGURE 1).

In line with systematic review standards, the robustness of the retained corpus is supported by comprehensive Scopus/Web of Science cov-erage, PRISMA-S reporting of search strategies, duplicate screening, and MMAT-based quality appraisal. The review was reported in line with PRISMA 2020 and PRISMA-S, with particular attention to transparency in study identifica-tion, screening, eligibility assessment, and methodological appraisal procedures.

The databases retained are Scopus and Web of Science only, due to their multidisciplinary coverage and the quality of their metadata. The time window spans from 1 January 2018 to 1 October 2025 in order to encompass the post-Transformer era and the rise of LLMs/foundation models, while integrating recent consolidation publi-cations. Only peer-reviewed journal articles and literature reviews in English are retained. In addition, the search follows PRISMA-S, with explicit reporting of search equations, fields, filters, and timestamps. The queries are constructed around two conceptual cores: generative artificial intelligence and management/strategy. Accordingly, Boolean syntax was adapted to each database and pilot-tested; the database-specific strings and filters are documented in the table below (see TABLE 1).

The review includes exclusively works published in English in peer-reviewed journals, encompassing both empirical research articles and literature reviews, and explic-itly addressing generative artificial intelligence, foundation models, or large language models (LLMs) applied to orga-nizational and strategic issues such as business functions and sectors, governance arrangements, business models, and organizational capabilities. Studies with a strictly technical focus and no organizational scope are excluded, as is grey literature including preprints, reports, non-peer-reviewed pro-ceedings, editorials, and letters, as well as duplicate records and all documents falling outside the predefined temporal or linguistic boundaries.

These choices follow systematic literature review recommendations in management, and are aligned with the objective of ensuring external valid-ity and comparability of the results. This choice strengthens comparability and methodological robustness, but it may under-represent the earliest evidence that first appears in preprints or conference proceedings, which we acknowledge as a limitation.

Data extraction is performed in duplicate by two inde-pendent reviewers, with subsequent consolidation. In addition, when reported, we coded enabling condi-tions such as experimentation/R&D, skills upgrading, and organizational learning routines. Empirical studies, whether quantitative, qualitative, or mixed-methods, were appraised using the Mixed Methods Appraisal Tool (MMAT). This appraisal was used to identify study-level methodological limitations and potential sources of bias, and to support a cautious interpretation of the synthesized evidence. Conceptual studies were evaluated according to the quality of their theoretical contribution in terms of clarity, novelty, scope, and usefulness, in line with and. Two reviewers carried out this appraisal independently, with discrepancies resolved by consensus.

Rather than relying on appraisal as a purely exclusionary mechanism, we used it to distinguish more robust evidence from findings requiring greater inter-pretive caution and to inform sensitivity checks in accordance with the recommendations of.

Due to the heterogeneity of indicators and contexts, no meta-analysis is conducted; instead, we employ a the-matic synthesis in the sense of and a SWiM synthesis for non-commensurable quantitative findings, following the BMJ and Cochrane Handbook guidelines proposed by and. The findings are organized into matrices linking GAI capabilities, mechanisms, outcomes, moderators, and risks, with the aim of informing a comprehensive conceptual framework. Given the heterogeneity of study designs, indi-cators, and empirical contexts, we did not apply a formal outcome-by-outcome certainty grading procedure. Instead, confidence in the body of evidence was assessed narratively, taking into account methodological quality, consistency of findings, and contextual diversity across the reviewed studies.

To ground the conceptual framework in the evidence base, we iteratively consolidated the extracted capability– mechanism–outcome patterns across studies into a parsimo-nious architecture. Hypotheses were formulated only when convergent relationships recurred across functions, sectors, and study designs. In line with the PRISMA flow diagram (see FIGURE 1), the inventory below (see TABLE 2) pro-vides a standardized tabular overview of the peer-reviewed studies retained after the screening and eligibility stages. Each entry reports the corresponding reference and the main variables investigated, thereby improving the transparency and traceability of the evidence base underlying the review.

Building on this study-level inventory, TABLE 3 synthe-sizes the main hypothesized relationships derived from the reviewed literature and translates the accumulated evidence into a structured conceptual model.

III. LITERATURE REVIEW.

A. FOUNDATIONS AND ASSOCIATED CAPABILITIES OF GAI

Generative artificial intelligence (GAI) refers to the set of computational techniques capable of producing seemingly novel content from learned distributions. Whether dealing with text, images, audio or code, it represents an extension of artificial intelligence from ‘‘recognition’’ to ‘‘generation’’. This shift made possible by the Transformer architecture and, on the language side, by the scaling up of large language models, constitutes a qualitative leap in the ability to solve general-purpose tasks in a few-shot setting and to transfer these capabilities to diverse organizational uses.

At the conceptual level, GAI relies on foundation models trained on large, heterogeneous corpora and subsequently specialized through prompting or fine-tuning, which explains their functional versatility and their rapid percolation into business processes. This versatility and its implications for information systems and management are well documented in the information systems (IS) literature. From a typological perspective, three main families dominate: large language models, which enable the generation of text, code, summaries and dialogues; latent diffusion models dedicated to images, used for synthesis, editing and upsampling; and multimodal models, which build bridges between text, image, audio and video.

Several disciplinary anchors structure the managerial dis-cussion. In innovation studies, GAI is described as a catalyst for ideation, rapid experimentation and business model inno-vation, with effects on novelty and time-to-market. In mar-keting, GAI is positioned as a driver of personalization, content generation and insight production, delivering effi-ciency gains. In the field of digital platforms, GAI is analyzed through value-creation mechanisms such as complementar-ity, orchestration and the democratization of creation, which contribute to reshaping the roles of ecosystem actors. Building on this synthesis, the present review retains four capability dimensions: personalization, deci-sion augmentation, automation, and ecosystem orchestration.

These dimensions are also aligned with prior conceptual work, including, which framed generative artificial intelligence as a strategic lever for value creation and compet-itive advantage. On this basis, and as illustrated in the figure below (see FIGURE 2), we retain four operational capability dimensions that are particularly relevant for strategy and organization:

Thus, the expanding boundaries of GAI come with height-ened requirements for transparency and explainability, now brought together under the GenXAI paradigm, which aims to strengthen trust, regulatory compliance and organizational accountability. The goal is to make the generative process intelligible by specifying why and how a given output has been produced, on the basis of which data, according to which models, and with what margins of error or bias. Recent literature emphasizes that only truly actionable explanations, combining verifiability, interactivity, cost control and secu-rity guarantees, make it possible to translate these technical capabilities into genuine managerial value.

B. ADOPTION AND INTENSITY/QUALITY OF USE IN ORGANIZATIONS

In the Information Systems literature, adoption refers to acceptance, intention to use, and subsequent actual use, which are classically explained by perceived benefits and costs, social influence, and facilitating conditions. Inten-sity and quality of use capture, beyond a simple binary ‘‘yes/no’’, the frequency, diversity and depth of interac-tions with the tool, as well as their alignment with task requirements, in line with the task–technology fit (TTF) per-spective. These two theoretical anchors remain the core foundation for analyzing the organizational appropriation of GAI and are complemented by work that reconceptual-izes the construct of ‘‘use’’ and emphasizes the need for context-sensitive measurement, particularly through the selection of usage indicators that are explicitly aligned with the expected outcomes.

To characterize the quality of use, the literature increas-ingly distinguishes between contexts in which GAI augments human capabilities and those in which it can degrade decision-making if poorly deployed. A recent meta-analysis on human–AI systems shows that the human–AI combi-nation tends to outperform especially in content creation, whereas in decision-making it may perform worse than the best of the two taken separately. This study high-lights the importance of aligning the nature of the task with the model being used, as well as the need for usage protocols that are clearly framed by supervision and verifica-tion mechanisms. Moreover, experimental work on creativity indicates that GAI tends to increase individual creativity while reducing the collective diversity of ideas.

This find-ing calls for explicit organizational choices regarding how to articulate individually GAI-assisted work with group-based creative dynamics.Governance research empha-sizes that trust, transparency, explainability, data quality, and the management of risks related to hallucinations, bias and security shape the acceptability and sustainable use of GAI.

At the strategic level, recent models describe adoption dynamics characterized by network externalities and investment trade-offs, while empirical studies link adoption to exploratory and exploitative innovation and to performance, with these relationships moderated by environmental dynamism and ethical dilemmas. Taken together, these con-tributions justify the inclusion in our model of variables capturing intention to adopt, RAI governance, and contextual contingencies.

Beyond average adoption levels, two contextual variables recur in the reviewed studies. First, several contributions show that trust informed by an understanding of model limi-tations, verification procedures and governance safeguards, what we term ‘‘informed trust in GAI’’, supports more appropriate patterns of use than blind reliance or gener-alized distrust. Second, the complexity and ambiguity of tasks condition the performance of human–AI systems: mod-els tend to perform better on structured, well-benchmarked tasks, whereas highly complex and open-ended decisions are more fragile. These observations lead us to conceptualize informed trust in GAI and task complexity as moderators of the relationship between GAI capabilities and the adoption and effective use of GAI.

Accordingly, we model adoption as a multi-stage construct (intention, intensity, and quality of use) and position informed trust and task complexity as boundary conditions shaping the translation of GAI capabilities into effective use.

C. ENABLING ORGANIZATIONAL CAPABILITIES

From a strategic perspective, GAI does not create value in and of itself; it activates organizational capabilities that enable the firm to sense opportunities and risks, seize investment and usage choices, and reconfigure assets, processes, and routines. These three micro-foundations of the dynamic capabilities view (DCV) explain why organizations exposed to the same technology do not achieve the same lev-els of performance: value depends on the speed of detection, the quality of resource allocation, and organizational plastic-ity. In turbulent informational and regulatory environments, the ongoing renewal of dynamic capabilities becomes the core mechanism linking the adoption of GAI to sustainable competitive advantages.

Applied to digital transformations, the literature shows that firms build processual dynamic capabilities through dig-ital monitoring and exploration, decision-making capabilities through portfolios of experiments and make–buy–ally trade-offs, and organizational capabilities through the reallocation of roles and the modularization of flows, which condition the industrialization of GAI use cases in terms of scaling, relia-bility, and the governance of models and data. This stream of research highlights a cumulative sequence: micro-activities such as scanning, experimenting, iterating, and standardiz-ing feed reconfiguration capabilities that, in turn, make the organization better able to capture the value of technical foundations such as large language models, diffusion models, and multimodal architectures.

In addition, absorptive capacity (ACAP) accounts for the conversion of external knowledge such as prompts, correc-tions, best practices and usage data into actionable learning: it first involves acquiring and assimilating this knowledge as a potential stock, then transforming and exploiting it as realized learning. ACAP depends on the historical pattern of investments in R&D and IT and on collective learning routines; it therefore mediates the impact of GAI on inno-vation, decision making and productivity by accelerating the internal diffusion of the knowledge generated. Importantly, this implies a conditional conversion logic: investments in AI (data, compute, models, and imple-mentation efforts) generate innovation outcomes only when accompanied by sustained innovation efforts such as experi-mentation and R&D, skills upgrading, and institutionalized learning routines.

These complements raise the organiza-tion’s ability to internalize external knowledge embedded in vendors, platforms, and ecosystems (e.g., prompts, best practices, fine-tuning signals), and to translate it into real-ized innovation and performance improvements rather than isolated technological deployment.

This logic is particularly relevant for internationally exposed firms, where the capacity to absorb external knowl-edge also conditions the extent to which innovation efforts generate beneficial spillovers from global partnerships, cross-border ecosystems, and foreign direct investment. In such contexts, absorptive capacity helps transform external tech-nological exposure into realized innovation rather than passive dependence on imported solutions.

Beyond its role as a determinant of absorptive capac-ity, R&D contributes to competitive advantage by building firm-specific knowledge stocks, experimentation routines, and innovation trajectories that are difficult to replicate. In practice, R&D strengthens a firm’s ability to translate generative AI from generic tooling into differentiated applica-tions through domain adaptation, process-specific redesign, and the codification of proprietary know-how into prompts, workflows, and control protocols. It also supports appro-priation mechanisms by coupling GAI-enabled innovation with complementary assets such as proprietary data pipelines, specialized talent, and protected process designs, thereby lim-iting pure imitation effects.

Accordingly, R&D operates as a strategic complement that amplifies the ACAP/DCV pathway and increases the likelihood that GAI usage yields durable competitive advantage rather than short-lived productivity gains. In this sense, sustained R&D and experimentation also reinforce the firm’s character as a learning organiza-tion, as repeated knowledge codification, feedback integra-tion, and capability renewal support longer-term strategic adaptation.

Recent work on GAI confirms this coupling between DCV and ACAP: value stems from orchestration through API integrations, ecosystems and platforms, from standardization through use case maps, quality controls and traceability, and from organizational learning through feedback loops and continuous improvement, far more than from the mere deployment of a model. Model governance covering version-ing, security and compliance and data governance covering quality, lineage and access are co specified here with dynamic capabilities; together they make ecosystem orchestration pos-sible and enhance the convertibility of GAI outputs into business outcomes. Accordingly, innovation effort functions as a key enabling condition that strengthens ACAP and amplifies the downstream effects of GAI capabilities and use intensity on innovation-related outcomes.

Then, we assume that dynamic capabilities and absorp-tive capacity mediate the effect of GAI capabilities and of usage intensity and quality on outcomes such as operational performance, decision quality, customer experience and com-petitive advantage, partly through GAI-related innovation as an intermediate mechanism. These mediating effects are stronger when the organization has established clear micro-foundations of scanning, experimentation and recon-figuration, together with ACAP routines that begin with acquisition and assimilation and continue with transforma-tion and exploitation.

Finally,,, connect the technology to plat-form orchestration and to the governance of models and data. Although conceptually related, the mediators are ana-lytically distinct. Absorptive capacity captures knowledge processes of acquisition, assimilation, transformation, and exploitation. Dynamic capabilities refer to higher-order sens-ing, seizing, and orchestrating strategic change. Process and organizational reconfiguration reflects realized redesign of workflows, roles, routines, and control arrangements through which GAI use is stabilized and scaled. GAI-related innovation denotes novelty outcomes at the product, ser-vice, or business-model level. Future empirical work can disentangle these effects by testing alternative mediation structures (parallel vs. serial), using temporal separation, and combining organizational- and process-level measures.

To clarify these interactions, the following figure (see FIGURE 3), presents our proposed conceptual framework. Building on these contributions, we propose a conceptual model grounded in the existing literature.

This framework positions dynamic capabilities and absorp-tive capacity as central mediating mechanisms between GAI capabilities, usage intensity and organizational outcomes. Accordingly, our model retains DCV and ACAP as media-tors and positions GAI-related innovation as an intermediate mechanism, while complementary process and organizational mechanisms are developed before.

D. RESPONSIBLE AI GOVERNANCE (RAI) AND DATA REGULATION

Recent contributions to the field of responsible artificial intel-ligence converge in defining governance as a coordinated set of structural, relational and procedural practices designed to ensure the safety, ethics and accountability of artificial intelligence systems within the organization. This governance is reflected in the establishment of dedicated policies, clearly identified roles and committees, monitoring mechanisms, training programs and processes for auditing and account-ability. Such an approach goes beyond mere legal compliance and aligns strategy, processes and risk management in order to create the conditions for a sustainable use of generative artificial intelligence.

In the same perspective, formally propose this struc-tural, relational and procedural triptych as an integrative framework for responsible artificial intelligence governance.

They examine its main antecedents, including data and IT maturity, organizational culture and leadership, as well as its effects on trust, adoption and performance.

Applied to generative artificial intelligence, governance must address a set of specific risks that include hallucina-tions, the circumvention of safeguards known as jailbreaking, the opacity of training pipelines, the leakage of sensitive information, uncertainties related to intellectual property and the difficulty of maintaining effective human oversight at scale. In the public policy literature, these risks also call for instruments that combine technical standards, processes of standardization, requirements for transparency and forms of supervision that are proportionate to the level of risk, as highlighted by.

From a regulatory perspective, the European AI Act establishes a risk-based approach and relies extensively on standardization to operationalize requirements related to safety, data governance, documentation, human oversight and system robustness, as highlighted by and. In the same context, the ISO/IEC 42001:2023 standard introduces the first dedicated artificial intelligence management system known as AIMS, which defines a structured set of policies, processes and controls intended to establish, maintain and improve artificial intelligence governance within organizations. This standard provides an operational bridge between regulatory requirements and internal practices, and its adoption is discussed as a lever for compliance and organizational performance, as noted by and.

Two transversal pillars therefore structure the governance of responsible artificial intelligence in generative artificial intelligence contexts, namely transparency and explainabil-ity on the one hand and accountability on the other. The first requires actionable explanations supported by source traceability, the articulation of underlying rationales and the clarification of usage limitations, all adapted to the generative and multimodal nature of the outputs. The second demands a clear definition of roles together with explicit duties of explanation and justification throughout the entire life cycle of artificial intelligence systems, especially as these systems gain autonomy and influence, as noted by and.

Data governance represents the other side of the equation, since data quality, lineage, protection, security and the rights attached to training datasets determine the fairness, robust-ness and lawfulness of generative artificial intelligence uses. Reviews on governance emphasize concrete mechanisms such as data catalogues, access controls, minimization poli-cies, metadata management and impact assessments, which function as trust-building infrastructures for generative artificial intelligence. Recent studies further underline that governance must explicitly encompass both data and systems and must coordinate privacy policy with artificial intelligence policy, as highlighted by and.

Responsible artificial intelligence governance therefore emerges as a decisive moderator of the effects of gen-erative artificial intelligence. The clearer the roles, stan-dards, transparency requirements and human oversight mechanisms are, the more productivity, quality and customer experience gains tend to materialize, while biases, errors, non-compliance and forms of organizational dependency remain contained. The sustainable scaling of generative artificial intelligence use cases requires a coherent alignment between regulatory expectations, standards such as the AI Act and ISO IEC 42001, explainability provisions and the effective accountability of the relevant actors.

Within this perspective, accountability becomes a genuine organizational design task that involves the definition of processes, met-rics, escalation thresholds and playbooks for red teaming and auditing, rather than a simple documentary exercise, as highlighted by,, and. This body of work supports treating RAI and data governance not only as a direct enabler of adoption and safe use, but also as a moderator that conditions the magnitude and reliability of performance effects.

E. PROCESS AND ORGANIZATIONAL RECONFIGURATION AND HUMAN-AI COLLABORATION

The diffusion of GAI within organizations produces lasting gains only when it is accompanied by a reconfiguration of processes, understood as a redesign of workflows, a redistri-bution of roles and a standardization of human control points. The literature in operations and supply chain management shows that the introduction of GAI is embedded in a capabil-ity framework that makes it possible to industrialize use cases and to secure quality in production. Without these artefacts, the effects remain local and difficult to scale up.

Importantly, this capability infrastructure is often mate-rialized through organizational innovation. Organizational innovation refers to the introduction of new managerial practices, routines, coordination mechanisms, structures, and control systems that reshape how work is organized and governed. In GAI contexts, organizational innovation enables the stabilization and scaling of use cases by formalizing roles (human oversight), redesigning workflows, institu-tionalizing verification checkpoints, and embedding GAI into performance management and accountability arrange-ments. Organizational innovation converts experimentation into durable organizational learning and performance effects.

Beyond efficiency and scaling effects, prior research also shows that organizational innovation can foster green inno-vation by reshaping coordination routines, control systems, and resource-allocation logics around sustainability-oriented objectives. This broader perspective suggests that the orga-nizational arrangements supporting GAI may also enable environmentally oriented innovation pathways when governance and performance criteria explicitly incorporate such goals.

At the micro-organizational level, randomized controlled trials and field studies converge to show that generative tools can substantially increase productivity and quality for routine and standardizable tasks, particularly professional writing and customer support report a reduction of around 40 per cent in the time required and an improvement of nearly 18 per cent in quality, while observe an average efficiency gain of about 15 per cent in a deployment involving 5 172 support agents. These effects, which are more pronounced among less experienced workers, suggest mechanisms of accelerated learning and diffusion of best practices through the tool.

As established before, the benefits of human–AI collab-oration depend on task–model fit and on supervision, veri-fication, and escalation protocols, particularly for decision-making tasks. Process and organizational reconfiguration therefore matter because they formalize roles, checkpoints, and accountability arrangements that stabilize and scale these usage conditions. At the socio-technical level, the adoption of GAI triggers measurable job crafting. Professionals reconfig-ure their tasks through new sequencing, delegation to AI and redesigned control points, and they redefine their role bound-aries. Field observations among early-career professionals reveal recurring patterns of cognitive crafting and signal craft-ing, which are mobilized to make the responsible use of GAI explicit and to assert the human added value.

These dynamics call for a rethinking of training arrangements, evaluation criteria and the management of career trajectories.

Lastly, reconfiguration must fully integrate the human factor, taking into account workloads, boundaries between professional and private life, and psychological safety. In the same vein, other studies point to a double effect, with on the one hand an increase in productivity and a possible reduction in technostress through task optimization, and on the other a heightened risk of overload when control mechanisms and escalation thresholds, particularly for high impact sensitive decisions, are not clearly defined. This underlines the need to specify human stopping points, to document the underlying rationales, and to calibrate the level of autonomy granted to generative agents according to criticality.

F. INNOVATION AND KNOWLEDGE MANAGEMENT

From an innovation management perspective, recent liter-ature identifies the domains in which GAI reconfigures ideation, development and commercialization processes, and proposes a research agenda. This work shows that it acts simultaneously on creativity, rapid experimentation and the codification of learning, thereby reducing frictions in scal-ing up. Yet these innovation benefits remain contingent on organizational complements, experimentation capability, skills development, and institutionalized learn-ing routines, that strengthen absorptive capacity. Without these efforts, GAI may accelerate local ideation speed but yields weaker cumulative innovation at the organizational level.

Regarding business models,, indicate that AI drives business model innovation based on data and con-tent generation. This dynamic results in a recompositing of the value proposition around forms of assisted co-creation, the emergence of new value capture mechanisms in which products are conceived as ‘‘content,’’ and the orchestra-tion of platforms that combine APIs, plug-ins and AI asset marketplaces. GAI is thus associated with roles of plat-form, creation tooling, intermediation and governance that are likely to open ‘‘product-to-platform’’ trajectories.

Recent empirical studies paint a mixed picture for inno-vation. On the one hand, access to GAI boosts individual creativity and the perceived quality of outputs, especially for less experienced contributors. On the other hand, the shared use of similar AI-generated suggestions tends to narrow the range of ideas at the group level, which creates a challenge in keeping enough variety in ideation processes. These patterns reinforce the need for task–model alignment and explicit verification routines, as detailed before.

In marketing, systematic comparisons indicate that AI-generated visual content can achieve, or even exceed, superhuman-level performance on criteria such as appeal, purchase intention and engagement metrics. This potential paves the way for shorter design cycles, intensified A/B testing and creative personalization at low marginal cost, while reconfiguring the trade-offs between human design and assisted generation rather than removing the requirements of brand control, consistency and safety.

At the level of knowledge management, GAI restructures the four classical processes of creation, storage, transfer and application by facilitating the externalization of tacit knowl-edge into textual form, the synthesis of heterogeneous infor-mation and internal dissemination through prompt libraries, playbooks and usage guidelines, as shown by. From this perspective,, emphasize the need to co-specify generative tools and knowledge practices through policies for explication, traceability and curation, clear editorial gov-ernance, safeguards for reuse and monitoring mechanisms, in order to build augmented knowledge-management ecosys-tems in which GAI, combined with human roles of curation, verification and arbitration, reinforces organizational learn-ing and sustainable innovation.

Such arrangements are also consistent with the learning-organization perspective, insofar as they transform dispersed AI-assisted practices into collec-tive, cumulative, and reusable organizational knowledge.

G. RISKS, LIMITATIONS AND SUSTAINABILITY CONDITIONS

Cognitive uncertainty associated with generative models remains a major risk, as hallucinations, understood as plausi-ble but factually incorrect responses, persist even in advanced technical configurations and expose managerial decisions to errors that are difficult to detect in the absence of system-atic external verification protocols. In the same vein, show that detection and mitigation are progressing but remain incomplete, particularly in out-of-distribution settings and in contexts that require a high level of precision.

Algorithmic bias and automation effects represent a second major limitation. Inside organizations, generative systems can reproduce or even amplify existing data biases and shape human judgement, including that of experienced decision-makers. In the same line, confirm these effects and recommend safeguards based on explainability and criti-cal training. Security vulnerabilities are also becoming more serious as models are increasingly exposed to external con-tent. Prompt injection attacks can hijack agents, exfiltrate data or bypass internal policies. Recent taxonomies proposed by and describe transferable attack scenarios and argue for multilayered defence strategies.

From an ethical and regulatory point of view, the European AI Act (Regulation (EU) 2024/1689) requires transparency, risk management, data quality and human oversight for high-risk systems, in line with the standardization effort around AI management systems (ISO/IEC 42001). For firms, this translates into obli-gations of traceability, incident logging and regular audits of robustness and bias.

An often-underestimated organizational risk is techno-logical dependence and the atrophy of skills, as cognitive offloading and automation can reduce critical acuity and the mastery of complex tasks over time, particularly in the absence of ‘‘AI-in-the-loop’’ strategies focused on deliber-ate practice. References,, and report a measurable decline in reasoning capabilities or collective creativity when assistance becomes systematic. In terms of competitiveness, the advantage generated by GAI is rapidly diluted through imitation, tooling standards and sectoral diffusion, and performance gains persist only when firms combine AI with organizational complements such as pro-prietary data, processes, skills and relational assets, together with governance that supports high-quality decision mak-ing.

Consistent before, sustained innovation effort, including R&D, supports continuous learning and safer scal-ing, which may help preserve competitive advantage despite rapid diffusion of similar foundation models.

Finally, environmental sustainability is becoming a strate-gic determinant as the energy and water footprint of training, and especially of large-scale inference, is rapidly increasing. Recent estimates and commentaries underline the importance of inference frugality and low-carbon supply. The robust and lasting integration of GAI rests on a coherent set of technical, organizational and environmental condi-tions. It requires a reliable architecture, governance structured by an AIMS-type management system with risk registers, regular tests and audits, as well as secure decision-making arrangements that combine human validation, operational explainability and the monitoring of uncertainty.

These requirements must be supported by the development of critical skills in the face of automation and by careful management of the energy footprint through carbon budgets, the measurement and optimization of consumption and the use of decarbonized infrastructures. In fact, the risks of GAI are not reasons for abstention but design and governance constraints, and sustainable value emerges when the tool is embedded in distinctive data, learning routines and risk management aligned with regulatory and environmental requirements. Beyond risk containment, the sustainability stream also connects GAI to innovation path-ways that can support environmentally oriented outcomes

(e.g., efficiency, waste reduction, circular reporting and mon-itoring), when such outcomes are explicitly examined in the reviewed studies. However, these effects depend on organi-sational innovation and governance that align GAI use with sustainability goals, clarify accountability, and avoid rebound effects.

IV. RESULTS AND DISCUSSION.

Taken as a whole, this state of the art suggests a causal chain in which the capabilities of GAI, namely personalization, deci-sion augmentation, automation and ecosystem orchestration, drive adoption and, above all, the intensity and quality of use, provided that there is a satisfactory fit between tasks and technology. This use, however, does not automatically generate value. Its conversion into durable organizational outcomes depends on complementary innovation efforts, including R&D, skills development, experimentation, and learning routines, which reinforce absorptive capacity and dynamic capabilities.

Through these mechanisms, organizations become better able to acquire, assimilate, transform and exploit knowledge, to adapt models to proprietary pro-cesses, and to stabilize and scale GAI-enabled work through process and organizational reconfiguration, including new routines, coordination mechanisms, control systems, and broader forms of organizational innovation.

To synthesize the hypothesized relationships derived from this causal chain, TABLE 3 presents the research hypotheses associated with our conceptual model.

It is important to distinguish the role of H7 from that of H6, H8, H9, and H10 in the structure of the model. H7 captures the direct overall relationship between the adoption and use of GAI and organizational performance. By contrast, H6, H8, and H9 specify the mediator-to-performance links through which the indirect effects of adoption and use are expected to operate via process and organizational reconfiguration, absorptive capacity, and dynamic capabilities, respectively. H10 reflects a distinct indirect pathway in which GAI capa-bilities are expected to influence organizational performance through GAI-related innovation. However, the expected effect is likely to be stronger in organizations characterized by higher digital maturity, stronger data governance, and greater alignment between generative AI applications and operational processes.

Conversely, the effect may be weaker in contexts marked by fragmented data, limited managerial support, or insufficient employee readiness.

As reflected in the table above (see TABLE 3), these mediating mechanisms, together with GAI-related innova-tion at the level of products, services and business models, explain how GAI capabilities can translate into operational performance, decision quality, customer experience, and competitive advantage. At the same time, these benefits remain conditional rather than guaranteed, since their relia-bility and transferability depend on the quality of governance arrangements. In this respect, RAI governance, based on transparency, human oversight, data quality, and safety, acts as a structuring condition that both amplifies value creation and secures it against the risks associated with generative systems.

This conversion chain is particularly critical for internationally exposed firms, where competitive advantage also depends on the ability to absorb external knowledge from transnational networks, global partnerships, and innovation spillovers associated with foreign investment and cross-border ecosystems.

First, we posit that GAI capabilities positively influence the adoption and use of GAI systems. Second, we expect that the adoption and use of GAI positively influence process reconfiguration, absorptive capacity and dynamic capabilities, while GAI capabilities also positively influence GAI-related innovation. Here, reconfiguration includes not only workflow redesign and role redistribution, but also organizational innovation in the form of new routines, struc-tures, coordination arrangements and management control practices required to scale and govern GAI-enabled work. Furthermore, we hypothesize that process and organizational reconfiguration, the adoption and use of GAI, absorptive capacity, dynamic capabilities and GAI-related innovation each positively influence the four dimensions of organizational performance.

More specifically, the relationship proposed in H7 should be read as the direct performance link associated with adoption and use, whereas H6, H8, and H9 represent the downstream links that make the corresponding mediated pathways explicit. In turn, H10 captures the innovation-based route to performance, which is conceptually distinct because it originates from GAI capabilities through GAI-related inno-vation rather than from adoption and use alone.

We treat innovation effort (experimentation/R&D, capa-bility building, and learning routines) as a foundational micro-basis of absorptive capacity rather than as a stan-dalone construct, which makes explicit the investment-to-innovation conversion mechanism emphasized by prior learning research. Taken together, the hypotheses summa-rized in Table above (see TABLE 3) form a coherent impact chain that links GAI capabilities and contextual contingencies (including complementary innovation assets such as R&D and learning investments, as well as data/IT maturity and regulatory exposure) to organizational performance. In this structure, H7 expresses the direct adoption/use–performance relationship, whereas H6, H8, H9, and H10 specify the distinct downstream paths that underpin the model’s indi-rect explanatory logic.

To make these relationships easier to understand, we synthesize them in the conceptual model presented in Figure below (see FIGURE 4).

This framework visually organizes the upstream drivers, the mediating organizational mechanisms and the down-stream outcomes in terms of operational performance, customer experience, competitive advantage and decision quality. It is less likely to apply, or may yield neutral/negative outcomes, when task–technology fit is low, when data quality and security constraints prevent reliable deploy-ment, or when organizations lack complementary skills and change readiness to institutionalize human–AI workflows. In highly regulated or reputationally sensitive settings, adoption may remain superficial or generate negative outcomes through compliance risks, error propagation, bias exposure, information leakage, or accountability failures. Moreover, rapid diffusion can erode advantage unless organizations continuously renew learning, reconfiguration practices, and governance safeguards.

The consolidation of empirical findings highlights value creation that is first particularly pronounced for routine and standardizable tasks, through gains in productivity and qual-ity, and then more differentiated value when the organization manages to industrialize human–AI workflows and activate innovation. These observations lead to an explicit distinc-tion between the links in the impact chain. The quality of outcomes also depends on governance that can reduce hal-lucinations, biases and information leakage. We therefore argue that a more robust framework of RAI governance is likely to strengthen the impact of the adoption and use of GAI on process and organizational reconfiguration and to reinforce the positive effects of process and organizational reconfiguration on value creation outcomes by increasing the reliability and transferability of the gains.

Formally, we hypothesize that RAI governance positively moderates the relationship between the adoption and use of GAI and process and organizational reconfiguration, and positively moderates the relationships between process and organizational recon-figuration and each performance dimension.

Finally, the adoption and effective use of GAI for decision augmentation varies according to the nature of the task and the perception of the tool. We therefore expect informed rather than naïve trust in GAI to positively moderate the rela-tionship between GAI capabilities and the adoption and use of GAI, whereas task complexity, understood as the ambiguity and difficulty of the task, is expected to negatively moderate this relationship, with stronger negative effects when tasks are more complex and ambiguous. We conceptualize organi-zational performance as a higher-order outcome reflected by four dimensions: operational performance, customer experi-ence, competitive advantage, and decision quality.

V. CONCLUSION.

This systematic review examined, based on Scopus and Web of Science sources published between 2018 and 01 October 2025, how generative artificial intelligence is reshaping orga-nizational strategies, processes, and outcomes. Overall, the review shows that the organizational effects of GAI depend less on adoption alone than on the conditions that enable its effective and sustainable deployment. In response to this, the article develops a sequential conceptual framework and formulates fourteen hypotheses to guide future empirical validation.

From a scientific perspective, this article advances the literature by clarifying and operationalizing four core capa-bilities of generative artificial intelligence, namely personal-ization, decision augmentation, automation, and ecosystem orchestration, and by linking them to analytically distinct organizational mediation mechanisms grounded in absorptive capacity, dynamic capabilities, process and organizational reconfiguration, and GAI-related innovation. In doing so, the article does not merely aggregate dispersed findings but pro-poses a more structured and testable explanation of how GAI capabilities are converted into organizational outcomes under specific governance and use conditions.

This contribution helps reposition GAI not simply as a technological artifact, but as a strategic and organizational phenomenon whose effects depend on how use, capabilities, and governance are articulated.

The managerial implications that follow from this synthe-sis are straightforward. The adoption of generative artificial intelligence alone is insufficient; organizations need to design complete value chains in which use is framed, monitored, and supported by appropriate governance, human oversight, and organizational learning mechanisms. Senior managers should pay particular attention to the diversity of ideas in creative contexts, to decision quality in sensitive contexts, and to the sustainability of inference as usage volumes increase.

This work has limitations that open up avenues for further research. The corpus is restricted to two databases and to a recent time window, and the heterogeneity of contexts and metrics has led us to favor a narrative synthesis. Longitudinal and quasi-experimental studies conducted across multiple sectors are now needed to test hypotheses from H1 to H14, refine the measurement of use quality and task-model align-ment, assess the effects of governance on decision quality, and map the environmental costs of generative artificial intel-ligence at the level of inference.

Future research should also examine more closely how organizational and innovation-related investments shape the long-term stabilization of GAI-enabled capabilities across sectors and contexts. The proposed research agenda aims to move from a body of convergent evidence to robust causal demonstrations that more tightly articulate technolo- gies, organizations and outcomes. It should also clarify under which boundary conditions GAI generates durable benefits, limited effects, or even negative consequences.

Ultimately, generative artificial intelligence constitutes a powerful strategic accelerator when organizations are able to align its deployment with appropriate capabilities, gov-ernance, and learning processes. By putting forward an integrated conceptual framework, this article provides a cumulative basis for empirical validation and managerial decision making and helps organizations transform local experimentation into sustainable advantage.