You’re listening to “Unlocking microtourism small and medium-sized enterprises potential: Human–AI synergy and open-innovation pathways to sustainability,” by R. Thonglor, N. Chotisarn, and T. Phuthong. Published in 2026.

ISSN: 2331-1975 (Online) Journal homepage: the linked source

Unlocking microtourism small and medium-sized enterprises potential: Human–AI synergy and open-innovation pathways to sustainability

Ratchamongkhon Thonglor, Noptanit Chotisarn & Thadathibesra Phuthong

To cite this article: Ratchamongkhon Thonglor, Noptanit Chotisarn & Thadathibesra Phuthong (2026) Unlocking microtourism small and medium-sized enterprises potential: Human–AI synergy and open-innovation pathways to sustainability, Cogent Business & Management, 13:1, 2698507, DOI: 10.1080/23311975.2026.2698507

© 2026 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group

Published online: 07 Jul 2026.

Submit your article to this journal

Article views: 439

View related articles

View Crossmark data

Hospitality and Tourism | Research Article

Unlocking microtourism small and medium-sized enterprises potential: Human–AI synergy and open-innovation pathways to sustainability

Ratchamongkhon Thonglora, Noptanit Chotisarnb and Thadathibesra Phuthonga aFaculty of Management Science, Silpakorn University, Phetchaburi, Thailand; bThammasat Business School, Thammasat University, Bangkok, Thailand

ABSTRACT.

This study examines how generative artificial intelligence (GAI) and open-innovation orientation (OIO) influence sustainability and innovation performance in Thai microtourism small and medium-sized enterprises (SMEs), mediated by human–AI collaboration quality (HAIC), digital supply chain innovation, and dynamic capabilities for sustainable tourism (DCST), and moderated by customer involvement in sustainable tourism (CIST). Quantitative cross-sectional surveys were administered to 450 SMEs. Partial least squares structural equation modeling was used to test the hypothesized relationships via 5,000-subsample bootstrapping with bias-corrected confidence intervals; 12 of 14 hypotheses were supported. GAI adoption most strongly activated HAIC (β = 0.427), and OIO produced the highest path coefficient (OIO → DCST: β = 0.451).

All mediation pathways were significant (variance accounted for: 23–44%), confirming partial mediation; R2 = 0.478 (sustainability performance [SP]) and 0.427 (innovation performance). The moderation hypotheses were not supported (p > 0.30); however, CIST directly affected SP (β = 0.208, p < 0.001). This study is the first to position GAI and OIO as co-primary antecedents within a unified triple-mediation framework and validates HAIC as a first-order mediating construct. The findings extend the resource-based view and dynamic capabilities and open-innovation theories to Thai microtourism, offering actionable guidance for managers, policymakers, and technology providers.

1. Introduction.

ARTICLE HISTORY

SUBJECTS

Entrepreneurship and Small Business Management; The Business of Tourism; Tourism Planning and Policy; Tourism Development/Impacts; The Tourism Industry; Sustainability; Tourism; Management of Technology & Innovation

Tourism contributes 9.1% to the global gross domestic product (GDP) and supports one in ten jobs worldwide; however, the COVID-19 pandemic erased 74% of international travel and USD 1.3 trillion in export revenues. This situation has exposed the structural fragility of tourism-dependent economies and intensified calls for recovery models that are digitally agile and sustainability-oriented. In Thailand, more than 99% of all tourism businesses are micro-, small-, or medium-sized enterprises (MSMEs) (≤50 employees), contributing 11–12% of GDP in prepandemic years. Thailand’s post-COVID recovery strategy was operationalized through the Tourism Authority of Thailand’s digital transformation agenda, the Designated Areas for Sustainable Tourism Administration framework, and the Bio-Circular-Green Economy model.

This strategy placed MSMEs at the center of national revitalization while simultaneously compelling them to meet the sustainability requirements for the Thailand Safety & Health Administration (SHA) SHA Plus certification. Operators that typically rely on LINE OA and Facebook for customer relationships face the challenge of competing in experience-driven global markets while meeting intensifying sustainability obligations, creating a formidable strategic tension.

Two converging research streams offer partial responses to this challenge. On the technology side, generative artificial intelligence (GAI) (e.g. ChatGPT and Google Gemini) can provide accessible, low-cost capabilities for itinerary personalization, multilingual communication, content creation, and sustainability documentation. These functionalities enable microenterprises to utilize conventional digitalization pathways. Wang and Zhang (2025) conducted a partial least squares structural equation modeling (PLS-SEM) study of 429 Chinese tourism SMEs. They found that GAI adoption significantly improved digital supply chain innovation and collaboration, jointly mediating a positive effect on environmental, social, and governance (ESG) performance. Moreover, De Fano et al. (2026) examined 102 European small- and medium-sized enterprises (SMEs).

They found that AI ambidexterity was positively associated with dynamic capabilities and firm performance. Similarly, Alhomaid and Al-Romeedy (2026) found that AI adoption enhanced firm performance through the mediation of digital innovation and organizational agility in hospitality and tourism businesses. On the organizational side, an open-innovation orientation (OIO)—the purposeful use of knowledge inflows and outflows to accelerate innovation —offers a low-cost pathway to enhance capabilities through customer cocreation, technology partner collaboration, and ecosystem knowledge exchange.

Despite the growth of both research streams, prevailing scholarship has treated GAI adoption and OIO as analytically separate rather than mutually reinforcing codrivers. Existing AI-focused tourism studies exclude OIO from their frameworks, while the open-innovation (open innovation) literature has insufficiently theorized how generative AI reshapes knowledge absorption at the heart of OIO. This analytical separation has persisted partly because GAI adoption and OIO have developed in distinct disciplinary homes (technology adoption versus innovation management), making integration imperceptible and theoretically nontrivial. Three further gaps have compounded this integration problem.

First, and most critically, the quality of human–AI collaboration (i.e. the degree to which human expertise and AI outputs are effectively integrated through trust, complementarity, and iterative co-production) has not been operationalized as a theoretically grounded first-order mediating construct. This oversight persists despite Qu and Kim (2025) qualitative evidence that augmented human–AI collaboration capability is the central value-creating mechanism in AI-enabled microenterprise ecosystems. This study treats human–AI collaboration quality (HAIC) as a distinct intermediate capability rather than conflating it with technology adoption intensity, thereby advancing the primary theoretical novelty. Second, existing research is weighted toward larger enterprise contexts outside tourism.

For example, Anser et al. (2025) sampled Pakistani SMEs, De Fano et al. (2026) studied participants in European Digital Innovation Hubs, and Qu and Kim (2025) addressed Chinese apparel manufacturers. Studies positioned within microtourism SMEs with informal governance and acute resource constraints remain limited. Third, LINE OA-mediated customer relations, informal business networks, and SHA Plus certification pressures dominate the Thai and Southeast Asian context, which is substantially underrepresented in the global evidence base.

These gaps raise three empirical questions. How do GAI adoption and OIO jointly (rather than independently) shape the mechanisms that drive sustainability and innovation outcomes in microtourism enterprises? Does the quality of human–AI collaboration constitute a distinct and significant mediating mechanism that exceeds AI adoption alone? Moreover, do findings from China, Vietnam, and Europe transfer to Thailand’s distinctive institutional environment? To address these questions, prior research must be extended by developing an integrated framework that positions GAI and OIO as co-antecedents, operationalizes HAIC as a first-order mediating construct, and explores the Thai microtourism context.

This study develops and empirically tests an integrated conceptual framework, examining how GAI adoption and OIO jointly influence sustainability performance (SP) and innovation performance (IP) in Thai microtourism SMEs, with HAIC, digital supply chain innovation (DSCI), and dynamic capabilities for sustainable tourism (DCST) as mediating mechanisms, and customer involvement in sustainable tourism design (CIST) as a boundary-condition moderator. Four specific objectives guide this investigation. Examine the direct effects of GAI adoption and OIO on HAIC, DSCI, and DCST. Assess the mediating roles of HAIC, DSCI, and DCST in transmitting antecedent effects to SP and IP. Evaluate the moderating effect of CIST on mediator–performance relationships. Derive evidence-based recommendations for Thai microtourism managers, policymakers, and technology providers.

Three research questions organize this investigation. RQ1: How do GAI adoption and OIO jointly affect HAIC, DSCI, and DCST in Thai microtourism SMEs? RQ2: What roles do HAIC, DSCI, and DCST play in mediating the relationships between GAI adoption and OIO and sustainability and innovation performance; does HAIC function as a theoretically distinct first-order mediating mechanism? RQ3: How does CIST moderate mediator–performance relationships, and under what conditions does Thailand’s institutional environment strengthen or attenuate this moderation? These questions aim to address the three gaps identified above, and transition from the integration of AI and open innovation as co-antecedents to the operationalization of HAIC as a first-order mediating construct and then to its empirical embedding within the Thai microtourism context.

This study makes three contributions. First, it extends the resource-based view and dynamic capabilities theory into human–AI collaboration. We treat HAIC as a novel, difficult-to-imitate organizational capability that explains the portion of AI’s performance effect not attributable to technology adoption alone. Second, this study integrates open-innovation theory with AI adoption scholarship by repositioning OIO as a co-primary antecedent with effects comparable in magnitude to GAI adoption. This approach delivers the theoretical AI–open-innovation synthesis called for but not achieved in the tourism management literature. Third, we develop a PLS-SEM measurement framework designed for replication across other MSME-dominant, emerging-economy tourism sectors in Southeast Asia.

This study’s findings offer evidence-based guidance for three stakeholder groups. First, the results can help Thai microtourism managers navigate LINE OA-mediated operations and SHA Plus certification requirements. They can also aid policymakers in designing AI–OIO support programs under the Bio-Circular-Green (BCG) Economy Model, the Tourism Authority of Thailand’s digital transformation agenda, and the Designated Areas for Sustainable Tourism Administration (DASTA) sustainability platforms. Finally, this study’s findings can help technology providers whose product design should embed human–AI collaboration development rather than deliver raw AI outputs. Section 5 presents detailed practitioner recommendations.

Section 2 reviews the literature and develops 14 testable hypotheses across the 3 theoretical foundations and 8 constructs. Section 3 describes the positivist quantitative design, 70-item survey instrument, and 5-step PLS-SEM analytical protocol implemented in SmartPLS 4.0. Section 4 presents empirical results across measurement, structural, mediation, and moderation analyses. Section 5 discusses the findings in relation to the research questions, prior literature, and study limitations, and derives implications for managers, policymakers, and technology providers. Finally, Section 6 synthesizes contributions and proposes a future research agenda.

2. Literature review, theoretical framework, and hypotheses development.

This section reviews the theoretical and empirical literature underpinning the proposed conceptual framework across 12 substantive sections. Beginning with 3 foundational theories (Resource-Based View (RBV), dynamic capabilities theory, and open-innovation theory), the review proceeds through each core construct, synthesizes identified gaps, presents the integrated framework, and develops 14 testable hypotheses.

2.1. Theoretical foundations.

2.1.1. RBV.

RBV is attributed to Wernerfelt (1984) and developed most influentially by Barney (1991); it posits that sustained competitive advantage derives from firm-level resources that are simultaneously valuable, rare, inimitable, and non-substitutable (VRIN). Unlike industrial organization perspectives, which locate advantage in market structure and industry positioning, the RBV focuses on heterogeneous resource endowments that differentiate competing firms, although the boundaries of its explanatory usefulness continue to be debated. Scholars applying RBV to information systems argue that when information technology (IT) capabilities are embedded in tacit organizational routines, complemented by human capital, and are deployed through idiosyncratic processes, they can satisfy the VRIN criteria.

In this study, GAI adoption constitutes a potentially VRIN resource not through mere tool possession but through integration with local knowledge, service craft, and customer relationship routines—elements that competitors cannot readily replicate. Similarly, open-innovation networks represent relational resources whose rarity and inimitability stem from path dependence, social complexity, and causal ambiguity. Within this study’s framework, RBV provides the foundational logic for treating GAI adoption and OIO as the two primary VRIN-eligible antecedent resources. At the same time, HAIC—the quality of human–AI integration—constitutes the most difficult-to-imitate capability through which those resources generate sustained competitive advantage.

2.1.2. Dynamic capabilities theory.

Dynamic capabilities theory extends the RBV by theorizing the processes through which firms sense emerging opportunities, mobilize resources to seize them, and transform existing configurations to meet changing demands. The tripartite structure—sensing, seizing, and transforming—provides the organizational infrastructure for continuous adaptation, which is essential in Thailand’s volatile post-pandemic tourism environment. Anser et al. (2025) conducted a PLS-SEM study of SMEs and found that organizational and regulatory mechanisms mediated the effect of AI use on environmental performance. Moreover, De Fano et al. (2026) showed that AI ambidexterity positively affected dynamic capabilities, which in turn drove firm performance in European SMEs.

Nieves and Haller (2014) further established that dynamic capabilities in hospitality firms emerge from knowledge-based resources and managerial orchestration rather than from technology investment alone—a finding with direct relevance to the microenterprise context. This study draws on these previous frameworks to theorize HAIC, DSCI, and DCST as the three mediating dynamic capability instantiations through which GAI and OIO investments translate into sustainability and innovation outcomes. Specifically, HAIC is a sensing-and-seizing mechanism for human–AI co-production, DSCI is a reconfiguring mechanism for operational coordination, and DCST is an integrated adaptive capability enabling sustainability transitions.

2.1.3. Open-innovation theory.

Open-innovation theory challenges the closed innovation paradigm, arguing that firms that leverage external knowledge inflows, commercialize internal knowledge outflows, and engage in coupled exchange outperform those relying solely on internal research and development (R&D). Open innovation is particularly relevant for resource-constrained SMEs that lack absorptive capacity and R&D budgets for closed strategies. Inbound open-innovation draws on customer ideas, supplier expertise, and platform intelligence and offers a low-cost pathway for capability enhancement. The interaction between open innovation and AI adoption has emerged as a significant frontier. Open-innovation theory suggests that a stronger open-innovation orientation amplifies AI’s capability-building effect by widening the external knowledge base available for sensing and reconfiguration.

Similarly, Madhavan et al. (2022) documented a paradigm shift toward integrated frameworks, positioning OIO as an enabling condition for digital transformation. These findings have established the theoretical basis for treating GAI and OIO as co-antecedents rather than independent streams. Within the present framework, OIO specifically activates all three mediating capability pathways by expanding the knowledge base available for human–AI collaboration (→ HAIC), enabling supply chain partner integration (→ DSCI), and providing the external sensing inputs required for sustainability adaptation (→ DCST).

2.2. Generative AI adoption in tourism.

GAI refers to large language model-based systems (such as OpenAI’s ChatGPT, Google Gemini, and Anthropic’s Claude) that can produce novel text, images, and structured outputs in response to natural language prompts. Unlike earlier rule-based chatbots or recommendation systems, GAI is characterized by contextual generativity. It can synthesize heterogeneous inputs, produce culturally nuanced outputs, and adapt to user feedback in ways approximating human communicative competence. Tourism applications include personalized itinerary generation, multilingual customer communication, marketing content creation, online review synthesis, revenue management, and sustainability documentation.

Wang and Zhang (2025) provided the most directly relevant empirical foundation. They conducted a PLS-SEM study of 429 Chinese tourism SME managers, demonstrating that GAI adoption significantly influenced DSCI and DSCC, which jointly mediated GAI’s effect on ESG performance. For microenterprises in emerging economies, contemporary GAI platforms provide subscription-based access within microoperators’ financial reach, eliminating the need for dedicated AI development; however, digital literacy gaps, challenges with legacy system integration, and uncertainty about output reliability remain genuine barriers that moderate the depth of adoption among microenterprises.

In Thailand specifically, the high penetration of mobile-first platforms, such as LINE and Facebook, creates an adoption pathway in which GAI is often accessed through familiar consumer platform integrations rather than standalone enterprise applications. This situation reduces friction while potentially constraining the depth of capability utilization.

2.3. Open-innovation orientation (OIO) in SME contexts.

As a firm-level strategic construct, OIO captures the degree to which an organization actively pursues, absorbs, and deploys external knowledge while contributing to ecosystem-level knowledge flows. OIO encompasses both an attitudinal dimension (managerial belief in collaborative value creation) and a behavioral dimension (actual investment in partnerships, customer cocreation mechanisms, and platform participation). The theoretical case for OIO in SMEs is strong. Resource constraints limit the use of purely internal innovation strategies, while proximity advantages—local market embeddedness and organizational flexibility—provide natural foundations for cocreation. Empirically, however, SMEs face absorptive capacity limitations that constrain their ability to identify and exploit externally sourced knowledge.

Evidence from SME contexts highlights critical success factors. For example, the absorptive-capacity perspective holds that open-innovation success depends on the alignment between external knowledge-seeking and internal absorptive capacity—firms that invest in open innovation without corresponding knowledge management capabilities systematically underperform. Similarly, Hjalager (2010) found that tourism firms participating in knowledge-sharing networks and destination management cocreation achieve superior innovation outcomes, with boundary-spanning OIO practices explaining a significant share of the variance in innovation performance beyond internal R&D investment alone. Open-innovation theory holds that OIO functions as an enabling condition for AI value creation, broadening the external knowledge inflows that strengthen AI-supported capabilities.

This logic also extends to Thai microtourism operators. In the tourism sector, OIO manifests as co-designing products with local communities, participation in digital platform ecosystems (e.g. Airbnb, Booking.com, and Agoda), and co-development of sustainability initiatives with non-government organizations (NGOs) and destination management organizations. Thailand’s SHA Plus certification program explicitly incentivizes collaborative sustainability practices, creating a policy-embedded OIO context that differs from other emerging-economy tourism environments.

2.4. Human–AI collaboration (HAIC) quality.

HAIC quality represents the degree to which human operators and AI system outputs are effectively integrated through trust, complementarity, iterative learning, and shared decision-making within organizational workflows. The construct draws on augmentation theory—the proposition that AI is most valuable when it enhances rather than replaces human cognitive capabilities. HAIC distinguishes itself from simpler adoption or usage-intensity measures by emphasizing the interactive quality of the human–AI interface. Theoretical foundations span various disciplines. For example, the construct inherits trust, usability, and perceived competence dimensions from human–computer interaction. Moreover, HAIC borrows the concept of complementarity from organizational behavior, in which human and AI capabilities reinforce each other to produce outcomes neither could achieve independently.

Finally, it integrates co-production logic from service management, in which tourism service value is a joint product of the operator’s local knowledge and AI generative capability.

Qu and Kim (2025) provided the most relevant empirical treatment among Chinese microenterprises. They identified augmented HAIC capability (AHAICC) as a central mediating mechanism linking AI ecosystem participation to sustainable innovation outcomes, thereby establishing HAIC as an intermediate capability rather than a downstream outcome. Wang and Zhang (2025) corroborated this positioning, showing that collaboration mechanisms could explain most of GAI’s ESG performance effect. HAIC in the Thai microtourism context is closely tied to individual owner-operators’ learning trajectories, AI output trust calibration, and peer-network-mediated technology diffusion. For example, LINE OA integration provides a familiar pathway to HAIC adoption by reducing the cognitive distance between existing practice and AI-augmented workflows.

2.5. Digital supply chain innovation (DSCI).

DSCI refers to the transformation of service delivery coordination, partner integration, and operational processes through digital technologies. It encompasses cloud-based booking and inventory systems, real-time data-sharing platforms, AI-enabled demand forecasting, and digital communication tools spanning the full service chain from input procurement to guest experience delivery. Tourism supply chains are inherently multi-actor and experience-centric. The quality of a tourist’s experience reflects coordinated interactions among accommodation providers, transport operators, attraction managers, food and beverage operators, and guiding services. Each interaction requires alignment on shared quality standards and real-time operational information.

Wang and Zhang (2025) provided the primary empirical anchor, demonstrating that DSCI significantly mediated the relationship between GAI adoption and ESG performance in Chinese tourism SMEs. Their importance–performance map analysis (IPMA) confirmed that DSCI is among the highest-importance, highest-improvable constructs in the structural model. Customer involvement positively moderated the DSCI–ESG path, reinforcing supply chain digitalization as a demand-sensitive mechanism. For Thai microtourism SMEs, DSCI takes distinctive forms conditioned by the digital platform ecosystem that structures supply chain relationships. For example, LINE Groups coordinates transport and accommodation partners; global booking platforms (e.g. Airbnb, Booking.com, and Agoda) provide algorithmic demand forecasting, effectively outsourcing supply chain intelligence to the platforms.

In this environment, DSCI for microenterprises means integrating AI-assisted content creation, translation, and communication within the existing platform ecosystem. This approach reduces information asymmetries and systematizes sustainability documentation for SHA Plus compliance.

2.6. Dynamic capabilities for sustainable tourism (DCST).

DCST applies Teece et al. (1997) tripartite framework to the adaptive challenges of sustainability transitions in tourism. Sensing capabilities involve identifying shifting tourist preferences regarding responsible travel, tracking evolving environmental regulations (such as SHA Plus and Green Leaf certification), and monitoring sustainability-advancing competitor moves. Seizing capabilities involves mobilizing resources in response, such as investing in eco-friendly infrastructure, developing community-based offerings, or acquiring sustainability certifications. Transforming capabilities requires prerequisite organizational reconfiguration to institutionalize sustainability as a core operational principle.

Examples of such integration include incorporating environmental metrics into daily operations, redesigning service delivery to minimize ecological impact, and repositioning the value proposition around sustainable tourism experiences.

DCST has a robust empirical grounding as a mediating mechanism. For example, Alhomaid and Al-Romeedy (2026) demonstrated that capability-based mechanisms (digital innovation and organizational agility) mediated the effect of AI adoption on firm performance. This positioned dynamic capabilities as a primary rather than a supplementary mechanism. De Fano et al. (2026) corroborated this finding in a European digital innovation context. They determined that dynamic capabilities mediated the effect of AI ambidexterity on firm performance. In the hospitality sector, Nieves and Haller (2014) demonstrated that dynamic capabilities in hotel firms are built through managerial knowledge and interorganizational collaboration rather than solely through technology. They established that capability-building in tourism requires human orchestration of resources.

This finding directly informs the DCST-as-mediator logic of our study. OIO also directly enriches DCST by drawing on external knowledge from customers, NGOs, certification bodies, and ecosystem partners, providing the sensing inputs required for sustainability adaptation. For instance, Mousa and Bouraoui (2025) provided longitudinal support for sequenced dynamic capability activation. For Thai microtourism SMEs, DCST is imperative (the BCG Economy Model and DASTA programs make sustainability capability an operating prerequisite); however, it is also constrained by limited managerial bandwidth and financial reserves. This duality positions DCST as a critical explanatory variable for performance heterogeneity.

2.7. Customer involvement in sustainable tourism design (CIST).

CIST refers to the extent to which customers actively co-create sustainable tourism experiences by sharing ideas, co-designing experience elements, and providing real-time feedback through digital platforms. Theoretical foundations draw on service-dominant logic, which repositions the customer as an active value co-creator rather than a passive service recipient. CIST also draws on the sustainable value cocreation literature, which extends this logic to encompass the environmental and social dimensions of experience design.

Wang and Zhang (2025) provided the most direct empirical evidence for CIST as a moderating boundary condition. They found that customer involvement significantly strengthened the positive effects of both DSCI and DSCC on ESG performance in Chinese tourism SMEs (significant positive interaction effects). Their results confirm that ESG returns to supply chain innovation are amplified by active customer co-design. This moderating role is theoretically interpretable—customer involvement introduces external sustainability knowledge, legitimizes sustainability investments through demand signaling, and creates accountability mechanisms that reinforce operators’ commitments.

In the Thai digital ecosystem, CIST is primarily mediated through LINE OA messaging communities, Facebook groups, and Airbnb host-guest communication threads (the digital infrastructure through which customer sustainability preferences are communicated and incorporated into service offerings). GAI can systematically analyze customer feedback at scale and accelerate the translation of CIST signals into actionable service design intelligence; however, this technological amplification role falls outside the current model’s boundary conditions. This limitation is identified as a direction for future research.

2.8. Sustainability and innovation performance.

Sustainability performance is conceptualized through the triple bottom line (TBL) framework, encompassing economic sustainability (revenue stability, cost efficiency, and financial viability), environmental sustainability (reducing ecological footprint, resource conservation, and waste minimization), and social sustainability (community welfare, cultural preservation, and fair labor practices). In the Thai context, TBL sustainability maps onto several United Nations Sustainable Development Goals (SDGs): SDG 8 (Decent Work and Economic Growth), SDG 12 (Responsible Consumption and Production), SDG 13 (Climate Action), and SDG 17 (Partnerships for the Goals). The government of Thailand has embedded these SGDs in its BCG Economy Model and national tourism strategies. The empirical literature provides strong benchmarks.

For example, Wang and Zhang (2025) reported strong explanatory power for ESG performance in their sample of Chinese tourism SMEs, with GAI adoption, DSCI, and customer involvement. Collectively, their findings explain a substantial share of the variance in sustainability outcomes—a robust structural model result by PLS-SEM standards.

As a complementary outcome construct, innovation performance captures the degree to which firms create competitive differentiation via new services, processes, and business model configurations. In tourism, service innovation encompasses new experience-product designs and digitally enabled personalization. At the same time, process innovation encompasses operational efficiency improvements enabled by AI and digital supply chain tools, and business model innovation encompasses reconfiguration of value propositions and partnership architectures. This study treats innovation performance as a second dependent variable alongside sustainability performance. This approach reflects the theoretical position—supported across RBV and dynamic capabilities literature—that innovation and sustainability capability are mutually reinforcing rather than competing objectives.

Peiró-Signes et al. (2026) corroborated this finding, demonstrating that resource allocation decisions, mediated by internal capability variables, account for a significant portion of the variance in innovation outcomes, reinforcing the general proposition that intermediate organizational mechanisms are critical explanatory links between technological antecedents and performance outcomes.

2.9. Thai microtourism SME context.

Thailand’s tourism industry contributed 11–12% to the gross domestic product (GDP) in the prepandemic years, employing 8 million workers (both directly and indirectly). More than 99% of tourism businesses are classified as MSMEs. Moreover, microenterprises (≤50 employees, ≤50 million THB annual revenue) form the operational backbone across accommodation, guiding, experiential tourism, food and beverage, and transport sub-sectors. The pandemic reduced international arrivals from 39.8 million in 2019 to 6.7 million in 2021, disproportionately affecting microenterprises with limited financial reserves and no access to enterprise-grade risk management tools. Recovery targets (i.e. returning to prepandemic levels by 2025) rely on repositioning around sustainability, digital quality assurance, and experience differentiation.

Operationally, Thai microtourism SMEs are characterized by family ownership, personalized governance, conservative resource allocation, and incremental organizational change. Digital adoption has advanced rapidly through consumer-grade platforms, such as LINE, Facebook, and Google My Business; at the same time, enterprise-grade tools remain inaccessible due to cost and complexity. The SHA Plus certification program has created structured sustainability reporting requirements that incentivize microtourism operators to systematize environmental and social practices. This framework provides a natural policy entry point for AI-assisted documentation. Simultaneously, DASTA’s sustainable tourism area programs and the TAT’s digital transformation agenda further reinforce this institutional landscape.

Wangtueai et al. (2022) documented a significant acceleration in OI-related themes among Thai-affiliated researchers during the pandemic. They identified the need for empirically tested frameworks linking OIO to SME resilience and sustainable growth—a call the present study directly addresses by developing a framework calibrated to the Thai microtourism operating environment.

2.10. Synthesis and research gaps.

Overall, Sections 2.1–2.9 presented a body of literature with considerable empirical depth but unresolved architectural fragmentation. Sections 2.2 and 2.3 demonstrated that GAI adoption and OIO have generated rich but separate research programs. For example, Wang and Zhang (2025) advanced GAI without theorizing OIO, while the open-innovation literature has typically modeled openness as a moderating boundary condition rather than a co-constitutive antecedent. Section 2.4 showed that HAIC has been conceptually gestured toward through qualitative evidence; however, it has never been operationalized as a reflective first-order latent construct with psychometric validation in quantitative tourism research.

Sections 2.5 and 2.6 confirmed that both DSCI and DCST perform robustly as mediators in adjacent contexts (Chinese tourism SMEs; Vietnamese and European general SMEs); however, no study has incorporated all three mediators within a single unified framework alongside dual antecedents. Section 2.9 established that the Thai microenterprise context is the world’s most MSME-dependent tourism economy and that this remains the most consequential gap in the global evidence base. Together, these observations reveal the absence of an integrated empirical framework but also the specific theoretical work that integration must perform.

GAI and OIO must be considered jointly operating codrivers; HAIC must be evaluated from a conceptual placeholder to a validated mediating mechanism; and the entire structure must be anchored within a microenterprise context where resource constraints and governance informality fundamentally alter capability dynamics. Four gaps consolidate the rationale for the proposed framework. First, AI adoption and open innovation remain separate research programs. Wang and Zhang (2025) examined GAI without an OIO construct, while the open-innovation literature has generally modeled openness as a moderator rather than a co-antecedent; no published study has simultaneously modeled both within a unified performance framework. Second, HAIC has been conceptually identified; however, it has not been operationalized as a psychometrically validated mediating construct in quantitative tourism research.

Third, microenterprises have been consistently conflated with the broader SME category in empirical studies, thereby obscuring their distinct capabilities and resource constraints. Fourth, the Thai and broader Southeast Asian context remains substantially underrepresented in the global evidence base. This study’s proposed framework directly addresses each gap.

2.11. Conceptual framework overview.

The proposed framework synthesizes RBV, dynamic capabilities theory, and open-innovation theory to explain how Thai microtourism SMEs achieve superior sustainability and innovation performance through the strategic integration of GAI and OIO. The framework positions GAI and OIO as complementary antecedent resources whose competitive value is realized through three capability-building pathways: HAIC, DSCI, and DCST. CIST serves as a contextual moderator amplifying the translation of HAIC into SP and IP.

This study’s framework comprises eight constructs organized across four functional layers. First, at the antecedent layer, GAI and OIO represent foundational strategic inputs generating competitive distinctiveness. Second, at the mediating layer, HAIC, DSCI, and DCST constitute the operational pathways that convert antecedent resources into value-creating capabilities. Third, at the moderating layer, CIST captures the degree to which customer cocreation amplifies the productivity of human–AI collaboration at the firm-market interface. Fourth, at the outcome layer, SP and IP together provide a multidimensional assessment of competitive and societal value creation. The asymmetric moderation design (in which CIST moderates HAIC–performance relationships but not those between DSCI or DCST) is theoretically motivated by construct-level specificity rather than analytical parsimony.

CIST captures active customer co-design in real-time service delivery. This process is directly interfaced with the HAIC workflow through digital communication channels, such as LINE OA messaging and review platforms. Customers who actively co-design experiences shape how owner-operators calibrate, trust, and iterate upon AI-generated recommendations. In contrast, DSCI primarily operates through interorganizational supply chain coordination with partner firms rather than with end customers. At the same time, DCST reflects long-horizon capability reconfiguration driven by regulatory and environmental sensing rather than immediate customer co-production. Wang and Zhang (2025) found that customer involvement moderated supply chain-to-performance pathways in the Chinese tourism SME context and provided partial boundary-condition support.

The scope of CIST moderation in the present model is accordingly constrained to the HAIC–performance interface, in which customer–operator cocreation is most theoretically proximate and empirically tractable.

Grounded in the RBV, this study’s framework (Figure 1) treats GAI adoption as a potentially VRIN resource whose sustained advantage depends on the quality of integration with human capabilities rather than on possession alone. Dynamic capabilities theory explains how micro-SMEs sense new market opportunities, seize them through digital process reconfiguration, and sustain competitive relevance in an ecologically turbulent environment. Open-innovation theory completes the theoretical architecture by explaining how boundary-spanning knowledge exchange (including customer cocreation, supplier collaboration, and ecosystem participation) expands the productive scope of AI capabilities beyond what resource-constrained microenterprises could achieve independently.

This integration responds to calls in the literature for frameworks that exceed technology adoption and examine capability-mediating mechanisms in resource-constrained contexts.

2.12. Hypotheses development.

This study proposes 14 hypotheses, organized into 3 sets: direct effects from antecedents to mediators (H1a–H2c), direct effects from mediators to performance outcomes (H3a–H5b), and moderation effects of customer involvement (H6a–H6b). Each hypothesis is theoretically grounded in the reviewed literature and empirically anchored in current high-impact evidence.

2.12.1. Direct effects: antecedents → mediators.

GAI systems augment decision-making capacity, automate routine cognitive tasks, and enable personalized guest interactions at scale, thereby altering the nature of human work in microtourism enterprises. In micro-SMEs facing severe resource constraints, the productivity of GAI adoption depends critically on the quality of the human–AI interface, i.e. whether employees develop complementary working skills, trust AI-generated recommendations, and actively collaborate with AI tools in service delivery. Qu and Kim (2025) examined Chinese microenterprises and demonstrated that augmented human–AI collaboration capability is a central mediating mechanism linking AI ecosystem participation to innovation outcomes. Wang and Zhang (2025) confirmed that the collaboration pathway explains the primary share of GAI’s ESG effect. Accordingly, this study proposed the following hypothesis:

H1a: Generative AI adoption positively affects HAIC Quality in Thai microtourism SMEs.

GAI introduces new capabilities for supply chain orchestration, including AI-powered demand forecasting, dynamic inventory optimization, personalized supplier matching, and real-time disruption monitoring. Wang and Zhang (2025) provided direct empirical evidence that GAI usage significantly improves ESG performance via DSCI; their results confirm partial mediation and underscore DSCI as a substantive transmission channel. De Fano et al. (2026) extended this by demonstrating that AI ambidexterity drives dynamic capabilities, with significant effects on firm performance. Moreover, they found that supply chain reconfiguration is a key locus of capability. Accordingly, this study proposed the following hypothesis:

H1b: GAI adoption positively affects digital supply chain innovation in Thai microtourism SMEs.

Dynamic capabilities theory posits that superior performance in turbulent environments requires sensing emerging opportunities, seizing them through strategic investment, and reconfiguring asset bases. Accordingly, GAI enables sensing through market intelligence analytics; seizing through rapid service innovation and personalization; and reconfiguring through AI-assisted resource reallocation. Mousa and Bouraoui (2025) provided longitudinal support for activating dynamic capabilities through digital technology. Similarly, Alhomaid and Al-Romeedy (2026) demonstrated that AI adoption positively affects firms’ dynamic capabilities, including digital innovation and organizational agility. Therefore, this study posits the following hypothesis:

H1c: GAI adoption positively affects dynamic capacity for sustainable tourism in Thai microtourism SMEs.

OIO reflects the systematic leveraging of external knowledge sources to supplement internal innovation processes. For micro-SMEs with limited internal knowledge bases, OIO provides access to AI platform providers, industry associations, and digital tool vendors that offer training resources, implementation templates, and collaborative frameworks to elevate the quality of human–AI working relationships. Open-innovation theory holds that OIO strengthens AI-to-capability pathways by supplying the external knowledge that resource-constrained SMEs lack internally. Similarly, Qu and Kim (2025) demonstrated that AI-enabled innovation ecosystems characterized by strong external collaboration networks significantly enhance augmented human–AI collaboration capabilities. Accordingly, this study proposed the following hypothesis:

H2a: OIO positively affects HAIC Quality in Thai microtourism SMEs.

Digital supply chain innovation in microenterprises depends inherently on the quality and breadth of interorganizational knowledge flows: co-developing digital supply chain solutions with technology partners, sharing real-time data with suppliers, and accessing emerging supply chain platforms through ecosystem participation require OIO-enabled relational capital. Wang and Zhang (2025) found that GAI-driven DSCI effectiveness depends on collaborative relationships with supply chain partners, while open-innovation theory holds that OIO expands the knowledge base available for supply chain adaptation in SME contexts. Accordingly, this study proposed the following hypothesis:

H2b: OIO positively affects digital supply chain innovation in Thai microtourism SMEs.

OIO enriches all three DCST sub-capabilities. First, it enables micro-SMEs to sense sustainability opportunities through external knowledge that would be invisible within narrowly bounded internal information systems. Second, it helps seize innovations through ecosystem partnerships, and third, it assists in reconfiguring service portfolios by incorporating sustainability knowledge from customers, NGOs, regulatory bodies, and destination management organizations. Peiró-Signes et al. (2026) demonstrated that external information sources and interfirm cooperation are among the strongest predictors of innovation performance in resource-constrained SMEs, while open innovation is theorized to be a significant enabler of dynamic capabilities. Accordingly, this study proposed the following hypothesis:

H2c: OIO positively affects dynamic capacity for sustainable tourism in Thai microtourism SMEs.

2.12.2. Direct effects: mediators → performance outcomes.

When HAIC is high, AI systems augment human judgment in sustainability-consequential decision domains, such as energy management, waste minimization, supply sourcing, and community engagement. Such augmentation enables sustainability outcomes that exceed what human-only micro-SME operations could achieve. Wang and Zhang (2025) confirmed that the collaboration pathway explains the primary share of GAI’s ESG effect. Qu and Kim (2025) demonstrated that augmented human–AI collaboration significantly mediates the AI ecosystem’s effects on sustainability outcomes in microenterprises. Accordingly, this study proposed the following hypothesis:

H3a: HAIC quality positively affects SP in Thai microtourism SMEs.

HAIC expands the ideation and prototyping capacity of microenterprises without dedicated R&D resources. AI generates, screens, and iterates service concepts at a speed and breadth that human teams alone cannot achieve. In contrast, human employees provide contextual judgment and cultural sensitivity that AI cannot replicate. De Fano et al. (2026) demonstrated that AI ambidexterity drives firm performance through dynamic capabilities, with human–AI interface quality as the enabling condition. Qu and Kim (2025) confirmed that augmented human–AI coworking capability significantly mediates the effects of the AI ecosystem on innovation performance. Accordingly, this study proposed the following hypothesis:

H3b: HAIC positively affects IP in Thai microtourism SMEs.

DSCI encompasses AI-integrated process innovations in procurement, logistics, inventory management, and partner coordination that reduce waste, improve environmental transparency, and enable responsible sourcing. DSCI offers a particularly potent sustainability pathway for microtourism SMEs whose environmental footprint focuses on supply chain activities. Wang and Zhang (2025) provided sector-direct evidence that DSCI significantly mediates the GAI–ESG relationship, with environmental and social dimensions improving most substantially through digital supply chain transparency. Accordingly, this study proposed the following hypothesis:

H4a: DSCI positively affects SP in Thai microtourism SMEs.

DSCI enables micro-SMEs to develop new service offerings more efficiently through digitally coordinated supply chains, respond to demand signals faster through real-time partner integration, and access novel product inputs through AI-mediated supplier discovery. Peiró-Signes et al. (2026) demonstrated that information source diversity and interfirm cooperation are among the strongest predictors of product and process innovation in resource-constrained SMEs, with the supply chain knowledge network serving as a critical input to innovation. Accordingly, this study proposed the following hypothesis:

H4b: DSCI positively affects IP in Thai microtourism SMEs.

Firms with higher DCST possess adaptive routines that systematically direct resource allocation toward environmental, social, and governance improvement, enabling superior sustainability performance by reconfiguring operations in response to regulatory changes, eco-conscious consumer preferences, and climate-related destination risks. Alhomaid and Al-Romeedy (2026) demonstrated that dynamic capabilities significantly predict firm performance. Mousa and Bouraoui (2025) provided longitudinal evidence that the sequential activation of dynamic capabilities drives adaptability. Accordingly, this study proposed the following hypothesis:

H5a: DCST positively affects SP in Thai microtourism SMEs.

Dynamic capabilities constitute the microfoundational architecture through which firms generate novel resource combinations that produce market-distinguishing innovations. Firms with higher DCST possess the capability to reconfigure, translating sustainability insights (such as evolving eco-preferences, emerging certification frameworks, and destination standards) into new service innovations and operational processes that constitute genuine competitive differentiation. De Fano et al. (2026) demonstrated that dynamic capabilities mediate the effect of AI ambidexterity on firm performance, with a significant indirect effect. Accordingly, this study proposed the following hypothesis:

H5b: DCST positively affects IP in Thai microtourism SMEs.

2.12.3. Moderation hypotheses.

CIST captures the degree to which customers actively co-design, co-evaluate, and covalidate sustainability-oriented tourism products through digital platforms and participatory processes. The moderation hypothesis posits that CIST amplifies HAIC’s translation into sustainability performance by providing demand-side sustainability intelligence—explicit eco-preferences, local knowledge, and sustainability feedback—that directs human–AI collaboration toward the highest-value improvement priorities. Wang and Zhang (2025) provided direct empirical evidence that customer involvement significantly strengthens the positive impact of collaboration mechanisms on ESG performance in Chinese tourism SMEs, with significant positive interaction effects. For Thai microtourism SMEs, where SHA Plus certification and

DASTA-led sustainability standards make customer-validated sustainability credentials increasingly central to market positioning, CIST is expected to be a particularly powerful moderator. Accordingly, this study proposed the following hypothesis:

H6a: Customer involvement in sustainable tourism design positively moderates the relationship between HAIC quality and SP, such that the positive effect of HAIC on SP is stronger when CIST is high.

The second moderation hypothesis extends customer involvement logic to innovation performance, proposing that CIST amplifies HAIC’s innovation returns by enriching ideation inputs available to human– AI collaborative systems. Customers can actively co-create sustainable tourism design by providing inspiration for new experience concepts, evaluating prototypes, and co-creating sustainability narratives. In this way, human–AI collaboration operates with richer, contextually grounded demand-side information, enabling innovations that are technically feasible and market-validated. Open-innovation theory explicitly recognizes customers as a primary source of demand-side innovation knowledge whose integration accelerates and improves innovation outcomes.

Wang and Zhang (2025) confirmed that customer involvement significantly moderates the collaboration-to-performance relationship, and that the innovation performance dimension is particularly sensitive to cocreation quality. Accordingly, this study proposed the following hypothesis:

H6b: Customer involvement in sustainable tourism design positively moderates the relationship between HAIC quality and IP, such that the positive effect of HAIC on IP is stronger when CIST is high.

3. Research methodology.

3.1. Research design.

This study adopts a positivist quantitative paradigm, grounded in the epistemological tradition that social phenomena can be measured, tested, and explained through empirical observation and statistical inference. Positivism is the most defensible philosophical position for this inquiry for three interconnected reasons. First, all eight constructs of interest (GAI adoption, OIO, HAIC, DSCI, DCST, CIST, SP, and IP) are operationalized through observable, self-report indicators whose systematic co-variation can be modeled statistically. Second, the hypothesized relationships are derived from theories with established empirical records (e.g. RBV, dynamic capabilities theory, and open-innovation theory), making deductive hypothesis testing the methodologically appropriate approach.

Third, this study aims to produce findings transferable across Thai microtourism SMEs for nomothetic generalization; it does not aim to provide an ideographic understanding of individual firm experience. Critical realism is the nearest philosophical alternative to the methodology framework; however, its commitment to unobservable generative structures and retroductive reasoning is less appropriate when constructs are directly operationalized through validated reflective indicators and theoretical mechanisms are already specified in prior literature. The deductive approach proceeds from propositions derived from the RBV, dynamic capabilities theory, and open-innovation theory to 14 hypotheses, which are subjected to empirical testing. This study’s findings are interpreted against these propositions to evaluate and refine their explanatory reach in the Thai microtourism context.

Causal language throughout this manuscript—including terms such as ‘affects’ and ‘mediates’—reflects theoretically derived directionality rather than empirically established causation, which cross-sectional data alone cannot confirm.

Within the methodology framework’s layered architecture, the methodological choice is a mono-method quantitative approach, the research strategy is a cross-sectional survey, and the primary data collection technique is a self-administered structured questionnaire. Each occupies a distinct layer that warrants separate justification. As a strategy, the cross-sectional survey provides systematic, standardized data collection across a geographically dispersed population. This approach enables the statistical generalization required by the deductive positivist design. The survey is employed at the firm level, with the microtourism SME as the unit of analysis. Cross-sectional surveys are well-suited to studies mapping relationships among latent constructs at a single point in time and have been widely adopted in tourism management research.

As a technique (distinct from the survey strategy itself), a self-administered, structured questionnaire is used to operationalize the construct measurement. This instrument is the standard data collection tool for PLS-SEM-based research in emerging-economy SME contexts; its structured format ensures replicability and cross-regional comparability across Thailand’s six geographic regions.

3.2. Population and sampling.

3.2.1. Target population.

The target population comprises microtourism SMEs operating in Thailand. Qualifying firms must be classified within the tourism and hospitality services sector, encompassing tour operators, travel agents, accommodation providers, attractions, transportation, food and beverage establishments with a tourism orientation, and community-based tourism enterprises. They must employ no more than 50 full-time equivalent employees with annual revenues not exceeding THB 50 million, consistent with the Thai SME Promotion Act B.E. 2543. Qualifying firms must also have been in operation for a minimum of 1 year at the time of data collection. Finally, they must currently use or plan to use AI technologies within the next 12 months. The target population is theoretically finite; however, the exact size cannot be precisely enumerated.

OSMEP (2023) reports that more than 99% of Thailand’s tourism enterprises are micro-, small-, or medium-sized, but no official business, licensing, or association register provides a complete and current count of microtourism enterprises that simultaneously satisfy all four inclusion criteria. This is most consequential for the fourth criterion, because current or planned AI use is not recorded in any registry, and a substantial proportion of microenterprises operate informally outside these databases. Thus, no exhaustive list of the eligible population exists, and its size cannot be stated with precision from which a probability sample could be drawn. This deliberate delimitation to microenterprises—rather than the broader SME population—reflects their qualitatively distinct resource constraints, limited technological infrastructure, and competitive vulnerabilities.

These constraints make microenterprises an important yet underexamined unit of analysis in the AI adoption and sustainability literature. Key informants are owners, managing directors, and department heads; their strategic decision-making authority makes them the most appropriate sources of firm-level perceptual data.

3.2.2. Sample size determination.

Two complementary approaches were used to confirm the adequacy of the sample size. The primary criterion follows Hair et al. (2019) ten-times rule for PLS-SEM, requiring a minimum of ten observations per maximum number of paths directed at any single endogenous construct. With the most complex endogenous constructs (SP and IP) each receiving paths from 3 mediators and one moderated path, the rule yields a minimum of 50 observations; however, 300 is recommended for models of this complexity. The secondary criterion uses GPower 3.1 power analysis, specifying 80% statistical power to detect a medium effect size (f2 = 0.15) at α = 0.05 with a maximum of 5 predictors, yielding a minimum n of 92.

To accommodate anticipated attrition and non-response, and to meet the requirements of multi-group sensitivity analyses across Thailand’s 6 geographic regions, the target sample is set at 400–500 valid responses. This approach is consistent with comparable PLS-SEM tourism studies (Wang & Zhang, 2025: n = 429).

3.2.3. Sampling strategy.

This study adopted a multistage purposive sampling strategy to ensure theoretical relevance and geographic representativeness. The primary channel operates through the Association of Thai Travel Agents (ATTA) and Thai Tourism Association (TTA) membership databases, which provide access to registered microtourism SMEs across all regions. A secondary channel uses snowball sampling via TAT regional offices and DASTA network coordinators; this approach reaches rural and community-based tourism segments that are underrepresented in association registers. A supplementary channel uses targeted recruitment through Facebook groups dedicated to Thai tourism business operators to reach informal microenterprise networks beyond formal association structures.

This study’s theoretical population is specifically defined—not all tourism firms, but microenterprises with current or planned AI engagement—and a comprehensive national sampling frame with AI adoption status does not exist; thus, purposive sampling is justified. Probability sampling—whether simple random, stratified, or cluster—presupposes an enumerable sampling frame that lists every member of the target population, but for our criterion-defined population, no such frame is available. Eligibility on the fourth criterion (current or planned AI engagement) is a latent characteristic that cannot be read from any pre-existing list and must instead be established through screening at the point of recruitment. This is inherently a purposive operation, and even a random draw from a general tourism register would, therefore, still require purposive screening to identify eligible firms.

Multistage purposive sampling is the methodologically appropriate strategy for a specifically defined population that lacks a complete frame, and the regional stratification and multichannel recruitment described below are used to approximate the representativeness that probability selection would otherwise provide. The resulting non-probability design is acknowledged as an explicit boundary on statistical generalizability (Section 5.5). The multichannel approach mitigates single-source self-selection bias, while geographic stratification ensures representation across Thailand’s north, Central, northeast, East, South, and West regions. Nonetheless, three recruitment channels carry meaningfully different selection probabilities.

Facebook group recruitment will likely overrepresent digitally active, younger, and technology-engaged microenterprise owners (the segment with the highest AI adoption rates), thereby reducing variance in the GAI adoption construct and potentially inflating path coefficients for this predictor. We assess this channel-specific risk by comparing respondent demographic profiles (age, digital platform use, and AI adoption stage) across the three channels at the analysis stage; any statistically significant channel differences will be reported and interpreted as a boundary condition on generalizability.

3.3. Instrument development.

3.3.1. Questionnaire structure.

The survey instrument is a structured, self-administered questionnaire comprising 70 items organized into 3 sections: a screening section (4 items), a main construct measurement section (51 items across 8 latent variables), and a demographic and contextual profile section (15 items). All construct items use a 7-point Likert scale (1 = strongly disagree; 7 = strongly agree). This approach provides sufficient response variance for PLS-SEM estimation and is consistent with standard practice in tourism management research; estimated survey completion time is 12–15 minutes. The instrument was developed through a three-stage process: theoretical domain specification based on the existing literature, item generation and adaptation from validated scales, and content validation using the index of item-objective congruence procedure.

The four screening items verify tourism/hospitality sector operations, employee count within the micro-SME threshold (≤50), at least 1 year of operation, and current or planned AI adoption within 12 months. Respondents failing any criterion were redirected to a polite termination screen. The questionnaire was approved by the Human Research Ethics Committee of Silpakorn University under Protocol COE 68.1224-123, dated 24 December 2025. The IOC assessment involved a panel of three academic experts; content validity was established using the IOC threshold of ≥0.50 per item. The Thai-language version was produced through a full forward-translation, back-translation, and reconciliation protocol following Brislin (1970) to ensure conceptual equivalence; Section 3.4.1 presents the full procedural details.

3.3.2. Measurement scales and operationalization.

All eight constructs are measured using reflective measurement models. This approach is consistent with the theoretical logic that each item is a manifestation of an underlying latent variable rather than a formative cause. Consistent with the positivist measurement paradigm set out in Section 3.1, each of the eight constructs is operationalized not as a directly observable objective fact but as a latent variable inferred from the converging perceptual judgments of knowledgeable key informants (owner-managers and department heads). The operational definitions in Table 1 specify

Note: All items are rated on a 7-point Likert scale (1 = Strongly Disagree; 7 = Strongly Agree). GAI = Generative AI Adoption; OIO = Open Innovation Orientation; HAIC = Human–AI Collaboration Quality; DSCI = Digital Supply Chain Innovation; DCST = Dynamic Capabilities for Sustainable Tourism; CIST = Customer Involvement in Sustainable Tourism Design; SP = Sustainability Performance; IP = Innovation Performance.

each construct’s conceptual domain, while the indicators capture the informant’s assessment of that domain. The 7-point agreement format records the degree to which an informant endorses each indicator statement and is appropriate for behavioral, factual, and evaluative referents, rather than being confined to the measurement of affect. This perceptual key-informant operationalization of firm-level constructs is well established in management, information systems, and tourism research, and its adequacy is treated as an empirical question rather than an assumption. The reflective measurement model reported in Section 4 (indicator loadings, average variance extracted, composite reliability, and heterotrait–monotrait discriminant validity) constitutes the formal test of whether each indicator set validly and reliably represents its latent construct.

Items were drawn from validated scales in the extant literature and adapted to the Thai microtourism SME context. Existing items were developed in non-tourism settings; modifications ensured contextual face validity while preserving construct domain coverage. Table 1 presents a comprehensive summary of all eight constructs—operational definitions, theoretical roles, measurement approaches, item counts, survey items, and primary source references. This specification warrants explicit justification for two constructs that could plausibly be argued as formative. First, OIO encompasses inbound, outbound, and coupled knowledge flows; thus, a reflective model is justified.

All items share a common attitudinal-behavioral disposition toward boundary-spanning collaboration, and the dimensions co-vary as interchangeable manifestations of a single underlying OIO disposition, as confirmed in prior reflective PLS-SEM operationalizations. A formative specification would be appropriate only if the dimensions were non-interchangeable, independent causes of the construct, which is inconsistent with the unidimensional attitudinal framing adopted here. Second, DCST encompasses sensing, seizing, and transforming sub-processes; thus, a reflective model is similarly justified. The three sub-processes co-vary as expressions of an integrated underlying adaptive capacity; firms with high-DCST exhibit all three sub-capabilities simultaneously. Moreover, removing any single item does not alter the construct’s conceptual domain.

This reflective specification is consistent with the closest empirical precedents.

3.4. Pilot study and content validity.

3.4.1. Index of item-objective congruence assessment.

Content validity was established through the IOC procedure before pilot testing. A panel of 3 academic experts (1 tourism management and sustainability scholar, 1 AI and digital innovation specialist, and 1 quantitative survey methodologist with Thai organizational research expertise) used a three-point scale (+1 = item clearly measures the construct; 0 = uncertain; −1 = item does not measure the construct) to independently rate all 51 construct items against their respective construct definitions. Items with IOC values below the conventional threshold (0.50) were revised or removed. The panel also assessed item clarity, cultural appropriateness for Thai respondents, and the accuracy of Thai-language translation; this step employed the back-translation procedure recommended for cross-cultural survey research.

3.4.2. Pilot study.

A pilot study was conducted with 30 microtourism SMEs drawn from the target population to assess psychometric properties before full-scale data collection. All 30 cases were valid with no missing values across the 51 construct items. Internal consistency reliability was evaluated using Cronbach’s alpha (α ≥ 0.70; Hair et al., 2019; Nunnally, 1978), and corrected item-total correlations were inspected; values below 0.30 were flagged for revision or removal.

All eight constructs achieved reliability that met or substantially exceeded the threshold. Generative AI adoption (GAI: α = 0.790) returned an acceptable value consistent with its broad operational scope. OIO (α = 0.876), HAIC (α = 0.882), DCST (α = 0.887), and CIST (α = 0.854) fell within the good range. DSCI (α = 0.904), SP (α = 0.918), and IP (α = 0.937) attained excellent reliability. Item-level RITC values ranged from 0.394 (CIST3) to 0.913 (IP3), with all 51 items exceeding the 0.30 retention threshold. CIST3 produced the lowest RITC (0.394); however, its removal would have improved α by only 0.014, and the item captures a theoretically distinct feedback-influence mechanism. Therefore, it was retained on grounds of content validity. No items were removed from the instrument.

Respondent feedback confirmed that the instructions were clear, the terminology was accessible, and the completion time was appropriate (12–15 minutes); thus, the questionnaire proceeded to full-scale data collection without modification.

3.5. Data collection.

Consistent with the high internet penetration and smartphone adoption among Thai SME operators and following the approach adopted by Wang and Zhang (2025), Data collection was conducted via an online self-administered questionnaire hosted on Google Forms. The questionnaire was distributed through three parallel channels: direct email distribution through ATTA and TTA member lists; dissemination through TAT regional coordinators and DASTA network managers; and targeted posting in closed Facebook groups for Thai tourism business operators. Personalized cover letters were sent with all distributions to explain study objectives, ethical protections, and participant rights.

The data collection period spanned from January 1 to March 31, 2026 (12 weeks), with 3 reminder communications sent to non-responding contacts at weekly intervals. To incentivize participation, 10 gift vouchers of THB 500 each were offered through a post-collection lottery draw; contact information for lottery entry was collected separately from survey responses on a distinct form to preserve response anonymity. Four data quality controls were embedded. Four screening questions automatically redirected ineligible respondents; 2 attention-check items identified careless responding; a minimum completion time threshold of 8 minutes flagged responses for review; and IP address monitoring prevented duplicate submissions.

Non-response bias was assessed using the temporal wave comparison procedure recommended by Armstrong and Overton (1977), in which demographic and key construct scores of early respondents (first tertile by submission date) were compared with those of late respondents (final tertile) using independent-samples t-tests. Late respondents were treated as proxies for non-respondents; statistically significant differences will be interpreted as indicative of non-response bias and discussed as a boundary condition on generalizability. Section 4.1 presents the results.

3.6. Data analysis approach.

3.6.1. Justification for PLS-SEM.

PLS-SEM, implemented in SmartPLS 4.0, was selected as the primary analytical method for five reasons. First, the structural model is complex—8 latent constructs, 14 hypothesized paths, 3 mediating mechanisms, and 1 moderated relationship. Under these conditions, PLS-SEM’s sequential partial estimation avoids convergence failures that affect covariance-based SEM. Second, the research objective is causal-predictive rather than strictly confirmatory; thus, PLS-SEM’s orientation toward prediction, assessed through Q2 blindfolding and PLSpredict, directly serves this dual objective. Third, non-normal data distributions are expected in an emerging-economy SME context, where owner-manager responses may exhibit floor or ceiling effects on technology adoption items; PLS-SEM’s distribution-free bootstrapping is robust to violations of multivariate normality.

Fourth, the method ensures direct comparability with the most proximate empirical studies. For example, Wang and Zhang (2025), Anser et al. (2025), and De Fano et al. (2026) all employed SmartPLS for structurally analogous models. Fifth, all constructs are measured reflectively, and PLS-SEM’s well-validated heterotrait–monotrait (HTMT)-based procedures for discriminant validity and composite reliability provide an appropriate and rigorous framework for measurement evaluation.

3.6.2. Five-step analysis sequence.

Following the PLS-SEM evaluation guidelines of Hair et al. (2019), the analysis proceeds through five sequential steps.

Step 1 (preliminary analysis) screens for missing values (≤5% per item), univariate and multivariate outliers, and distributional characteristics through descriptive statistics and normality assessment via skewness and kurtosis. This step also assesses common method bias (CMB) through two complementary statistical procedures: (a) Harman’s single-factor test and (b) full collinearity variance inflation factor (VIF) assessment. Harman’s single-factor test requires that the first unrotated factor explain less than 50% of the total variance; under the full collinearity VIF assessment, values below 3.3 indicate the absence of common method variance. Procedurally, CMB risk is attenuated by separating predictor and criterion items into distinct questionnaire sections with an intervening demographic block.

Marker-variable analysis is used as a supplementary check, and an attention-check item review is also conducted at this stage. Section 4.2 presents the complete CMB results.

Step 2 (measurement model assessment) evaluates four psychometric properties in sequence. Indicator reliability is assessed using outer loadings, with a threshold of ≥0.70; items with loadings between 0.40 and 0.70 are evaluated for their impact on average variance extracted (AVE) before finalizing removal decisions. Internal consistency is assessed using Cronbach’s α (≥0.70) and Composite Reliability (CR ≥0.70; ≤0.95), with the upper bound guarding against item redundancy. Convergent validity requires that each construct’s AVE be ≥0.50, indicating that each construct explains more than half of its indicators’ variance. Discriminant validity is established through the HTMT ratio (HTMT <0.85 for conceptually distinct constructs) and the Fornell–Larcker criterion.

Step 3 (Structural Model Assessment) begins with a collinearity check among predictors (VIF <3.3 preferred; <5.0 maximum). Path coefficients are then estimated via bootstrapping using 5,000 subsamples and bias-corrected accelerated 95% confidence intervals. Explanatory power is interpreted against R2 benchmarks of 0.25 (weak), 0.50 (moderate), and 0.75 (substantial). Effect sizes are quantified using Cohen’s (1988) f2 thresholds of 0.02 (small), 0.15 (medium), and 0.35 (large). Predictive relevance is assessed via Q2 (>0 required; >0.25 medium; >0.50 large), and overall model fit is evaluated using the standardized root mean square residual (SRMR <0.08).

Step 4 (mediation analysis) tests specific indirect effects via bootstrapping (5,000 subsamples, BCa 95% CIs) following the procedures of Nitzl et al. (2016), with significance confirmed when the confidence interval excludes zero. The variance accounted for (VAF) statistic classifies the type of mediation: VAF <20% indicates no mediation, 20–80% indicates partial mediation, and >80% indicates full mediation. Total indirect effects are also computed for the supplementary IPMA.

Step 5 (moderation analysis) employs the product indicator approach for the HAIC × CIST interaction term. The interaction path coefficient was tested via bootstrapping (5,000 subsamples, BCa 95% CIs); positive and significant coefficients confirmed H6a and H6b. Finally, simple-slope analysis at ±1 standard deviation (SD) of the CIST moderator and Johnson–Neyman floodlight analysis were conducted. These analyses aim to identify the range of CIST values over which the HAIC-to-outcome relationship is statistically significant.

3.7. Ethical considerations.

The research protocol received formal ethical approval from the Human Research Ethics Committee of Silpakorn University’s Research, Innovation, and Creative Work Administration Office (Protocol REC 68.1219-234-11346; COE 68.1224-123; dated 24 December 2025) under the Exemption Review procedure. This approval confirmed compliance with Belmont Report standards and applicable national legislation. Informed consent was obtained from all participants via a bilingual information sheet (Thai and English) at the start of the questionnaire. This sheet explains the study’s purpose, the voluntary nature of participation, the right to withdraw without penalty, and data-handling procedures; participation is conditional on explicit consent. Anonymity was maintained throughout—no individual firm or respondent is identifiable in any reported output.

All data are stored in encrypted form on secure institutional servers accessible only to the research team. Respondent contact information for lottery purposes is stored separately from survey responses and will be destroyed after prize distribution. In accordance with Silpakorn University’s institutional data governance policies, research data will be destroyed upon completion of the study.

4. Results.

This section presents the empirical findings following the 5-step PLS-SEM analysis of 450 valid responses from Thai microtourism SMEs. The presentation proceeds sequentially through: respondent profile and preliminary analysis; CMB assessment; measurement model evaluation; structural model assessment; mediation analysis; and moderation analysis. All analyses were conducted in SmartPLS 4.0 using 5,000 bootstrap subsamples and bias-corrected accelerated 95% confidence intervals. The findings are organized to directly address the study’s three research questions. The structural model assessment (Section 4.4) answers RQ1 by quantifying how GAI adoption and OIO jointly affect HAIC, DSCI, and DCST.

The mediation analysis (Section 4.5) answers RQ2 by establishing the transmitting roles of the three mediators, and testing whether HAIC operates as a distinct first-order mechanism, and the moderation analysis (Section 4.6) answers RQ3 by evaluating how CIST conditions the mediator–performance relationships.

4.1. Respondent profile and preliminary analysis.

4.1.1. Sample characteristics.

Section 3.2.3 specifies the intended sampling design, but this section reports the achieved sample together with the preliminary data screening. A total of 498 questionnaire responses were received across the three distribution channels. After applying the four screening criteria and removing 48 incomplete or failed attention-check responses, 450 valid cases were retained for analysis, yielding a valid response rate of 90.4%. Of the 48 cases removed, 21 failed the micro-SME size criterion (≤50 employees or ≤50 million THB revenue), 14 were removed for incomplete responses (≥10% missing items), 9 failed the minimum completion time threshold (under 8 minutes), and 4 failed 1 or both attention-check items.

No responses were excluded based on sector eligibility or AI adoption planning criteria, confirming that the three recruitment channels effectively prescreened for the target population on these dimensions. This response exceeds the target sample of 400– 500 and is consistent with comparable PLS-SEM tourism studies (Wang & Zhang, 2025: n = 429).

Note: n = 450. Percentages may not sum to 100 due to rounding.

Table 2 presents the complete demographic profile, showing that respondents were distributed across all six Thai geographic regions, with the largest representation from the Central (25.1%) and Southern (23.8%) regions. Regarding business size, 54.3% of respondents employed 10 or fewer staff, confirming the sample’s microenterprise composition. Accommodation (22.4%) and tour operators/travel agents (20.9%) were the dominant business types. Owners and cofounders constituted the largest respondent category (44.9%), followed by senior executives (27.1%), ensuring strategic-level informants in most cases. Regarding AI adoption status, 62.2% of respondents were currently using AI tools, and the remaining 37.8% were planning adoption within 12 months—consistent with the population eligibility criteria.

To assess whether including AI-planning respondents affected the structural estimates, a sensitivity analysis was conducted by re-running the PLS-SEM model on the subset of current AI users only (n = 280). Path coefficients for all 12 supported hypotheses remained directionally consistent and statistically significant (p < 0.001) in the subsample, with a maximum difference in coefficients of Δβ = 0.031 relative to the full-sample estimates. These results confirm that the findings are not materially driven by respondents with aspirational rather than actual AI adoption. Moreover, the full-sample estimates are appropriate for interpretation.

4.1.2. Descriptive statistics.

Table 3 presents the construct-level descriptive statistics computed from the 450-case dataset. All construct mean scores ranged from 3.66 (GAI) to 4.07 (HAIC) on the 7-point scale, indicating moderate to moderate-high levels of AI adoption, open-innovation orientation, and capability development. Standard deviations ranged from 1.01 to 1.11, confirming adequate variance for the PLS-SEM estimation. Skewness and excess kurtosis values were within the acceptable thresholds (skewness: <2; excess kurtosis: <7), indicating no severe violations of distributional assumptions. The final dataset contained no missing values (missing = 0 for all 51 items), confirming complete data integrity. All 450 retained cases passed both embedded attention-check items, indicating no careless or random responding in the analyzed sample.

4.1.3. Non-response bias assessment.

Consistent with the procedure specified in Section 3.5, non-response bias was assessed using the temporal wave comparison method. The respondents were ordered by their response sequence and divided into an early-response group (first tertile, n = 150) and a late-response group (final tertile, n = 150), with late respondents treated as proxies for non-respondents. Independent-samples t-tests compared the two groups across the eight latent constructs, and chi-square tests compared the distributions of five demographic characteristics (firm size, AI-adoption status, respondent role, geographic region, and business type), for 13 comparisons. None of the eight construct comparisons was statistically significant (all p > 0.05; absolute t-values from 0.30 to 1.61), and four of the five demographic comparisons were likewise non-significant.

The sole exception was geographic region (χ2 = 12.48, df = 5, p = 0.029), which did not survive a Bonferroni correction for the 13 comparisons (adjusted α = 0.004) and was therefore attributable to chance. The eight latent constructs—the variables that enter the structural model—exhibited no systematic differences across response waves, and non-response bias was, therefore, unlikely to materially threaten the validity or generalizability of the findings.

4.2. Common method bias (CMB) assessment.

Two complementary procedures were used to assess CMB. The first procedure included Harman’s single-factor test, the Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy, and Bartlett’s test of sphericity. Harman’s single-factor test was conducted in SPSS 27 by forcing all 51 measurement items into a single-component principal component analysis. The single extracted factor accounted for 38.48% of total variance—substantially below the 50% threshold. This outcome indicates that no single factor dominated the variance structure; thus, CMB does not represent a critical threat to the data. The KMO measure of sampling adequacy was 0.968, and Bartlett’s test of sphericity was statistically significant (χ2(1275) = 15,719.73, p < 0.001), confirming the factorability of the correlation matrix.

Second, following Kock (2015), the full collinearity VIF procedure was applied within SmartPLS. All indicator-level VIF values for the full collinearity assessment were below 3.3 (range: 1.734–3.130), well below the threshold for serious common method variance. These two convergent results provide reasonable assurance that the CMB is absent in the structural estimates reported in subsequent sections.

4.3. Measurement model assessment.

4.3.1. Indicator reliability and internal consistency.

Table 4 presents the comprehensive measurement model results. The table integrates indicator-level and construct-level evidence across four validity and reliability dimensions: convergent validity (outer loadings, indicator reliability, and AVE), internal consistency reliability (Cronbach’s alpha, ρA, and ρC), and discriminant validity (HTMT). All 51 indicator outer loadings exceeded the 0.70 threshold across all 8 reflective constructs, with loadings ranging from 0.741 (SPSOC3) to 0.894 (IP1). These results confirm that each item explains more than 50% of its latent variable’s variance. Indicator reliability values (λ2) ranged from 0.549 (SPSOC3) to 0.799 (IP1). These results are all above the 0.50 benchmark, providing item-level evidence that none of the indicators is dominated by measurement error.

Removing the single borderline case, HAIC1 (λ = 0.744, IR = 0.554), would have marginally reduced AVE while impairing content coverage of the human–AI complementarity dimension; therefore, this case was retained.

Internal consistency reliability was uniformly strong across all three reported indices. First, Cronbach’s alpha ranged from 0.892 (CIST) to 0.928 (IP). Second, Dijkstra–Henseler’s rho (ρA) ranged from 0.893 (CIST) to 0.930 (IP), closely mirroring alpha values and confirming the robustness of the reliability estimates to the equal-weighting assumption of Cronbach’s alpha. Third, Composite reliability (ρC) ranged from 0.917 (CIST) to 0.943 (IP) (all within the recommended 0.70–0.95 range), eliminating redundancy concerns. Regarding discriminant validity at this stage, the maximum HTMT value for each construct (reported in the final column of Table 4) ranged from 0.548 (CIST) to 0.673 (SP and IP). All values were substantially below the 0.85 threshold, providing construct-level confirmation that each latent variable is empirically distinct from all others in the model.

4.3.2. Convergent validity.

Convergent validity was confirmed for all eight constructs: AVE values ranged from 0.619 (SP) to 0.735 (IP), all exceeding the 0.50 threshold established by Fornell and Larcker (1981). The SP construct recorded the lowest AVE, reflecting the theoretically justified heterogeneity of its three-dimensional TBL structure (environmental, social, and economic sub-dimensions); however, it still meets the minimum criterion.

4.3.3. Discriminant validity.

Discriminant validity was assessed using the HTMT ratio. Table 5 shows that all HTMT values were below the conservative 0.85 threshold for conceptually distinct constructs, ranging from 0.392 (GAI–CIST) to 0.673 (SP–IP). The highest value, between SP and IP (HTMT = 0.673), theoretically reflects expected relatedness between the 2 performance outcomes rather than discriminant validity

Note: n = 450. IR = Indicator Reliability (λ2). AVE = Average Variance Extracted. ρA = Dijkstra–Henseler’s rho. ρC = Composite Reliability. HTMT = Heterotrait-Monotrait ratio; ‘Yes’ indicates all pairwise HTMT values are significantly lower than the 0.85 threshold, confirming discriminant validity. All loadings > 0.70, AVE > 0.50, α and ρC within 0.60–0.90/0.60–0.90 range.

failure; it also remains well within the acceptable range. The Fornell–Larcker criterion was also satisfied: the square root of each construct’s AVE (diagonal values: 0.787–0.857) exceeded all interconstruct correlations. Collectively, these results confirm adequate discriminant validity across all construct pairs.

The estimated model fit was acceptable: SRMR = 0.075 (below the 0.08 threshold), and Bentler–Bonnet normed fit index (NFI) = 0.896. The NFI value falls marginally below the conventional 0.90 threshold; however, NFI is sensitive to model complexity and sample size in PLS-SEM and is not the primary recommended fit criterion for this estimator. Instead, the

SRMR of 0.075 satisfies the recommended criterion, and the saturated model SRMR of 0.036 confirms a strong global fit when all possible paths are estimated. Combined with the individual-level reliability and validity results above, these indicators confirm that the measurement model provides an adequate foundation for structural model assessment.

4.4. Structural model assessment.

4.4.1. Predictor collinearity.

Before examining path coefficients, predictor collinearity was assessed using within-model VIF values. All indicator-level VIF values were below 3.3 (range: 1.734–3.130), with DSCI2 (VIF = 3.020) and DSCI5 (VIF = 3.130) recording the highest values (both are within the recommended threshold). These values are consistent with the full collinearity VIF range described in Section 4.2, which used the same 51-indicator set assessed within SmartPLS. No multicollinearity concerns were identified at either the indicator or construct level.

4.4.2. Path coefficients and hypothesis testing.

Addressing RQ1 (how GAI adoption and OIO jointly affect HAIC, DSCI, and DCST), Table 6 presents the complete results of the structural model. All 6 antecedent-to-mediator paths (H1a–H2c) were positive

Note: n = 450. HTMT = Heterotrait-Monotrait ratio. Diagonal values represent the square root of the Average Variance Extracted (√AVE) for each construct, enabling direct verification of the Fornell-Larcker criterion: each √AVE (range: 0.787–0.857) exceeds all off-diagonal correlations in its row and column. All off-diagonal values < 0.85 threshold, confirming discriminant validity for all construct pairs. Interaction terms (CIST × HAIC; CIST × DSCI) are excluded as single indicators.

Note: n = 450; bootstrapping with 5,000 subsamples; BCa = bias-corrected accelerated confidence intervals. β = standardized path coefficient; SE = standard error; t = t-statistic; p = two-tailed p-value. R2 benchmarks: 0.25 = weak; 0.50 = moderate; 0.75 = substantial. All Q2predict > 0 confirm predictive relevance; PLS-SEM RMSE < Linear Model RMSE for all indicators, confirming predictive superiority of the non-linear PLS-SEM model. Bold values indicate the standardized path coefficients (β).

and statistically significant (p < 0.001). GAI adoption demonstrated the strongest direct effect on HAIC (β = 0.427; t = 9.934), confirming that the deployment of generative AI is the primary driver of HAIC quality. OIO exerted the strongest effect on DCST (β = 0.451; t = 12.597), the highest path coefficient in the entire antecedent layer. This result underscores open-innovation orientation as the dominant driver of dynamic capability development for sustainable tourism.

All six mediator-to-outcome paths (H3a–H5b) were positive and statistically significant (p < 0.001). Among the mediators, HAIC demonstrated the largest effect on IP (β = 0.287; t = 6.812). In comparison, DSCI exerted comparable effects on both SP (β = 0.249; t = 5.532) and IP (β = 0.252, t = 5.363). DCST significantly contributed to SP (β = 0.216; t = 4.761) and IP (β = 0.244; t = 5.594). The direct effect of CIST on SP was also positive and significant (β = 0.208; t = 5.411), confirming that customer involvement is an independent contributor to sustainability performance and exceeds its hypothesized moderating role.

4.4.3. Explained variance and predictive relevance.

The structural model explained 35.5% of the variance in HAIC (R2 = 0.355), 34.3% in DSCI (R2 = 0.343), and 42.1% in DCST (R2 = 0.421); all outcomes were within the moderate range. For the primary outcome variables, SP achieved R2 = 0.478 (approaching substantial), and IP achieved R2 = 0.427 (moderate). These results indicate that the theorized antecedents, mediators, and moderators collectively account for half of the variance in sustainability performance and 43% of the variance in innovation performance. The factors underlying the unexplained variance in the HAIC and DSCI mediators are considered in Section 5.5.

PLSpredict Q2predict values exceeded 0 for all endogenous constructs (DCST = 0.415, DSCI = 0.336, HAIC = 0.349, SP = 0.369, and IP = 0.285), confirming medium predictive relevance. Critically, PLS-SEM root mean square error (RMSE) values were lower than the naïve linear model (LM) RMSE for all indicator-level predictions. This outcome indicates that the nonlinear PLS-SEM model outperforms a simple benchmark in out-of-sample prediction—a key criterion for establishing practical relevance. Across all 51 indicators, PLS-SEM RMSE ranged from 0.761 to 1.042, compared with LM RMSE values of 0.798 to 1.089 for the same indicators. This disparity represents a consistent predictive advantage of the PLS-SEM specification over the linear baseline.

4.4.4. Effect sizes.

Effect size (f2) calculations from the f-square matrix revealed a medium effect for GAI on HAIC (f2 = 0.218) and a medium-to-large effect for OIO on DCST (f2 = 0.270)—the largest in the model, approaching the large-effect threshold of f2 ≥0.35. Finally, the calculations revealed a small-to-medium effect for OIO on DSCI (f2 = 0.141, just below the medium threshold of 0.15). These distinctions confirm that these pathways are statistically significant and substantively meaningful. Moreover, OIO’s dynamic capability-building effect (OIO → DCST) is the most substantively impactful relationship in the entire structural model. GAI’s effects on DCST (f2 = 0.118) and DSCI (f2 = 0.131) were medium in magnitude.

Mediator-to-outcome effects were predominantly small to medium, including HAIC on IP (f2 = 0.093), DSCI on IP (f2 = 0.067), DSCI on SP (f2 = 0.069), DCST on SP (f2 = 0.051), and DCST on IP (f2 = 0.062). The direct CIST-to-SP effect registered a small-to-medium effect (f2 = 0.063). This result is consistent with CIST’s positioning as a moderator whose primary explanatory role is boundary-condition specification rather than main-effect magnitude.

4.5. Mediation analysis.

Addressing RQ2 (the mediating roles of HAIC, DSCI, and DCST, and whether HAIC functions as a theoretically distinct first-order mechanism), Table 7 presents the specific indirect effects for all 12 mediation pathways, showing that all 12 indirect effects were statistically significant (BCa CIs excluded 0; p < 0.001). These results provide strong support for the mediating roles of HAIC, DSCI, and DCST in transmitting the effects of GAI adoption and OIO on sustainability and innovation performance.

Note. n = 450; 5,000 bootstrap subsamples; BCa 95% confidence intervals. VAF = Variance accounted for = indirect effect/total effect × 100. Mediation type classification: VAF 20–80% = partial mediation. All indirect effects are statistically significant (BCa CIs exclude zero). No full mediation was detected across any pathway. Bold values indicate the standardized specific indirect effects (β).

VAF values ranged from 23.0% (OIO → HAIC → SP) to 43.9% (GAI → HAIC → IP), classifying all pathways as partial mediation (VAF 20–80%). The absence of full mediation (VAF >80%) across all 12 pathways classifies all relationships as partial mediation; thus, the 3 specified mediators account for a substantial but not exclusive portion of the antecedent-to-performance effects. To confirm the partial mediation classification empirically, we estimated direct paths from GAI adoption and OIO to SP and IP in a supplementary model alongside the mediated paths: GAI → SP (β = 0.148; p < 0.001), GAI → IP (β = 0.156; p < 0.001), OIO → SP (β = 0.122; p = 0.003), and OIO → IP (β = 0.139; p = 0.001).

The results were all positive and significant, confirming that residual direct antecedent-to-performance effects co-exist with the mediated pathways, empirically justifying the partial mediation classification, and suggesting the presence of additional transmission mechanisms beyond the three specified mediators.

GAI → HAIC → IP (indirect β = 0.123; VAF = 43.9%) was the strongest mediation pathway, confirming HAIC as the primary mechanism through which AI adoption translates into innovation performance. Regarding sustainability performance, OIO → DCST → SP (indirect β = 0.097; VAF = 40.6%) and GAI → HAIC → SP (indirect β = 0.091; VAF = 38.1%) were the most substantive pathways. The DSCI-mediated pathways showed VAFs ranging from 30.0% to 36.0%. In contrast, the DCST-mediated pathways for GAI ranged from 26.1% to 26.8%, reflecting somewhat weaker, but still significant, mediating contributions. OIO’s dynamic capability pathway to performance (OIO → DCST → IP: VAF = 40.4%; OIO → DCST → SP: VAF = 40.6%) was comparable in strength to the HAIC pathway. This result underscores DCST as a substantive mechanism for the performance effects of open innovation.

As a supplementary analysis specified in Step 4 of the analytical protocol, an IPMA was conducted using total indirect effects to identify which mediating constructs offer the greatest combined importance and performance-improvement potential for sustainability and innovation outcomes. The IPMA results indicate that HAIC assigns the highest importance to IP (total indirect effect = 0.197) and the second highest to SP (0.146). In contrast, DCST assigns the highest importance to SP (total indirect effect = 0.161) when OIO is the antecedent. DSCI occupies a consistently midrange importance position across both outcomes (SP: 0.169; IP: 0.172). In terms of performance (mean latent variable scores), HAIC (M = 4.07) and DCST (M = 3.82) present moderate improvement headroom relative to the scale maximum. In contrast, DSCI (M = 3.82) similarly indicates scope for targeted managerial intervention.

These IPMA findings corroborate the path coefficient results. They also highlight HAIC and DCST as the highest-leverage capability investments for Thai microtourism SME managers seeking to improve both sustainability and innovation performance through AI adoption and open-innovation strategies.

4.6. Moderation analysis.

Addressing RQ3 (how CIST conditions the mediator–performance relationships), Table 8 presents the results of the moderation analysis for H6a and H6b, showing that neither hypothesis was supported. The

Note. n = 450; 5,000 bootstrap subsamples. Neither moderation hypothesis (H6a, H6b) was supported; interaction terms CIST × HAIC → SP (β = −0.007, p = 0.867) and CIST × DSCI → SP (β = −0.039, p = 0.320) were non-significant. Simple slope analysis confirms the HAIC → SP relationship is stable across CIST levels (range: β = 0.207–0.220), and the DSCI → SP relationship shows a slightly decreasing pattern at higher CIST (+1 SD: β = 0.210 vs. −1 SD: β = 0.287), contrary to the hypothesized amplification effect. Conditional indirect effect columns represent post-hoc conditional process analysis re-estimated at ±1 SD of CIST; they confirm that CIST does not moderate the full indirect pathway from GAI to SP.

interaction term, CIST × HAIC → SP, produced a nonsignificant coefficient (β = −0.007; t = 0.168, p = 0.867; BCa 95% CI [−0.091, 0.067]). Similarly, the CIST × DSCI → SP interaction was also nonsignificant (β = −0.039; t = 0.995; p = 0.320; BCa 95% CI [−0.116, 0.036]). Effect sizes for both interaction terms were negligible (f2 = 0.000 and 0.002, respectively), confirming no moderating influence of customer involvement on the HAIC-to-SP or DSCI-to-SP relationships in this sample.

Simple-slope analysis at ±1 SD of CIST provides further interpretive insight (Figures 2 and 3). For HAIC → SP, slopes were stable across CIST levels (−1 SD: β = 0.220; mean: β = 0.214; +1 SD: β = 0.207), indicating a consistent positive direct effect of HAIC on sustainability performance across customer involvement intensities (Figure 2). For DSCI → SP, a marginal reversal is observed (−1 SD: β = 0.287; mean: β = 0.249; +1 SD: β = 0.210). This result suggests that at higher levels of customer involvement, the DSCI-to-SP relationship slightly attenuates rather than amplifies—a pattern contrary to H6a’s prediction but not statistically significant (Figure 3).

These findings suggest that, in the Thai microtourism SME context, customer involvement in sustainable design primarily functions as an independent driver of sustainability performance (direct β = 0.208, p < 0.001) rather than as a boundary condition that amplifies capability-mediated pathways.

As a complementary procedure specified in Step 5 of the analytical protocol, the Johnson–Neyman floodlight analysis was conducted to determine the precise range of CIST values over which the HAIC–SP and DSCI–SP relationships attain and lose statistical significance. This approach eliminates sole reliance on the arbitrary ±1 SD breakpoints used in simple-slope analysis. The Johnson–Neyman analysis identified no significant region for the CIST × HAIC → SP interaction. The HAIC → SP relationship was positive and statistically significant across the entire observed range of CIST (from −1.84 to +1.91 on the standardized scale), with significance levels consistently below α = 0.05 (Figure 2). This result confirms that the HAIC → SP pathway is unconditionally robust across CIST levels and that the nonsignificant interaction term does not mask any region-specific moderation.

For the CIST × DSCI → SP interaction, floodlight analysis similarly revealed no statistically significant interaction region. The DSCI → SP relationship remained significant across all observed CIST values, whereas the interaction coefficient itself remained nonsignificant throughout (Figure 3). Collectively, the Johnson–Neyman results corroborate the simple-slope findings, reinforcing the conclusion that CIST does not function as a moderator of capability-mediated sustainability pathways in this sample.

4.7. Summary of hypothesis testing.

Twelve of the 14 hypotheses tested were supported, while 2 were not. All 6 antecedent-to-mediator hypotheses (H1a–H2c) and all 6 mediator-to-outcome hypotheses (H3a–H5b) were statistically supported; their significant positive path coefficients were confirmed through 5,000 bootstrap subsamples and BCa confidence intervals. The two moderation hypotheses (H6a and H6b) were not supported, as the interaction terms were nonsignificant and the effect sizes were negligible. These findings confirm the core mediating architecture of the proposed framework. HAIC, DSCI, and DCST are theoretically coherent and empirically robust transmission mechanisms between AI adoption, open-innovation orientation, and dual performance outcomes in Thai microtourism SMEs.

At the same time, the results indicate that customer involvement in sustainable tourism design operates as a direct performance driver rather than a conditional amplifier of capability-mediated pathways.

5. Discussion.

5.1. Overview of the major findings.

The central finding of this study is that OIO and GAI adoption function as co-equal antecedents of sustainability and innovation performance in Thai microtourism SMEs. OIO produces the model’s highest path coefficient (OIO → DCST: β = 0.451) and its total effects on sustainability performance (0.239) and innovation performance (0.272), matching those of GAI adoption (0.239; 0.280); this result challenges the implicit primacy of technology adoption in the AI-performance literature. Three mediating mechanisms (HAIC, DSCI, and DCST) all transmitted these effects with full statistical support: 12 of 14 hypotheses were supported, and all mediation VAF values ranged from 23.0% to 43.9%, confirming partial mediation. The structural model explained R2 = 0.478 for SP and R2 = 0.427 for IP, with Q2predict values confirming medium predictive relevance across all endogenous constructs.

The study’s most unanticipated result was the non-significance of both CIST moderation hypotheses (H6a: CIST × HAIC → SP; β = −0.007; p = 0.867; H6b: CIST × DSCI → SP; β = −0.039; p = 0.320), while CIST simultaneously exerted a significant direct effect on SP (β = 0.208; p < 0.001). These results indicate that in this context, customer involvement functions as an independent additive sustainability driver rather than a conditional amplifier of capability-mediated pathways.

5.2. Interpretation of the findings.

Interpreted in the order of the study’s three research questions, the discussion first turns to RQ1, which asks how GAI adoption and OIO jointly affect HAIC, DSCI, and DCST. The strongest antecedent-layer path was GAI → HAIC (β = 0.427, t = 9.934), confirming generative AI deployment as the primary activator of human–AI collaboration quality. This result is consistent with Qu and Kim (2025) qualitative evidence that augmented human–AI collaboration capability mediates AI ecosystem effects on innovation outcomes. It also supports Wang and Zhang (2025) finding that the collaboration pathway accounts for the largest share of GAI’s ESG effect. Notably, OIO produced the highest path coefficient in the entire model: OIO → DCST (β = 0.451, t = 12.597).

This result exceeded all GAI antecedent effects, establishing open knowledge orientation as the pre-eminent driver of sustainability-relevant dynamic capabilities and positioning OIO’s total performance effects (SP: 0.239; IP: 0.272) in near-parity with GAI’s (SP: 0.239; IP: 0.280). All 12 mediation pathways were significant (VAF range: 23.0–43.9%); the strongest channel was GAI → HAIC → IP (VAF = 43.9%), and the strongest sustainability channel was OIO → DCST → SP (VAF = 40.6%). The consistent partial mediation (no pathway exceeded VAF = 80%) indicates that GAI and OIO also exert direct performance effects not fully channeled through HAIC, DSCI, or DCST. This outcome suggests the presence of additional transmission mechanisms, such as reputational signaling through booking platforms or direct AI-enabled cost efficiencies.

The second research question (RQ2) concerns the mediating roles of HAIC, DSCI, and DCST and whether HAIC operates as a theoretically distinct first-order mechanism. This study’s most theoretically distinctive finding is the asymmetry in capability specialization within the triple-mediation architecture. The three mediators do not operate as equivalent transmission channels but rather serve distinct functional roles, each calibrated to a different performance domain. HAIC most strongly mediates innovation performance (GAI → HAIC → IP; VAF = 43.9%), indicating its role as a generative ideation and prototyping capability that enables the co-production of novel service concepts through iterative human–AI cycles.

DCST most strongly mediates sustainability performance when OIO is the antecedent (OIO → DCST → SP; VAF = 40.6%), reflecting its regulating and environmental-sensing capability activated by external knowledge flows. DSCI occupies a balanced intermediate position across both outcomes (VAF: 30.0–36.0%), reflecting its operational coordination function that simultaneously reduces waste and enables service customization. This asymmetry implies that managers seeking returns on innovation performance should prioritize investments that strengthen the human–AI collaborative interface. In contrast, managers targeting sustainability performance should prioritize open knowledge partnerships that activate dynamic reconfiguration capacities, treating these as distinct intervention pathways rather than generic ‘AI adoption’ investments.

The third research question (RQ3) addresses how CIST conditions the mediator–performance relationships. The nonsignificant CIST moderation is this study’s most unanticipated finding and warrants careful interpretation. Three explanations are most plausible. Thai microtourism operators engage customers primarily through reactive LINE OA and Facebook channels rather than the proactive co-design processes required to condition capability effectiveness. The CIST scale captures general responsiveness rather than AI-mediated cocreation intensity sufficient to produce moderation effects. The strong direct CIST → SP effect (β = 0.208) suggests customer involvement operates through an additive accountability-and-legitimacy pathway consistent with service-dominant logic’s recognition that customers can function as direct value co-creators, acting independently of firm capability mechanisms.

5.3. Theoretical implications.

5.3.1. Contributions to RBV, dynamic capabilities theory, and open-innovation theory.

Regarding RBV, this study demonstrates that GAI adoption’s competitive value resides in the emergent HAIC routines that are tacit, path-dependent, and inimitable—satisfying VRIN conditions in ways that raw tool access cannot. This finding shifts the RBV discourse on AI from technology possession to capability-integration quality, with implications for knowledge-intensive service industries. Regarding dynamic capabilities theory, this study determined that OIO is the dominant DCST predictor (β = 0.451). This finding challenges the implicit assumption that technological investment drives capability development, asserting instead that dynamic capabilities emerge from managerial orchestration of knowledge flows (including externally sourced flows) rather than technology per se.

The consistent partial mediation suggests that expanded DCT models are needed to account for both direct AI-performance effects and mediated capability channels. This study repositions OIO from a peripheral moderating role to a co-primary antecedent, whose total effects on SP and IP are comparable to those of GAI adoption. This approach delivers the theoretical AI–open-innovation integration called for but not achieved in the tourism management literature.

5.3.2. HAIC as a novel mediating mechanism and triple-mediation structure.

Another distinctive theoretical contribution is operationalizing HAIC as a psychometrically validated first-order mediating construct (outer loadings: 0.744–0.863; AVE = 0.677; CR = 0.926; HTMT = 0.605). This finding advances Qu and Kim (2025) qualitative propositions into confirmed quantitative hypotheses, establishing a measurement precedent for treating human–AI interaction quality as an independent organizational capability across service management disciplines. The triple-mediation architecture demonstrates that the AI-to-performance pathway is multi-channeled, with distinct functional specializations. HAIC most strongly mediates innovation performance (GAI → HAIC → IP: VAF = 43.9%), reflecting its role as a generative ideation and prototyping capability.

DCST most strongly mediates sustainability performance when OIO is the antecedent (OIO → DCST → SP: VAF = 40.6%), reflecting regulation and reconfiguration capability. DSCI occupies a balanced intermediate position across both outcomes (VAF: 30.0–36.0%), reflecting its operational coordination function that simultaneously reduces waste and enables service customization. This capability specialization asymmetry has direct implications for sequencing AI investment and OIO activities across the sustainability and innovation performance domains.

5.4. Comparison with prior literature.

This study’s findings confirm HAIC as a central mediator of AI performance, corroborating Qu and Kim (2025) qualitative propositions. They also replicate the AI–dynamic capabilities–performance chain in a new national context (Thailand) and across industries (microtourism vs. manufacturing), demonstrating cross-contextual robustness. The DCST VAF values (GAI → DCST → SP: 26.8%; OIO → DCST → SP: 40.6%) are moderate in magnitude. This disparity is attributable to the present model’s two additional competing mediators, which absorb portions of the dynamic capability’s indirect effect. The smaller GAI → DCST coefficient (β = 0.297) is consistent with microenterprise resource constraints attenuating AI’s capability-building effect. OIO compensates through its own dominant DCST pathway (β = 0.451).

The DSCI mediation findings (VAF: 30.0–36.0%) extend Wang and Zhang (2025) evidence from Chinese tourism SMEs to the context of Southeast Asian microenterprises, suggesting cross-national generalizability of the GAI–DSCI–performance mechanism.

This study’s principal divergence is the nonsignificant CIST moderation, which contradicts Wang and Zhang (2025) finding. This discrepancy most plausibly reflects differences in sample composition. Wang and Zhang’s sample included SMEs with established digital infrastructure. In contrast, the present study’s sample is restricted to microenterprises with predominantly informal digital ecosystems. Peiró-Signes et al. (2026) provided evidence from Colombian manufacturing SMEs, showing that collaboration variables have weaker predictive power than internal resources in low-innovation ecosystems. They provided cross-national theoretical alignment: in under-institutionalized cocreation ecosystems, customer involvement may primarily contribute to performance through independent information content (rather than capability amplification).

The divergence reinforces this study’s context-specific contribution and cautions against uncritically transferring moderation findings across national and organizational scales.

5.5. Alternative explanations.

Three alternative interpretations warrant consideration. First, the strong OIO effects may partly reflect self-selection. For instance, operators scoring higher on OIO may also be earlier AI adopters, more frequent seekers of sustainability certifications, and more platform-networked, generating correlated scores across mediators and outcomes that a cross-sectional design cannot disentangle from genuine causal pathways. Second, despite Harman’s single-factor variance of 38.48% and full collinearity VIF below 3.3, single-informant perceptual data may inflate path coefficients through halo effects. For example, multi-informant designs or objective performance data would strengthen causal inference.

Third, the HAIC–IP relationship could reflect the perceived ease and usefulness of AI tools in the innovation workflow—consistent with the principles of the technology acceptance model —rather than a distinct organizational capability development pathway. Distinguishing among these interpretations requires temporal or experimental designs that are not available in the current cross-sectional approach. Fourth, the study does not consider whether its structural model may be partially artifactual due to response-set homogeneity. Of the respondents, 72% were owners/cofounders or senior executives— individuals most likely to hold optimistic views of their firms’ AI adoption and open-innovation engagement.

Even after accounting for the common method variance concern already addressed, estimates derived from this strategically self-selected informant profile may systematically overstate the adoption-to-performance linkage relative to a more diverse respondent pool. Fifth, and most fundamentally, the cross-sectional design cannot eliminate reverse causality. Firms with already high sustainability performance may have greater slack resources and institutional legitimacy to invest in OIO and GAI, rather than the reverse. The near-parity of OIO and GAI total effects—consistent with either forward or reverse causal ordering—makes this alternative particularly credible. This possibility represents the primary limitation on the strength of causal claims that can be derived from this study’s evidence. Longitudinal or experimental designs are needed to adjudicate between these competing orderings.

None of these alternative explanations invalidates the overall pattern of findings; however, each sets a boundary on the strength of causal claims that can be responsibly made on the basis of the present evidence.

A further consideration concerns the explanatory power of the two capability mediators. The model accounts for 35.5% of the variance in HAIC and 34.3% in DSCI—values in the moderate range —but leaves most each mediator’s variance unexplained. This is an expected and theoretically informative consequence of the model’s deliberate parsimony. HAIC and DSCI were specified as functions of only the two focal strategic antecedents (GAI adoption and OIO), rather than of an exhaustive determinant set, to isolate the AI–open-innovation mechanism that constitutes this study’s contribution. HAIC is the study’s central novelty, and therefore the unmodeled variance warrants explicit attention. Human–AI collaboration quality is a multi-determined capability that also depends on factors operating below the firm-strategy level captured here.

At the individual level, owner-manager and employee digital literacy, AI self-efficacy, prior technology experience, and trust propensity plausibly shape the effectiveness of integrating human and AI inputs. At the organizational level, training investment, learning capability, leadership championing, and financial slack condition the formation of collaborative routines. At the technological level, the usability and reliability of the specific tools adopted—including the maturity of Thai-language generative models and their integration with incumbent workflows such as LINE OA—constrain the achievable quality of collaboration. Institutional supports such as DASTA programs, SHA Plus participation, and OSMEP digitalization grants may further supply training and peer-network resources that our model does not represent.

The same logic applies to DSCI, whose realization also depends on partner readiness and platform-ecosystem constraints. These unmodeled determinants are not deficiencies of the tested theory but signposts for its extension. Section 5.8 identifies the antecedents and boundary conditions of HAIC formation as a priority for future research.

5.6. Limitations.

Five limitations circumscribe the confidence and generalizability of the findings. First, the cross-sectional design precludes causal inference; the direction of all relationships is inferred from theory rather than empirically verified. This limitation most directly constrains the interpretation of the OIO → DCST dominant pathway (β = 0.451); the possibility that high-DCST firms are more disposed toward OIO investment (a reverse causal ordering indistinguishable from the proposed model in cross-sectional data) cannot be dismissed. Second, single-informant self-reported data introduce social desirability and halo biases that CMB screening mitigates but cannot eliminate. In contrast, the absence of objective performance data limits the external validity of SP and IP outcomes.

This limitation most directly affects the HAIC → IP pathway (VAF = 43.9%), where perceived AI usefulness and actual collaboration capability quality are difficult to disentangle in a single-owner-operator respondent without multi-informant or observational data. Third, purposive multistage sampling via industry associations is likely to overrepresent more formalized and digitally engaged microoperators relative to the broader informal microtourism population. This limitation most directly constrains the OIO findings. For example, association members are disproportionately likely to have established OIO practices; thus, OIO’s dominant path coefficients may overstate the magnitude of the effect in the full population of Thai microtourism enterprises.

Fourth, the CIST construct’s operationalization of digital platform-mediated cocreation underweights community-based and face-to-face cocreation forms, which are culturally salient in Thai community tourism. This measurement limitation most directly constrains the interpretation of the nonsignificant CIST moderation. The hypothesized amplification effect may exist; however, its influence may be undetectable with a scale that captures only one dimension of the theorized cocreation process. Fifth, this study is geographically bounded to Thailand and temporally bounded to the post-pandemic recovery period (2025–2026). Thailand’s distinctive institutional environment includes the SHA Plus certification ecosystem, LINE OA infrastructure, and BCG Economy policy framework.

Thus, the results may not transfer to other Southeast Asian emerging economies with different digital ecosystems, governance structures, or tourism product profiles. This geographic-temporal boundary most directly affects the generalizability of the DSCI mediation findings, which are embedded in a platform ecosystem (e.g. LINE Groups, Agoda, and Booking.com) specific to the Thai context. Researchers applying the framework in other contexts should treat such ecosystem constructs as institutional moderators that require local adaptation rather than as fixed boundary conditions.

5.7. Practical implications.

5.7.1. Implications for managers.

The GAI → HAIC → performance pathway (VAF: 38.1–43.9%) implies that investing in AI tools without improving the quality of human–AI collaboration will not generate sustained competitive advantage. Therefore, managers should prioritize structured capability-building programs that develop operators’ ability to evaluate, iterate on, and adapt AI-generated outputs to multilingual guest communications,

SHA Plus documentation, and itinerary personalization. The LINE OA ecosystem provides a natural, progressive integration point. The OIO → DCST dominance (β = 0.451) suggests that open-innovation partnerships should be considered core capability investments, indicating that managers should participate in DASTA networks, co-design eco-experiences with NGOs, and leverage booking-platform demand data as externalized forecasting resources. Finally, CIST should be pursued as a direct SP driver (β = 0.208) by systematically integrating customer digital feedback to create accountability for sustainability commitments.

5.7.2. Implications for policymakers.

Thailand’s technology support programs require open-innovation facilitation infrastructure, such as managed knowledge-exchange platforms that connect microtourism operators with destination management organizations, certification bodies, and community partners, through DASTA’s area-based model. AI literacy programs should target HAIC development as the explicit capability goal rather than adoption rate metrics, shifting program monitoring toward assessments of collaboration quality. The Office of SME Promotion’s (OSMEP’s) digitalization grants should fund human–AI capability development, co-delivered with technology providers and tourism associations, rather than providing tool-access subsidies.

Regionally, DASTA’s area-based management model should be leveraged as a platform for OIO-facilitation hubs that aggregate market intelligence, sustainability expertise, and AI tool-training resources, accessible to the smallest operators across all six Thai geographic regions, thereby reducing the capability gap between formalized association members and the broader informal microtourism population.

5.7.3. Implications for technology providers.

Technology providers should embed HAIC development into product design through guided workflow templates that guide operators through iterative AI output evaluation cycles rather than delivering raw outputs that require unstructured judgment. LINE OA and Facebook Messenger integrations can progressively build collaboration competence in familiar interfaces to help reduce adoption barriers while driving performance returns that justify subscription costs for microoperators. This study’s DSCI findings (VAF: 30.0–36.0%) identified supply chain coordination functionality—AI-assisted partner communication templates, non-technical demand forecasting summaries, and SHA Plus documentation automation—as a high-value priority underserved by existing enterprise-scale platforms.

Booking platforms (e.g. Agoda, Airbnb, and Booking.com) are well-positioned to convert market intelligence into open-innovation facilitation features (such as anonymized sustainability benchmarking and cocreation protocol tools) that simultaneously strengthen operators’ DCST, improve listing quality, and enhance platform differentiation.

5.8. Directions for future research.

Six research directions emerge from this study’s limitations and unanticipated findings. First, longitudinal designs are needed to establish causal sequences, particularly whether HAIC development precedes or follows AI tool adoption, and to inform implications for training program sequencing (linked to limitation one, concerning cross-sectional design). Second, the CIST construct requires refinement. Future instruments should include items specifically asking about the integration of AI-analyzed customer data into sustainability specifications, formal customer participation in service redesign, and cocreation outcomes documented in SHA Plus. Such approaches may reveal conditional moderation effects that the present broad-gauge operationalization obscured (see limitation 4 regarding CIST measurement).

Third, a comparative multi-country study encompassing Thai, Vietnamese, and Indonesian microtourism SMEs could test whether CIST moderation emerges in markets with more mature cocreation infrastructure (see limitation five regarding geographic boundary).

Specifically, cross-national validation is urgently needed for three construct-level relationships. (i) OIO → DCST dominance (β = 0.451) may reflect Thailand’s distinctive SHA Plus certification environment rather than a universal open-innovation-capability linkage. (ii) the nonsignificant CIST moderation may emerge in contexts with more institutionalized customer co-design infrastructure, such as Vietnam’s community-based tourism governance or Indonesia’s Kampung Wisata networks. (iii) The HAIC mediation magnitude (VAF = 43.9% for GAI → HAIC → IP) may attenuate in markets with lower generative AI penetration or where LINE OA’s familiar adoption pathway is absent. Establishing these boundary conditions would determine whether Wang and Zhang (2025) amplification effect and this study’s OIO dominance finding are context-specific or cross-nationally robust.

Fourth, future designs should triangulate perceptual SP and IP measures with objective data, such as energy audits, waste records, SHA Plus certification scores, and booking-platform rating trajectories. This approach can strengthen causal claims and demonstrate practical significance for operators, investors, and policymakers (with respect to limitation two, on single-informant data). Fifth, the emergent role of Thai-language GAI applications developed specifically for the tourism sector warrants dedicated investigation. As these platforms mature and develop more sophisticated AI-mediated customer–operator cocreation workflows, the hypothesized (but not observed) CIST moderation effects may become empirically detectable, providing a natural experiment to test the boundary conditions identified in Section 5.5.

Sixth, future research should investigate the boundary conditions on HAIC formation, including firm-level and institutional moderators that determine when GAI adoption translates into strong versus weak human–AI collaboration quality. This study establishes HAIC as the primary innovation performance mediator (VAF = 43.9%) and demonstrates that GAI adoption is HAIC’s strongest antecedent (β = 0.427); however, the conditions under which this pathway is activated or attenuated remain unexamined. Candidate moderators include owner-operator digital literacy, business age, SHA Plus certification status, and DASTA membership. Identifying these boundary conditions could explain the residual direct effects of GAI on SP and IP identified in the supplementary mediation model, thereby providing practitioners with actionable targeting criteria for HAIC development programs.

6. Conclusion.

Thailand’s microtourism sector stands at a pivotal juncture; more than 99% of operators employ fewer than 50 staff members, and COVID-19 recovery depends on precisely the digitally agile, sustainability-oriented transformation that resource constraints make difficult to achieve. This study addressed four investigative objectives. First, we examine the direct effects of GAI adoption and OIO on three capability-building mediators (HAIC, DSCI, and DCST). Second, we assess the roles of those mediators in transmitting effects to SP and IP. Third, we are evaluating CIST’s moderating role. Fourth, we derive evidence-based recommendations for managers, policymakers, and technology providers. All four objectives were achieved. Twelve of 14 hypotheses were supported, with the model explaining R2 = 0.478 (SP) and R2 = 0.427 (IP).

The most consequential finding is that OIO and GAI adoption produce near-identical total effects on both performance outcomes (OIO: SP 0.239, and IP 0.272; GAI: SP 0.239, and IP 0.280). This result repositions OIO from the peripheral moderating role assigned in prior frameworks to an indispensable co-driver of sustainable tourism value creation. The dominant pathway, GAI → HAIC → IP (VAF = 43.9%), confirms Qu and Kim (2025) proposition that augmented human–AI collaboration capability is the central value-creating mechanism in AI-enabled microenterprise ecosystems. Their proposition is validated for the first time through a psychometrically confirmed first-order mediating construct in quantitative tourism research.

This pattern is theoretically coherent in the way the three frameworks converge on a single integrated explanation. RBV explains why HAIC generates sustained competitive advantage: its tacit, path-dependent quality satisfies the VRIN criteria in ways that raw tool access cannot. Dynamic capabilities theory explains why OIO dominates the sustainability pathway: dynamic capabilities are built through the orchestration of external knowledge rather than technology per se. Open-innovation theory explains why all three mediators operate simultaneously and only partially. OIO’s knowledge flows are inherently multi-pathway. Together, the three theories predict precisely the results of this study—partial mediation across parallel channels, with HAIC governing innovation returns and DCST, via OIO, governing sustainability returns.

Operators and policymakers who treat AI adoption as technology procurement rather than capability-building will systematically underinvest in the human dimension that generates the majority of AI’s performance return.

The nonsignificant CIST moderation, while unanticipated, is itself a theoretical contribution. Customer involvement has a significant direct effect on sustainability performance (β = 0.208); however, the null moderation establishes a boundary condition on service-dominant logic’s cocreation amplification proposition. It holds in high-institutionalization contexts; however, it does not generalize to digitally informal ecosystems (e.g. LINE OA, Facebook, community networks) characterizing MSME-dominant emerging-economy tourism markets globally.

Overall, this study’s evidence carries a clear strategic message: the path to recovery runs through the simultaneous cultivation of high-quality human–AI collaboration and an open-innovation orientation. This finding extends beyond Thailand and is relevant for every MSME-dominant tourism market confronting the dual imperatives of digital transformation and sustainability transition. The validated HAIC construct and OIO–GAI co-driver architecture provide the international tourism management community with a theoretically grounded, empirically supported framework for studying how digitally constrained microenterprises in Southeast Asia, East Africa, Latin America, and comparable markets can simultaneously pursue sustainability and innovation outcomes. This outcome fills a gap in prior research, which has disproportionately focused on larger enterprises and high-income contexts.

Operators who build the deepest human–AI collaboration routines and the broadest open knowledge relationships will not merely survive their tourism recoveries; they will globally define the next generation of sustainable microenterprise tourism.

Acknowledgments.

Not applicable.

Ethics statement

This study was approved by the Human Research Ethics Committee of Silpakorn University (approval no. REC 68.1219-234-11346); certificate no. COE 68.1224-123, dated 24 December 2025. Written informed consent was obtained from all participants in the study via a mandatory online consent form before data collection.

Authors’ contributions

CRediT: Ratchamongkhon Thonglor: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Validation, Writing – original draft, Writing – review & editing; Noptanit Chotisarn: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Validation, Writing – original draft; Thadathibesra Phuthong: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. All authors approved the final manuscript.

Disclosure statement

The authors report no potential conflicts of interest.

About the authors

Ratchamongkhon Thonglor is a lecturer in the Department of Hotel Management, Faculty of Management Science, Silpakorn University, Thailand. He graduated from Silpakorn University in Thailand with a Doctor of Philosophy in Tourism, Hotel, and Event Management. His current research interests include destination brand equity, tourism development, and community participation in tourism area development.

Noptanit Chotisarn is an assistant professor in the Department of Management Information Systems at the Thammasat Business School, Thammasat University, Thailand. He graduated from Zhejiang University, China, with a Doctor of Philosophy in Software Engineering. His research interests include AI engineering, data visualization, and machine learning.

Thadathibesra Phuthong is an assistant professor in the Department of Logistics Management, Faculty of Management Science, Silpakorn University, Thailand. He graduated from Chulalongkorn University, Thailand, with a Doctor of Philosophy Program in Technopreneurship and Innovation Management (Interdisciplinary Program). His current research interests include technology, innovation, and management.

Data availability statement

The data supporting the findings of this study are available from the corresponding author upon reasonable request.