Human-centered AI adoption for small and medium-sized enterprise competitiveness: a methodological perspective on digital capabilities and organizational adaptation

1 Introduction

Small and medium-sized enterprises (SMEs) account for most firms and much of employment in almost every economy, so their productivity has an outsized effect on aggregate competitiveness (OECD, 2021), and artificial intelligence (AI) is widely presented as a way for them to close the productivity gap with larger competitors (Brynjolfsson and Mitchell, 2017; Davenport and Ronanki, 2018). Yet adoption is uneven, and a recurring pattern is that firms acquire a tool, run a pilot, and fail to embed it in daily operations (Ayinaddis, 2025; Baabdullah et al., 2021; Schwaeke et al., 2024; Sharma et al., 2024). The gap between acquisition and embedding is where competitiveness is won or lost.

Existing scholarship addresses parts of this problem. Acceptance models explain the decision to adopt (Davis, 1989; Venkatesh et al., 2003); the technology-organization-environment framework widens the lens to firm and context factors (Tornatzky and Fleischer, 1990); and capability-based work shows that internal digital and dynamic capabilities predict AI adoption and performance more strongly than external support (Arroyabe et al., 2024; Cimino et al., 2025; Enholm et al., 2022; Mikalef and Gupta, 2021). Three gaps motivate this article: capability accounts rarely specify the human-centered design conditions under which employees keep using a tool long enough for capability to form; the organizational adaptation through which capability becomes performance is usually inferred rather than observed, so the mechanism is asserted; and the literatures on responsible AI, work design and algorithmic control, which address precisely those human conditions, are seldom joined to the capability account (Bankins et al., 2024; Kellogg et al., 2020; Mikalef et al., 2022; Parker and Grote, 2022). Table 1, Panel A, sets the gaps out construct by construct.

Table 1. Positioning against existing accounts (panel A), propositions with candidate indicators and test designs (panel B), and the methodological commitments they imply (panel C).

Objective 1 concerns how human-centered design conditions the processes through which SMEs convert AI access into capability: a human-centered foundation raises the likelihood that sensing, seizing and reconfiguring occur (P1.1); where it is absent, adoption stalls at symbolic use, with acquisition recorded but routines unchanged (P1.2); and the appropriate human-AI configuration is contingent rather than uniformly augmentative (P1.3). Objective 2 concerns how organizational adaptation converts capability into competitiveness: reconfiguring produces structural, cultural and human-capital adaptation as the realized state downstream of it, with an imperfect correspondence between the two (P2.1); adaptation carries a substantial part of the association between AI capability and competitiveness, so controlling for it attenuates, without necessarily eliminating, the direct association between tool ownership and performance (P2.2); and the competitiveness dimensions realize in a characteristic order, efficiency earliest and realized resilience latest (P2.3).

These propositions cannot be validated within this article, since the Perspective format excludes original data. Each is instead paired in Table 1, Panel B, with candidate indicators and a design capable of testing it. Figure 1 and Table 1 are the authors’ own elaboration and report no original or unpublished data.

2 What existing accounts explain, and what they leave open

Four bodies of work bear directly on AI in small firms, and the framework complements rather than displaces them: acceptance models, which predict willingness to use but not what an organization must reorganize once the system arrives; the technology-organization-environment framework, frequently applied as a static checklist of enablers and barriers; capability-based work, which shows that internal digital and dynamic capabilities outweigh external support (Arroyabe et al., 2024; Cimino et al., 2025); and human-centered and responsible AI research, which specifies what a workable human-AI configuration looks like (Mikalef et al., 2022; Raisch and Krakowski, 2021; Shneiderman, 2020). What remains open is the junction between them: capability accounts establish that dynamic capabilities matter without specifying the design conditions that let capability form where adoption is visible to every employee, and human-centered research specifies those conditions without connecting them to capability or performance. The contribution claimed here is that junction and its methodological consequences, not the constructs themselves. Panel A of Table 1 sets out, account by account, what each explains, what it leaves open, what the framework adds and which elements are recombinations.

3 How the framework was developed

The framework is a conceptual synthesis, not a systematic review. No protocol was registered and no comprehensiveness is claimed; the selection should be read as interpretive. Literature was identified through Scopus, Web of Science and Google Scholar, supplemented by citation tracing from the anchor works cited here and by OECD and World Bank policy sources, covering 2015–2026 for the AI-specific streams, with earlier foundational sources retained where a construct originates in them (Davis, 1989; Teece, 2007; Tornatzky and Fleischer, 1990). Four streams were retained because each answers a distinct question the phenomenon raises, and frameworks were treated as overlapping where their constructs shared referents rather than labels.

Two retention rules governed what entered the framework. A human-centered dimension was retained as a principle only where it recurred across the human-centered and responsible-AI sources consulted and where the SME literature offered a stated mechanism linking it to continued use at small-firm scale; dimensions meeting the first test but not the second enter as safeguards (Section 4.2). The three adaptation modes were derived from the objects reconfiguring acts upon in the digital transformation literature (Verhoef et al., 2021; Vial, 2019), and we claim they are analytically distinct rather than exhaustive. Supplementary Material S1 reports the databases, concept combinations, stream selection and overlap-handling rules in full. Claims elsewhere about the balance of evidence are our reading of it rather than established fact.

4 The human-centered AI adoption framework

4.1 Construct vocabulary

Definitions come first, because the propositions depend on keeping several loosely used terms apart. Dynamic capabilities are higher-order capacities to renew a firm's resource base, with sensing, seizing and reconfiguring as their microfoundations (Teece, 2007); digital capabilities are the technical and data assets the firm holds; and AI capability is the organizational capacity to deploy AI, distinct from the software itself (Mikalef and Gupta, 2021).

Four further terms are not equivalent explanatory variables and are not used interchangeably below: tool ownership is possession of, or contracted access to, an AI application; AI use is any recorded use of it; embedded use is use integrated into the routines through which the relevant work is done, evidenced by the behavioural markers in Figure 1; and AI capability is the organizational capacity defined above. P1.2 concerns the gap between ownership and embedded use, P2.2 the association between ownership and performance once adaptation is accounted for. Digital and AI resources are the endowment on which the microfoundations operate, not themselves dynamic capabilities. Figure 1 arranges these constructs into a chain of propositions rather than a causal chain: no relationship shown has been estimated here.

4.2 The human-centered foundation and its declared scope

Four principles sit at the foundation: meaningful human control, transparency and explainability, augmentation where appropriate, and trust and accountability. They were retained under the rule in Section 3, because each has a demonstrable mechanism at SME scale, where a tool that staff cannot understand or override is abandoned rather than contested.

The resulting scope is stated plainly rather than implied: the framework addresses a restricted subset of human-centered AI, concerned with control, comprehension, appropriate use and accountability, and the four principles are not offered as a comprehensive human-centered foundation. A wider set of requirements—fairness and non-discrimination, privacy and employee data governance, worker participation and contestability, safety and error escalation, limits on surveillance and algorithmic control, inclusion and accessibility, and the distribution of benefits and risks—is nonetheless incorporated into the framework as a cross-cutting worker and stakeholder safeguards dimension (Figure 1). These are normative and sociotechnical design requirements, and stating them as conditions on responsible deployment does not depend on prior SME-specific validation (Alfrink et al., 2023; Kellogg et al., 2020; Mikalef et al., 2022; Shneiderman, 2020). What the SME evidence does not support is a specification of their mechanisms at small-firm scale, so they are included as requirements without any claim that a subdimension has been validated there. Contestability illustrates the distinction: that a worker should be able to challenge a consequential automated output is a design requirement (Alfrink et al., 2023), whereas how contestation affects continued use in a firm with no compliance function is an open question, and the framework's principal extension.

Human welfare is not treated instrumentally. Worker-level outcomes—autonomy, workload, skill use and job quality—appear in Figure 1 as ends in their own right, capable of diverging from firm competitiveness rather than tracking it (Bankins et al., 2024; Parker and Grote, 2022). Customer and community effects are neither represented in Figure 1 nor operationalized in Table 1, so the outcome domain claimed here is limited to workers and firms; Section 9 treats their inclusion as the framework's most consequential extension.

Augmentation is a defensible default in many SME applications, but the appropriate configuration depends on task criticality, cost of errors, employee expertise, regulatory constraint and system type, and full automation is appropriate for some routine, low-stakes, high-volume tasks (P1.3).

4.3 Capabilities and adaptation

Reconfiguring and organizational adaptation are closely related but are treated here as distinct constructs, on a single basis maintained throughout the text, Figure 1, the propositions and Table 1. Reconfiguring is an activity: the work of redesigning roles, procedures and decision rights around a tool. Adaptation is a state: the realized configuration of formal arrangements, shared norms and skills that the activity produces. The distinction is the familiar one between a capacity exercised and a condition achieved, and it is empirically consequential because the correspondence is imperfect—firms attempt redesigns that revert, procedures are rewritten but not followed, training is delivered but roles are unchanged. That residual variance is informative rather than nuisance, since it is where symbolic use is located (P1.2). Adaptation is therefore positioned downstream of reconfiguring and never as an antecedent of it, which resolves the overlap in the original formulation while keeping the two measurable apart (P2.1). Because the non-identity is a claim and not an assumption, Panel B specifies a measurement model that tests it.

4.4 Competitiveness and its temporal order

Efficiency, responsiveness, innovation and resilience are distinct dimensions of competitiveness rather than interchangeable indicators, and they are expected to realize in a characteristic order rather than at fixed intervals: efficiency earliest, because it follows from task-level substitution once use is embedded; responsiveness and innovation later, because they require adapted routines and skills to be exercised on new problems; and realized resilience only where the firm is exposed to a disruption. The ordering is stated relatively and no numerical horizons are assigned, because the available evidence does not support particular intervals. Resilience requires one further distinction: resilience capacity—redundancy, substitutability of suppliers and staff, documented response routines—can be assessed before any disruption occurs, whereas realized resilience is observable only across one. AI adoption may also change what confers advantage rather than simply improving performance on existing terms (Krakowski et al., 2023).

5 Boundary conditions and alternative pathways

The framework is not universal. Firm size, sector, system type, task criticality, build-versus-buy, data maturity, owner centralization, access to external advisors, regulatory intensity and national digital infrastructure all condition the pathway, and the human-centered foundation binds hardest where decisions are consequential and least where errors are cheap and reversible. Supplementary Material S2 sets out how each condition is expected to operate.

Alternative orderings deserve equal attention. Prior adaptation may enable sensing and seizing rather than only follow them, since firms accustomed to revising routines recognize opportunities more readily. Competitiveness gains may generate slack that reinforces every capability dimension. And where adoption is imposed without participation, adaptation may occur as compliance or resistance rather than embedding, producing recorded use with no capability behind it (Kellogg et al., 2020). Figure 1 treats the stated ordering as the default on two grounds: capability cannot form from use that has already stopped, whereas capability can form without the reverse ever being observed; and adaptation, defined as the state produced by reconfiguring, cannot precede it within a single cycle, so a prior-adaptation effect is a second-cycle effect carried by the feedback path. These are arguments for a default, not evidence for it, and the alternatives remain the specifications against which the framework should be tested.

6 Methodological implications

If competitiveness is produced by adaptation, methods must capture change, configuration and context rather than a single cross-section of stated intention. Four commitments follow; Panel C of Table 1 states each with its operational requirement.

The framework spans levels: sensing and seizing are firm-level acts, reconfiguring and adaptation are observable at team and role level, and the principles and worker outcomes are experienced individually, so the unit of analysis must be specified rather than assumed. Behavioural and archival traces are preferable to self-report, because only they separate symbolic from embedded use; the indicators in Table 1 are candidates rather than validated instruments and require psychometric development, including the competing measurement models that test rather than presuppose the separation of reconfiguring from adaptation. The observation window must match the dimension under study: the three-wave, eighteen-to-twenty-four-month panel that suffices for embedding and efficiency effects is inadequate for the later-emerging dimensions of P2.3, which require a longer window and, for realized resilience, exposure to a disruption. And adoption is self-selected, so causal claims about the adaptation-competitiveness link require designs that address selection and reverse causality rather than a single-informant survey.

7 Emerging and transition economies

The setting is not a matter of illustration, and specific relationships should be expected to differ. The mechanisms below are context-sensitive expectations for testing: where the underlying condition is documented we cite it, and where it is not, the claim is marked as a hypothesis rather than presented as an established regional characteristic.

Supply-side dependence. Public cloud AI compute is concentrated in a small number of jurisdictions, and most countries consume AI services hosted elsewhere rather than hosting the infrastructure on which they run (Lehdonvirta et al., 2024). Where services are supplied predominantly by foreign platforms, we expect reconfiguring to be constrained by what those platforms permit. The supply structure is documented; the firm-level constraint is our expectation and is unmeasured in the region.

Language. Model performance in lower-resource languages remains materially weaker than in English across generation, comprehension and translation tasks (Ahuja et al., 2023), and Uzbek is a documented low-resource case in which basic annotated resources have only recently been constructed (Kuriyozov et al., 2022). We expect the transparency principle to bind harder in such settings, because staff cannot verify outputs they cannot fully read, and P1.2 to hold more strongly. The moderating claim is ours and is untested.

Territorial concentration of capability. AI adoption is diffusing unevenly within countries as well as between them, concentrating in the most innovative regions, in larger firms and in knowledge-intensive services (Kergroach and Héritier, 2025). In Uzbekistan, the digital skills of managers and employees are identified as a principal constraint on SME digitalization, with support delivered through a scattered set of agencies (OECD, 2023). We therefore expect regional firms to depend on external advisors, which makes the availability of universities and public support agencies a boundary condition rather than a background factor. That AI specialists in particular are concentrated in capital cities is a plausible extension of the documented territorial pattern and is stated here as a hypothesis.

Records and measurement. Where business records are incomplete, we expect sensing to depend more on peer networks and vendor demonstrations than on internal data, so that the sensing-seizing link weakens and intermediaries matter more. We have no regional evidence on record completeness in SMEs specifically; the firm-level indicators of the World Bank Enterprise Surveys, which cover Uzbekistan, are the appropriate instrument for testing it (World Bank, 2024).

Regulation. It would be inaccurate to describe safeguards as simply absent. Uzbekistan regulates personal data through a dedicated statute, amended in 2021 to require that citizens’ personal data be processed on infrastructure located in the country (Law No. ZRU-547, 2019), and approved a national AI development strategy to 2030 in October 2024 (Resolution No. PP-358, 2024). The expectation concerns orientation rather than absence: that framework addresses data localization and national AI capacity and, on our reading, does not yet specify obligations concerning algorithmic management, employee data within the employment relationship, or the contestation of automated decisions. Where that holds, the safeguards of Section 4.2 are least likely to be supplied by law and most likely to require design, and this reading would be revised by evidence of implementing regulation to the contrary.

For Central Asian economies, including Uzbekistan, these expectations imply that policy support should fund capability building, co-design, applied training and advisory capacity rather than technology procurement alone. Comparative firm-level evidence for the region remains scarce, and generating it is among the more useful contributions this agenda could make.

8 Practical implications

Three proposals are concrete enough to act on and to measure. Firms should conduct a readiness assessment with staff rather than for them, treating the proportion of employees able to explain what a tool does as a deployment gate; define per-tool success criteria before piloting, including a stated override rate at which the tool is reviewed, and record overrides; and schedule an explicit reconfiguring step, revising written procedures and role descriptions within a defined period after go-live, treating the absence of such revision as the diagnostic that P1.2 describes.

Which measure binds depends on the setting. In a small manufacturer deploying predictive maintenance the override measure binds first, because unexplained recommendations are disregarded and the tool falls out of use before adaptation occurs; in a professional-services firm using generative drafting the procedure-revision measure binds instead, because output is easy to adopt informally and hard to govern. Owner-managers can apply the measures as deployment gates, support agencies as conditions of assistance, and researchers as behavioural indicators of embedded use.

9 Limitations and future directions

Four limitations bound these claims, and each implies a specific remedy. The framework is conceptual and untested, so its propositions are candidates for validation rather than results; the remedy is the multi-wave design in Panel B for P2.2, estimated against the alternative orderings of Section 5. Its human-centered layer is deliberately restricted, and the safeguards of Section 4.2 are stated as requirements without specified mechanisms; the remedy is to specify those mechanisms at SME scale. Its constructs derive from literature developed largely in high-income settings, so applicability to informal and microenterprise contexts is open; the remedy is comparative work in transition economies, where Section 7 states the predicted departures and marks which are hypotheses. And the indicators in Table 1 are unvalidated; the remedy is instrument development.

The indicator set is deliberately not confined to firm performance: worker outcomes are measured directly rather than inferred from it. Customer and community effects are not represented in the present framework, and we regard their inclusion, with the corresponding revision of Figure 1 and Table 1, as the change most likely to move it toward a holistic account of what AI adoption does.

10 Conclusion

The contribution of this Perspective is a junction rather than a new construct: it connects the human-centered design conditions under which employees continue using AI to the capability and adaptation processes through which use becomes competitiveness, and specifies what would have to be observed for that connection to hold. An AI purchase is the beginning of a change project rather than the end of a decision, and studies should follow firms past the point of purchase. Whether the relationships survive testing is an empirical question.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fhumd.2026.1939008/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Digvijay Pandey, Department of Technical Education Uttar Pradesh, India

Reviewed by: Mircea Constantin Scheau, University of Craiova, Romania

Alexander Sánchez-Rodríguez, Equinox University of Technology, Ecuador