Why Small Firms Buy AI and Then Never Use It: A New Framework Explains the Gap

Small and medium-sized enterprises account for most firms and most employment in nearly every economy, which means their productivity quietly sets the ceiling on national competitiveness. Artificial intelligence is repeatedly presented as the great equalizer, a way for a twenty-person workshop to close the gap with corporate giants. Yet the pattern researchers keep observing is stubbornly familiar: a firm buys an AI tool, runs an enthusiastic pilot, and then fails to embed it in daily operations. A new Perspective article published in Frontiers in Human Dynamics argues that this gap between acquisition and embedding is precisely where competitiveness is won or lost, and that the missing ingredient is not technology at all but human-centered design.

The article, authored by Layli Navruz-zoda of Bukhara State University and colleagues at institutions across Uzbekistan, does not present new experimental data. Instead, it offers something arguably more consequential for the field: a carefully constructed conceptual framework, the Human-Centered AI Adoption Framework, that stitches together two research literatures which have largely evolved in isolation. One literature, rooted in the dynamic capabilities tradition, shows that internal digital and dynamic capabilities predict AI adoption and performance more strongly than external support. The other, drawn from human-centered and responsible AI research, specifies what a workable human-AI configuration looks like. What has been missing, the authors contend, is the junction between them, along with the methodological consequences that follow.

The framework rests on a deliberately restricted set of four human-centered principles: meaningful human control, transparency and explainability, augmentation where appropriate, and trust and accountability. The authors are explicit that this is a subset, not a comprehensive charter. Each principle earned its place through a retention rule requiring both recurrence across the human-centered AI literature and a stated mechanism linking it to continued use at small-firm scale. The logic is pragmatic: in a small enterprise, a tool that staff cannot understand or override is simply abandoned rather than formally contested. Broader requirements such as fairness, privacy, worker participation, safety, limits on surveillance and accessibility enter the framework as a cross-cutting safeguards dimension, acknowledged as normative design requirements whose SME-scale mechanisms remain unspecified.

At the technical heart of the framework lies a distinction that sounds subtle but carries real analytical weight: the difference between reconfiguring and adaptation. Reconfiguring is an activity, the work of redesigning roles, procedures and decision rights around a new tool. Adaptation is a state, the realized configuration of formal arrangements, shared norms and skills that the activity produces. The correspondence between the two is imperfect. Firms attempt redesigns that revert, procedures get rewritten but not followed, training is delivered while roles remain unchanged. That residual variance, the authors argue, is not statistical noise but the very location of symbolic use, the phenomenon where tool ownership is recorded while routines stay frozen.

This distinction underpins the article’s six propositions, organized under two objectives. The first objective concerns how human-centered design conditions the conversion of AI access into capability. Proposition P1.1 holds that a human-centered foundation raises the likelihood that sensing, seizing and reconfiguring occur, the classic microfoundations of dynamic capabilities. Proposition P1.2 predicts that where the foundation is absent, adoption stalls at symbolic use. Proposition P1.3 rejects the assumption that AI is uniformly augmentative, arguing instead that the appropriate configuration depends on task criticality, error cost, employee expertise, regulatory constraint and system type; full automation remains appropriate for some routine, low-stakes, high-volume tasks.

The second objective addresses how organizational adaptation converts capability into competitiveness. Proposition P2.1 asserts that reconfiguring produces structural, cultural and human-capital adaptation as a realized state downstream of it, and that the two are empirically separable. Proposition P2.2 goes further: adaptation carries a substantial part of the association between AI capability and competitiveness, so controlling for it should attenuate, though not necessarily eliminate, the direct link between tool ownership and performance. Proposition P2.3 proposes a temporal ordering in which competitiveness dimensions realize in sequence, with efficiency emerging earliest because it follows from task-level substitution once use is embedded, responsiveness and innovation arriving later, and realized resilience observable only across an actual disruption.

What elevates the article beyond conceptual synthesis is its methodological rigor. Each proposition is paired with candidate indicators and a research design capable of testing it. For P1.2, the authors suggest measuring the gap between tool ownership and the share of relevant transactions actually passing through the tool, using longitudinal panels with archival and log data. For P2.1, they propose comparing a one-factor measurement model against distinct but correlated factors, treating discriminant validity as a test of the claim rather than an assumption. For P2.2, they recommend three-wave panels with baselines, difference-in-differences designs around exogenous deployment timing, and matched comparisons to address the self-selection that plagues adoption research. The authors are candid that self-reported use cannot separate symbolic from embedded use, and that only behavioral and archival traces, such as system access logs and documented procedure revisions, can.

The framework’s engagement with emerging and transition economies adds a distinctive dimension often absent from AI adoption research. The authors note that public cloud AI compute is concentrated in a small number of jurisdictions, meaning most countries consume AI services hosted elsewhere, a structural constraint they expect to limit reconfiguring to what foreign platforms permit. Language compounds the challenge: model performance in lower-resource languages remains materially weaker than in English, and Uzbek is a documented low-resource case where basic annotated resources have only recently been constructed. The authors expect the transparency principle to bind harder in such settings, because staff cannot verify outputs they cannot fully read. They also flag that Uzbekistan’s regulatory framework addresses data localization and national AI capacity but does not yet specify obligations concerning algorithmic management or the contestation of automated decisions, meaning safeguards are least likely to be supplied by law and most likely to require deliberate design.

The practical implications are refreshingly concrete. Firms should conduct readiness assessments with staff rather than for them, treating the proportion of employees able to explain what a tool does as a deployment gate. They should define per-tool success criteria before piloting, including a stated override rate at which the tool is reviewed. And they should 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 signal that symbolic use has taken hold. Which measure binds first depends on context: in a small manufacturer deploying predictive maintenance, the override measure matters most because unexplained recommendations get disregarded; in a professional-services firm using generative drafting, procedure revision matters more because output is easy to adopt informally but hard to govern.

The authors acknowledge four limitations with unusual candor: the framework is conceptual and untested, its human-centered layer is deliberately restricted, its constructs derive largely from high-income settings, and its indicators are unvalidated. Each limitation comes paired with a remedy, from multi-wave designs to comparative work in transition economies. Perhaps the most consequential extension they identify is the inclusion of customer and community effects, which are currently absent from the framework entirely. Worker-level outcomes such as autonomy, workload, skill use and job quality, by contrast, appear as ends in their own right, capable of diverging from firm competitiveness rather than tracking it. The article’s closing message is deceptively simple: an AI purchase is the beginning of a change project rather than the end of a decision, and researchers should follow firms past the point of purchase. Whether the framework’s propositions survive empirical testing remains an open question, but the authors have done the harder preliminary work of specifying exactly what would have to be observed for the connection between human-centered design and small-firm competitiveness to hold.

Subject of Research: A conceptual framework linking human-centered AI design principles to capability formation and organizational adaptation in small and medium-sized enterprises

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

Article References: Navruz-zoda, L., Najmiddinov, S., Norboev, S., Ibadullaev, E., & Zakirov, M. T. (2026). Human-centered AI adoption for small and medium-sized enterprise competitiveness: a methodological perspective on digital capabilities and organizational adaptation. Frontiers in Human Dynamics, 8, Article 1939008. https://doi.org/10.3389/fhumd.2026.1939008

Image Credits: AI Generated

DOI: 10.3389/fhumd.2026.1939008

Keywords: artificial intelligence, SMEs, human-centered AI, dynamic capabilities, organizational adaptation, digital transformation, AI adoption, competitiveness, responsible AI, transition economies, work design, research methodology