The theoretical scaffolding of the research deserves some unpacking for readers unfamiliar with these traditions. The Diffusion of Innovation framework, originally articulated by Everett Rogers, holds that the rate and depth at which an innovation spreads through a social system depends on how potential users perceive it: whether it fits their existing ways of working, whether it seems easy to understand and operate, whether it outperforms what it replaces, and whether it can be trialled and observed before full commitment. The Technology-Organisation-Environment framework complements this by situating adoption within three layers of influence: the technological tools themselves, the internal resources and structures of the firm, and the external pressures and supports of the wider business environment. Combining the two allows the researchers to test whether perceptions of the technology and characteristics of the firm’s context jointly shape actual use behaviour, rather than treating either in isolation.

The choice of Partial Least Squares Structural Equation Modelling reflects both the nature of the data and the state of the field. PLS-SEM is well suited to studies that combine multiple latent constructs measured through survey items with moderately sized samples and complex path structures, particularly when the goal is exploratory or predictive rather than purely confirmatory. It estimates the measurement model, which checks that survey items genuinely capture the constructs they are meant to represent, and the structural model, which quantifies the strength and significance of relationships between constructs. The multi-group analysis layered on top is the methodological centrepiece: it partitions the sample into subgroups, here medium-scale and large-scale firms, and tests whether the same predictors carry the same weights in each. Before such comparisons are valid, measurement invariance must be established, which is precisely what the MICOM procedure reported in the appendix addresses. The finding of full invariance for most constructs and partial invariance for the remainder indicates that the two groups understood the survey constructs in sufficiently similar ways for cross-group comparisons to be meaningful, a step that many published studies unfortunately skip.

The pattern of results rewards close reading. Compatibility emerged as the only predictor significant in every group, with path coefficients ranging from 0.162 to 0.325, and it also carried the largest single effect in the pooled analysis. This suggests that the single most reliable lever for increasing technology use is ensuring that new equipment and workflows align with existing production practices, skill sets, and job designs. In garment production, where cutting, sewing, and finishing follow deeply embedded sequences, technologies that demand radical reorganisation of tasks face resistance not from stubbornness but from the genuine disruption they impose on throughput and piece-rate earnings. Technologies that can be folded into familiar routines, by contrast, are absorbed quickly and used consistently.

The divergence between firm sizes is equally instructive. Environmental context, encompassing external pressures such as customer expectations, regulatory signals, export market requirements, and industry norms, predicted technology use strongly in medium-scale firms but not in the pooled model’s other segments, while complexity mattered for large-scale firms. One plausible reading is that medium-scale firms operate in closer proximity to their customers and supply chains, making external demands more visible and immediately consequential to daily operations. Large-scale firms, with more formalised structures and dedicated technical staff, may internalise environmental pressures at the managerial level, leaving complexity, the perceived difficulty of operating the technology, as the salient barrier or facilitator for the workers who actually run the machines. This kind of size-dependent mechanism is exactly what multi-group analysis is designed to detect, and it cautions against one-size-fits-all technology promotion policies.

The non-significant findings are as informative as the significant ones. Relative advantage, the perceived superiority of the technology over existing methods, did not reach significance in any group, and organisational context likewise failed to predict use once other factors were controlled. The latter result may seem counterintuitive, since firm resources, leadership support, and internal structures are commonly cited as adoption drivers. But adoption and use are different behaviours: a firm may possess ample resources and supportive management, yet if the technology clashes with shop-floor realities, those organisational strengths never translate into sustained utilisation. The relative advantage result may reflect the possibility that workers judge technologies less on abstract superiority and more on whether those technologies fit into the concrete rhythm of their tasks, which is precisely what the compatibility construct captures.

Placing the study in its regional and sectoral context sharpens its significance. Ghana’s garment industry occupies a competitive middle ground: it faces pressure from low-cost imports, aspires to capture shares of global apparel value chains that demand certified sustainability practices, and simultaneously contends with energy costs, skills shortages, and limited access to finance. Sustainable garment production in this setting is not merely an environmental aspiration but a market-access requirement, as international buyers increasingly audit suppliers for energy efficiency, waste reduction, and chemical management. Modern technologies, from computer-aided design and automated cutting to water-saving dyeing processes, are the practical instruments through which such requirements are met. Understanding what drives their use therefore connects directly to the country’s industrialisation strategy and export competitiveness.

The sample of 516 production workers in the Greater Accra Region merits attention for what it implies about generalisability. Greater Accra concentrates much of Ghana’s formal manufacturing and hosts the largest cluster of medium and large garment enterprises, making it a reasonable site for studying these firm classes. At the same time, workers in smaller firms, informal workshops, or other regions may face different constraints, and the study’s conclusions should be read as applying to the formalised segment it sampled. The descriptive analytical survey design captures associations at a single point in time, which means the findings identify predictors that travel together with technology use rather than proving causal direction. Longitudinal or quasi-experimental follow-up could establish whether improving compatibility, for instance, actually raises utilisation over time.

The methodological transparency of the paper strengthens confidence in its conclusions. Harman’s single-factor test, reported in the appendix, addresses the concern that common method variance, artefactual covariance arising from collecting all measures through the same survey, might inflate the observed relationships. The extraction results show the first factor accounting for well under half of the total variance, which is the conventional threshold suggesting that no single factor dominates the response patterns. Combined with the MICOM invariance results, these checks indicate that the multi-group comparisons rest on solid psychometric ground rather than on artefacts of measurement.

For practitioners, the actionable implications are concrete. Managers introducing new production technologies should invest as much in workflow integration and worker training as in the hardware itself, since compatibility is the strongest and most consistent predictor of use. For medium-scale firms, engaging with external ecosystems, industry associations, buyer sustainability programmes, and regulatory initiatives, appears to reinforce technology utilisation, suggesting that participation in such networks is not merely reputational but operationally meaningful. For large-scale firms, simplifying the perceived complexity of new systems, through intuitive interfaces, staged rollouts, and peer mentoring on the floor, addresses the barrier that most directly suppresses use among their workforce.

For policymakers, the size-dependent findings argue for differentiated support instruments. Programmes that subsidise equipment purchases, a common industrial policy tool, address adoption but not utilisation, and the evidence here suggests the binding constraint lies downstream of purchase. Extension services, demonstration projects, and technical assistance that help firms adapt technologies to their existing processes would target the compatibility mechanism directly. Similarly, strengthening environmental enablers, such as sustainability certification support, export promotion tied to green production, and reliable infrastructure, would disproportionately benefit the medium-scale segment that responds most to external signals.

The study also contributes to theory by extending the DOI and TOE frameworks into a sector and setting where they had been under-applied. Most prior applications of these frameworks concern information technology adoption in services or knowledge-intensive firms; testing them in the context of industrial production technology among garment workers demonstrates their portability while revealing which constructs carry weight in a hands-on manufacturing environment. The demonstration that predictors vary by firm size adds a contingency dimension that the original frameworks treat only implicitly, and future research could probe additional moderators such as worker skill level, technology type, ownership structure, or export orientation.

Several avenues for further work follow naturally. Replication in other regions of Ghana and in neighbouring West African economies would test whether the compatibility primacy and the size-dependent pattern hold beyond Greater Accra. Studies linking measured technology use to objective sustainability outcomes, such as energy consumption per garment, material waste rates, or defect levels, would close the loop between the predictors studied here and the environmental and economic benefits that motivate the research. Qualitative work inside firms could illuminate the mechanisms behind the findings, revealing, for example, how compatibility is achieved in practice or how environmental pressures filter down to workers. Finally, intervention studies that deliberately manipulate compatibility or complexity, through training redesign or interface modification, would provide the causal evidence that survey designs cannot.

In sum, the research offers a carefully executed, theoretically grounded account of why modern production technologies are underused in Ghana’s garment sector and what would change that. Its central message, that the path to sustainable technology utilisation runs through fit between technology and workplace rather than through the technologies alone, and that this path differs by firm size, gives both managers and policymakers a more precise map than the generic adoption literature has typically provided. As sustainability requirements tighten across global apparel supply chains, the firms and economies that learn to convert technology ownership into technology use will be the ones that convert industrial investment into competitive advantage.

Subject of Research: Predictors of modern technology use for sustainable garment production in Ghanaian garment firms

Article Title: Predictors of modern technology use for sustainable garment production among medium and large-scale firms using a multi-group analysis

Article References: Lamptey, L., Amankwah, A. M., Asinyo, B. K., & Tawiah, B. (2026). Predictors of modern technology use for sustainable garment production among medium and large-scale firms using a multi-group analysis. Discover Sustainability. https://doi.org/10.1007/s43621-026-04959-z

Image Credits: AI Generated

DOI: 10.1007/s43621-026-04959-z

Keywords: sustainable garment production, technology adoption, Ghana, PLS-SEM, multi-group analysis, Diffusion of Innovation theory, Technology-Organisation-Environment framework, firm size, compatibility, complexity, environmental context, garment industry

The technical heart of the study is a safety-gated form of tabular Q-learning, a classic reinforcement learning algorithm in which an agent maintains a table of expected long-term rewards for each combination of state and action. What makes this implementation distinctive is the set of hard constraints wrapped around the learning process. The simulation represents safety, labour, and environmental performance as separate dimensions for each supplier, and the agent is never allowed to collect rewards when verified safety or labour performance falls below a threshold of 0.60. A minimum annual audit is imposed on every supplier regardless of what the agent would prefer. These gates are evaluated before any policy action is executed, and crucially, an unverified self-disclosure from a supplier cannot lift the gate. Only the last successfully verified safety and labour values count.

This design choice reflects a deep insight about the problem structure. Supplier sustainability is a partially observable process: the agent cannot see the latent true sustainability of a factory, only noisy proxies such as the most recent verified score, current disclosure, performance trend, audit recency, and any verified critical flags. The mathematical framework carefully separates these latent variables from observed ones, and the observable state is constructed as an approximate discrete representation of the underlying reality. By preventing disclosure alone from making a supplier reward-eligible, the model encodes a principle that governance experts have long advocated: self-reported claims are not evidence.

To test the approach, the researcher trained three independent Q-learning agents and evaluated them in 75 paired out-of-sample replications against six competing policies: a contextual bandit, a risk-based adaptive heuristic, a static threshold rule, a combined scheduled-audit-and-threshold policy, scheduled auditing alone, and random action selection. The paired replication design matters for statistical rigor. Rather than treating each of the 1,440 supplier-month transitions as an independent observation, the study aggregates incidents and actions at the replication level before running paired tests, avoiding the artificial precision that inflates significance in many simulation studies.

The results are striking. The Q-learning agents achieved a mean true compliance of 0.862, compared with 0.820 under the static threshold policy that represents a common industry baseline. The share of suppliers scoring at or above 0.80 rose from 58.2 percent under the baseline to 81.3 percent under the learning agent. Perhaps most dramatically, the mean number of severe incidents fell from 3.29 to 1.71, a reduction of nearly half. All of this was accomplished with 167 audits, against 362 for the combined scheduled-audit-and-threshold policy, a 53.8 percent reduction in audit burden. The trade-off was a modest one: mean procurement cost ran about 1.1 percent higher than under the static threshold.

The comparison with the contextual bandit is particularly instructive for machine learning practitioners. A contextual bandit is a simpler learning algorithm that chooses actions based on immediate observed context without modeling delayed consequences. In these experiments, the bandit produced lower compliance, at 0.742, and more severe incidents, at 2.96, than the full Q-learning agent. The gap supports a central argument of the paper: supplier governance is a delayed-action problem. The consequences of shifting sourcing share away from a factory, or of skipping an audit, unfold over months, and an algorithm that only optimizes immediate reward cannot capture those dynamics. Q-learning, which propagates future value backward through its state-action table, can.

The risk-based adaptive heuristic, a hand-crafted policy that allocates audits according to assessed risk, fared better than most alternatives. It produced statistically similar mean compliance to the Q-learning agent, but it required 55 additional audits and yielded a narrower share of compliant suppliers. In other words, the learned policy matched the performance of a well-designed human heuristic while using fewer resources and spreading compliance gains more broadly across the supplier base. For procurement organizations weighing the cost of sophisticated analytics against the cost of audits, that efficiency margin could be decisive.

The study also reports an unusually thorough set of robustness checks. Paired statistical tests are accompanied by effect sizes, convergence diagnostics confirm that the Q-tables stabilized during training, and results are broken down by supplier stratum, distinguishing direct exporters, subcontractors, and small or informal suppliers, the three tiers that characterize real garment sourcing portfolios. A credit-assignment ablation examines how the reinforcement signal drives learning, and time-state testing probes whether the policy generalizes across simulation periods. Multi-parameter robustness analyses vary key model settings to confirm the findings are not artifacts of a single configuration. The sourcing-share adjustment itself is formulated as a convex quadratic projection with linear equality and bound constraints, which guarantees a unique allocation that is independent of the order in which suppliers are processed, a subtle but important fairness property.

Several implementation details reveal how carefully the model guards against gaming and error. The annual audit floor is computed from elapsed months since the last successful verification, not since an attempted audit, so a failed inspection does not reset the clock and buy a supplier extra time. Audit recency and the supplier-specific reinforcement signal are stored as separate fields in the implementation, preventing conceptual conflation. The critical gate relies exclusively on verified values, closing the loophole through which a struggling factory might talk its way back into good standing. These choices make the simulation a meaningful test of adaptive governance rather than a naive optimization exercise.

The author is candid about the limits of the work. This is simulation evidence, built from a stochastic model calibrated with published aggregate statistics, not factory-level validation, and no real procurement manager has yet handed sourcing decisions to an algorithm. The study involved no human participants and received no external funding. Still, the implications are considerable. Ready-made garment supply chains face mounting legal, commercial, and stakeholder pressure to verify sustainability rather than trust supplier claims, and audit budgets are finite. A framework that lifts compliance, halves severe incidents, and cuts audit volume by more than half, even in a stylized setting, suggests that reinforcement learning could become a practical instrument of supply chain governance. The open-access paper, published on 5 October 2026 with DOI 10.1007/s43621-026-04745-x, provides a reproducible foundation for the field trials that must come next, and it signals a future in which the clothes we wear are policed not just by clipboard-wielding inspectors, but by algorithms that learn where the real risks hide.

Subject of Research: Reinforcement learning for adaptive supplier sustainability scoring and audit targeting in ready-made garment supply chains

Article Title: Reinforcement learning for adaptive supplier sustainability scoring and sourcing share adjustment in ready made garment supply chains

Article References: Anzum, K. M. T. (2026). Reinforcement learning for adaptive supplier sustainability scoring and sourcing share adjustment in ready made garment supply chains. Discover Sustainability. https://doi.org/10.1007/s43621-026-04745-x

Image Credits: AI Generated

DOI: 10.1007/s43621-026-04745-x

Keywords: reinforcement learning, Q-learning, supply chain management, sustainability, garment industry, supplier governance, audit targeting, sourcing share adjustment, Bangladesh, machine learning, simulation, compliance