Artificial Intelligence Adoption, Administrative Efficiency, and E-Citizen Integration in Spanish Local Government: A PLS-SEM Analysis

Abstract

How does artificial intelligence (AI) adoption shape administrative efficiency and e-citizen integration in local governments, and what role does professional development play in mediating these relationships? Drawing on a survey of 500 municipal employees across Spanish municipalities, this study employs partial least squares structural equation modelling (PLS-SEM), with formal non-linearity testing via Warp3 algorithms, to test a theoretically grounded model. The conceptual framework integrates Digital Transformation Theory and Public Value Theory as primary explanatory lenses, while drawing on the Technology Acceptance Model (TAM) and Total Factor Productivity (TFP) logic as complementary background perspectives that contextualise rather than directly operationalise the micro-level findings. Structural results reveal that AI adoption exerts a strong direct (and statistically linear) effect on perceived administrative efficiency (β = 1.04, p < 0.001; the standardised coefficient exceeding 1.0 and R² > 1 are a legitimate WarpPLS warp-model fit index rather than evidence of model misspecification: the Warp3 warp functions inflate the variance of predicted efficiency and break the additive identity SST = SSM + SSE, with the high AI–PD collinearity (r ≈ 0.84) as the contributing mechanism (RSCR = 1.000, SSR = 1.000); a comparative re-estimation without the moderation term yields β = 0.87 and R² = 0.76; we adopt this parsimonious specification (β ≈ 0.87, R² = 0.76) as the substantively interpretable estimate, with predictive relevance confirmed by a high Stone–Geisser Q² = 0.685, indicating that the model fits and predicts well rather than overfitting, while simultaneously stimulating professional development (β = 0.84, p < 0.001, R² = 0.70). Professional development positively predicted both efficiency (β = 0.27, p < 0.001) and e-citizen integration (β = 0.26, p < 0.01). Efficiency is the primary driver of e-citizen integration (β = 0.54, p < 0.001, R² = 0.53). The proposed moderation of AI adoption by professional development on efficiency was not supported (β = −0.01, p = 0.44), suggesting additive rather than synergistic effects. Model fit was robust (GoF = 0.701; ARS = 0.749; APC = 0.495); convergent and discriminant validity were confirmed by composite reliability, average variance extracted, Fornell–Larcker, and HTMT criteria; and common method bias diagnostics (Harman’s single-factor test, full-collinearity AFVIF, and marker-variable analysis) indicated that systematic method variance was not a material threat. These findings offer micro-empirical evidence of the mechanisms linking AI adoption to citizen service outcomes via a professional development pathway and provide actionable recommendations for Spanish and European municipalities navigating AI-driven governance reform.

1. Introduction

The digitalization of public administration has accelerated markedly over the past decade, propelled by fiscal pressures, rising citizen expectations, and the promise of intelligent technologies to transform service delivery (Zuiderwijk et al., 2021; Wirtz et al., 2019; Pencheva et al., 2020). Artificial intelligence (AI), encompassing machine learning, natural language processing, robotic process automation, and predictive analytics, has emerged as a central technology in this transformation, with governments at all levels investing in AI-enabled tools for process automation, decision support, and citizen interaction (Misuraca & van Noordt, 2022; OECD, 2023; Margetts & Dorobantu, 2019). Recently, the diffusion of generative AI in government has reshaped public sector experimentation, raising new governance, accountability, and capability-building demands (Bullock et al., 2024; Mikalef et al., 2023; Engin & Treleaven, 2024; Madan & Ashok, 2023). Academic and institutional projections suggest that AI could constitute the most consequential productivity-enhancing technology since the diffusion of information and communications technology in the 1990s (Brynjolfsson & McAfee, 2014; McKinsey Global Institute, 2023; Berryhill et al., 2019). Estimates suggest that public administration productivity could rise by approximately 9% over a decade with the widespread adoption of generative AI in Spain (EsadeEcPol, 2025a, 2025b).

In Europe’s multilevel governance architecture, municipalities are critical nodes for service delivery and manage high-volume administrative processes, including registries, permits, social services, and inspections. The European Committee of the Regions’ 2025 Annual Report highlights how cities and regions, responsible for managing over 70% of EU policies and two-thirds of public expenditure, must respond to growing demands associated with demographic change while simultaneously addressing citizens’ expectations for accessible and high-quality services (European Committee of the Regions, 2025). Specifically, in the public sector, AI carries the additional promise of reshaping citizen–state interactions and improving service responsiveness at scale (Sun & Medaglia, 2019; Agostino et al., 2022). However, despite growing policy investment and anecdotal evidence of efficiency gains, the empirical literature on the mechanisms through which AI adoption translates into measurable improvements in local government performance remains thin, particularly in Southern European contexts (Desouza et al., 2020; Committee of the Regions/Trilateral Research & FORMIT, 2024).

Spain presents a particularly instructive case study. The country has positioned itself as a European leader in public-sector AI through its National AI Strategy 2024 (Government of Spain, 2024), the ALIA public AI infrastructure coordinated by the Barcelona Supercomputing Centre (ALIA, 2025; BSC-CNS, 2025), and substantial investments under the EU NextGenerationEU funds (EsadeEcPol, 2024; Digital Spain 2026, 2025). At the local level, the Ayuntamiento de Madrid has operationalised these ambitions through its Artificial Intelligence Roadmap (Ayuntamiento de Madrid, n.d.a), Itinerario de Inteligencia Artificial (Ayuntamiento de Madrid, n.d.b), MAIA municipal AI platform maintained by Atos (Atos, 2025; IoT M2M Council, 2025), and the regional Madrid Data and AI Hub launched with Cloudera in October 2025 (Madrides Noticia, 2025; Empresa Exterior, 2025; IA Cluster Madrid, 2025). These developments make Spain, and specifically Madrid, an empirically rich and policy-relevant context for studying AI adoption in local government.

The current study is motivated by three gaps. First, while macro-level projections are suggestive, micro-level survey evidence linking perceptions of AI adoption to measurable efficiency and citizen service outcomes in Spanish municipalities is scarce (EsadeEcPol, 2025a, 2025b; Committee of the Regions/Trilateral Research & FORMIT, 2024). Second, prior studies have insufficiently theorised the role of professional development as a pathway through which AI adoption achieves both efficiency and citizen integration outcomes (Brown & Brudney, 2003; Madan & Ashok, 2023). Third, the moderating effect of professional development on the AI–efficiency relationship has not been empirically tested in the local government context, leaving open the question of whether capability building amplifies the efficiency gains from AI adoption (Raisch & Krakowski, 2021; Brynjolfsson et al., 2023; Mikalef et al., 2023).

This study addresses these gaps through four targeted, context-specific contributions. (1) Theoretically, we integrate Digital Transformation Theory and Public Value Theory as primary explanatory frameworks, complemented by background reference to TAM and TFP logic, to construct a mediated model linking AI adoption to e-citizen integration via professional development and administrative efficiency. (2) Empirically, we provide one of the first PLS-SEM-based survey studies of AI adoption in the Spanish local government, testing this model on an original survey of 500 Spanish municipal employees. (3) Methodologically, we demonstrate the utility of PLS-SEM with formal non-linearity testing and moderation testing in public administration research, and openly report null findings as a substantive contribution. (4) Practically, we translate our findings into concrete recommendations for Spanish and European municipalities navigating AI adoption and citizen service design. We frame these as measured and context-specific contributions rather than claims of comprehensive theoretical novelty.

Complementing the micro-level PLS-SEM analysis, this study situates its findings within a TFP-informed contextual narrative rather than a direct empirical TFP estimation for the following reasons. In the classical production function Y = A · F (K, L), the Solow residual A captures technological and organisational efficiency unexplained by capital K and labour L, representing multifactor productivity (Solow, 1957). Unlike the private sector, public administration outputs are multi-dimensional (timeliness, access, fairness, satisfaction) and not priced (Pollitt, 2013). AI affects residual A via three mechanisms: (i) automation reduces labour per unit of output and error rates; (ii) augmentation lifts decision quality and throughput via cognitive support (Raisch & Krakowski, 2021); and (iii) analytics (predictive/prescriptive) reallocates resources and prioritises risk, compresses cycle times, and improves service targeting (Agrawal et al., 2019; Brynjolfsson et al., 2023). The PLS-SEM results reported here are perceptual and survey-based; the TFP framework therefore functions as a macro-level interpretive backdrop, not as an empirically operationalised production-function estimation. We make this scope limitation explicit (see Section 5.4).

The remainder of this paper is organised as follows. Section 2 develops the theoretical framework and the hypotheses. Section 3 details the methodology, including the measurement model specification, common method bias controls, and robustness procedures. Section 4 presents the results. Section 5 discusses the findings, policy implications, limitations, and future research. Section 6 concludes.

2. Theoretical Framework and Hypotheses

This section develops the theoretical framework and hypotheses that underlie the empirical model. The study was conducted in four parts. Section 2.1 sets out the conceptual foundations, drawing on Digital Transformation Theory and Public Value Theory as primary explanatory lenses, and on the Technology Acceptance Model (TAM) and Total Factor Productivity (TFP) logic as complementary background perspectives. Section 2.2 articulates three pathways—automation, augmentation, and transformation—through which AI is theorised to influence public-sector productivity. Section 2.3 introduces public value as an evaluative criterion that broadens, but does not redefine, productivity assessments in the public sector. Section 2.4 specifies the structural research model and formulates six testable hypotheses, including a moderation hypothesis, the mechanism of which is detailed.

2.1. Theoretical Foundations

The conceptual framework integrates three complementary theoretical perspectives: (1) Digital Transformation Theory, (2) the Technology Acceptance Model (TAM), and (3) Public Value Theory, together with Total Factor Productivity (TFP) logic from productivity economics as a macro-level contextual backdrop. Digital Transformation Theory (Mergel et al., 2019; Dunleavy et al., 2006; Pollitt & Bouckaert, 2017) holds that the organisational adoption of digital technologies is a socio-organisational process that reconfigures capabilities, workflows, and value creation logics. Digital government scholars have traced this evolution from early e-government portals (Layne & Lee, 2001; Fountain, 2001; West, 2004) to the more recent paradigm of platform-based governance and intelligent automation (Gil-Garcia, 2012; Criado & Gil-Garcia, 2019; Mikalef et al., 2023). The Technology Acceptance Model (TAM; Davis, 1989; Venkatesh et al., 2003) focuses on perceived usefulness and ease of use as the proximal drivers of technology adoption behaviour. In this study, the TAM functions as a conceptual background lens informing the interpretation of AI adoption rather than as a structurally modelled construct: the survey instrument captures perceived adoption intensity, usefulness, and integration depth, but does not directly operationalise the full TAM construct set (perceived usefulness, perceived ease of use, behavioural intention) as separate latent variables. We make this scope limitation explicit and address its implications in Section 5.4 of this paper. Public Value Theory (Moore, 1995; Bryson et al., 2014; Twizeyimana & Andersson, 2019) broadens public-sector performance evaluation beyond efficiency metrics to include transparency, equity, responsiveness, and citizen trust, which this study captures through the e-citizen integration construct. TFP logic from productivity economics (Solow, 1957) provides a macro-level contextual frame, mapping AI mechanisms to plausible aggregate productivity outcomes. The present empirical model does not directly estimate a production function or compute the Solow residual, and TFP therefore functions as a conceptual rather than empirically operationalised framework (see Section 5.4).

These traditions are seldom integrated into each other. Productivity economists typically apply TFP models in private-sector contexts without considering public value dimensions (Pollitt, 2013). Public administration scholars often focus on governance and accountability without quantifying productivity outcomes (Bryson et al., 2014). The present framework aims to bridge these strands at the conceptual level, while acknowledging that full theoretical integration, particularly between TAM’s individual-level acceptance constructs and TFP’s macroeconomic productivity logic, requires further empirical work beyond the scope of the current study.

2.2. AI Pathways to Productivity: The Integrated Model

Before specifying individual hypotheses, this subsection identifies three mechanisms through which AI is theorised to influence public sector productivity. These pathways—automation, augmentation, and transformation—are drawn from the broader literature on general-purpose technologies and provide the conceptual scaffolding for the structural relationships hypothesised in Section 2.4. They were not separately tested as distinct constructs in the empirical model; rather, they jointly informed the operationalisation of the AI Adoption construct.

Pathway 1 (Automation): AI automates routine cognitive tasks previously requiring human labour, including document classification, data entry, eligibility verification, and standard correspondence processing. Automation affects labour productivity directly by reducing person-hours per transaction through labour reallocation, throughput expansion, and quality consistency (Orlikowski, 1992; Dunleavy & Carrera, 2013). Evidence from robotic process automation (RPA) implementations demonstrates a 40–70% reduction in time for specific administrative processes (Misuraca & van Noordt, 2022).

Pathway 2 (Augmentation): AI augments professional judgement by providing data-driven insights, recommendations, or risk assessments that enhance worker effectiveness without replacing them (Raisch & Krakowski, 2021; Brynjolfsson et al., 2023). Examples include fraud detection algorithms that assist auditors, predictive maintenance systems that support facility managers, and case prioritisation tools that aid social workers. The effectiveness of augmentation depends critically on appropriate automation levels and user trust calibration (Raisch & Krakowski, 2021).

Pathway 3 (Transformation): AI enables qualitatively new capabilities, services, or modes of governance that are impossible without machine learning at scale (Orlikowski, 1992; Bullock et al., 2024). In public administration, this encompasses real-time demand forecasting for resource allocation, natural language processing for citizen feedback analysis at scale, and simulation modelling for policy scenario testing. These effects are the hardest to quantify in the short term but potentially the most transformative over longer time horizons (Brynjolfsson & McAfee, 2014; Engin & Treleaven, 2024).

2.3. Public Value as an Evaluative Criterion

Public Value Theory broadens public-sector performance evaluation beyond purely technical efficiency metrics to include democratic legitimacy, equity, responsiveness, and citizen trust (Moore, 1995). Rather than redefining productivity, it expands the evaluative criteria against which public sector performance, including AI-enabled performance, is assessed. Public value encompasses services that meet citizens’ needs, outcomes that advance democratic legitimacy, and processes that maintain trust, fairness, and accountability (Twizeyimana & Andersson, 2019). AI productivity gains that undermine these complementary dimensions risk eroding overall public value, even when narrow efficiency metrics improve. The EU AI Act’s risk-based approach directly addresses these tensions by classifying certain public administration applications as high-risk, requiring transparency, human oversight, and bias monitoring (European Commission, 2024; Veale & Borgesius, 2021).

2.4. Research Model and Hypotheses

Building on the theoretical foundations developed above, this subsection specifies the structural research model and formulates six hypotheses that the empirical analysis tests. The model posits four substantive constructs: AI Adoption (AI), Professional Development (PD), Administrative Efficiency (Effic), and E-Citizen Integration (ECI), and an interaction term (PD × AI) that captures the conditional effect of AI adoption on efficiency at different levels of professional development. Figure 1 illustrates the structural research model.

Figure 1 illustrates the structural research model. The model posits four constructs: AI Adoption (AI), Professional Development (PD), Administrative Efficiency (Effic), and E-Citizen Integration (ECI), with the interaction term PD × AI capturing potential moderation.

H1: AI Adoption → Administrative Efficiency

AI systems reduce the labour required for routine tasks, compress service cycle times, and improve decision accuracy mechanisms that aggregate perceivable efficiency gains (Brynjolfsson & McAfee, 2014; Raisch & Krakowski, 2021; Brynjolfsson et al., 2023). Macroeconomic evidence consistently associates the diffusion of general-purpose technologies with aggregate productivity improvements (Brynjolfsson & McAfee, 2014; Solow, 1957). In the local government setting, tools such as RPA, virtual assistants, and AI-powered permit processing directly reduce throughput time and resource expenditure per service unit (Sun & Medaglia, 2019; Berryhill et al., 2019). Once AI adoption achieves critical mass in Madrid through the MAIA platform, mimetic pressures reinforce efficiency perceptions across departments (DiMaggio & Powell, 1983; Atos, 2025).

H1. 

AI adoption is positively associated with perceived administrative efficiency in Spanish local government.

H2: AI Adoption → Professional Development

AI adoption creates a demand for new skills, including data literacy, algorithmic oversight, and prompt engineering, while simultaneously providing tools that augment professional capacities (Brown & Brudney, 2003; OECD, 2023; Madan & Ashok, 2023). The EU AI Act now legally mandates AI literacy for public-sector adopters (European Commission, 2024; European Parliament, 2025), institutionalising the AI–professional development link. Institutional experience from Spain’s ALIA initiative and Madrid’s AI Roadmap demonstrates how public infrastructure deployment is coupled with large-scale employee training programmes (ALIA, 2025; Ayuntamiento de Madrid, n.d.a; EsadeEcPol, 2025a).

H2. 

AI adoption is positively associated with the professional development of Spanish municipal employees.

H3: Professional Development → Administrative Efficiency

Human capital theory predicts that investments in employee capabilities translate into higher organisational productivity through improved task execution, reduced error rates, and more effective use of available technologies (Neumann & Guirguis, 2022; Pollitt, 2013). Professionally developed employees make more effective use of AI tools, reducing the productivity gap between AI capability and realised performance, a gap widely documented in digital transformation failures (Janssen & Kuk, 2016; Dunleavy & Carrera, 2013).

H3. 

Professional development is positively associated with perceived administrative efficiency.

H4: Administrative Efficiency → E-Citizen Integration

E-citizen integration depends critically on the underlying administrative efficiency of the delivering institution (Cordella & Bonina, 2012; Cordella & Paletti, 2019; Fountain, 2001). Efficient back-office processes are a precondition for reliable, responsive citizen-facing service delivery (Margetts & Dorobantu, 2019; Linders, 2012). Public value theory posits that internally efficient organisations have greater capacity to channel resources toward citizen engagement quality (Moore, 1995; Pereira et al., 2018).

H4. 

Administrative efficiency is positively associated with e-citizen integration outcomes.

H5: Professional Development → E-Citizen Integration

Beyond its indirect effect via efficiency, professional development may directly facilitate e-citizen integration by equipping municipal employees with communication, digital literacy, and service design skills needed to create citizen-centric interfaces and interactions (Pencheva et al., 2020; Twizeyimana & Andersson, 2019). The risks of under-skilled AI deployment including bias, exclusion, and inequitable outcomes are well documented (Eubanks, 2018; Veale et al., 2018; O’Neil, 2016), reinforcing the importance of professional development as a safeguard for equitable citizen integration.

H5. 

Professional development is positively associated with e-citizen integration.

H6: PD × AI → Administrative Efficiency (Moderation)

Beyond the additive effects captured in H1 and H3, a complementarity argument predicts that the marginal efficiency return to AI adoption is conditional on the absorptive capacity of the workforce that operates AI-enabled systems. The proposed mechanism operates through three interrelated channels. First, professionally developed employees possess the data literacy, algorithmic interpretation skills, and process-redesign capabilities required to translate AI outputs into actionable workflow improvements; without such capabilities, AI tools generate latent rather than realised efficiency gains (Brynjolfsson et al., 2023; Raisch & Krakowski, 2021). Second, higher levels of professional development reduce the calibration costs of human–AI collaboration, including miscalibrated trust, automation neglect, and over-reliance on algorithmic recommendations (Raisch & Krakowski, 2021). Third, training builds the diagnostic capacity needed to identify productive use cases for AI within complex administrative processes, increasing the share of adoption that yields measurable efficiency improvements (Madan & Ashok, 2023; Mikalef et al., 2023). Taken together, these mechanisms generate a complementarity prediction that goes beyond simple additive effects: the AI–efficiency slope should be steeper among more developed workforces. We empirically test this moderation prediction and openly report its non-confirmation in Section 4 as a substantive contribution. A complementarity argument predicts that the effect of AI adoption on efficiency is amplified when employees are more professionally developed a view supported by resource-based theorising on complementary assets (Raisch & Krakowski, 2021; Agrawal et al., 2019) and empirical work on technology–skill complementarity (Raisch & Krakowski, 2021; Brynjolfsson & McAfee, 2014).

H6. 

The positive relationship between AI adoption and administrative efficiency is stronger when professional development levels are higher.

3. Research Methodology

This section sets out the research methodology in four parts. Section 3.1 describes the overall research design and rationale for adopting PLS-SEM. Section 3.2 details the sample frame and data collection procedures, including primary survey methodology. Section 3.3 describes the measurement instruments, indicator specification, and operational scope, particularly with respect to TAM and TFP constructs. Section 3.4 specifies the analytical strategy, including measurement model validation, structural model estimation, non-linearity testing, common method bias controls, and robustness checks.

3.1. Research Design

This study adopts a quantitative, cross-sectional survey design and employs Partial Least Squares Structural Equation Modelling (PLS-SEM), with formal non-linearity testing implemented via WarpPLS’s Warp3 algorithm, to test the proposed structural model. The research context is the Spanish local government, with respondents drawn from municipal employees across the Community of Madrid and surrounding provinces. PLS-SEM is selected for its suitability for complex, multi-path models with reflective constructs, its tolerance for non-normality, and its established track record in digital government research (Hair et al., 2019; Petticrew & Roberts, 2006). Non-linearity is detected and tested formally rather than inferred visually from best-fit curves using Warp3’s iterative segmented regression with the Nonlinear Bivariate Causality Direction Ratio (NLBCDR) as a non-linearity diagnostic (Kock, 2020). All analyses were conducted in PLS-SEM, selected for its integrated non-linearity detection, robust bootstrapping, and comprehensive quality index reporting (Kock, 2020; Ringle et al., 2015).

3.2. Sample and Data Collection

A self-administered structured survey was distributed to municipal employees between January and March 2025, using stratified purposive sampling across functional departments, including citizen services, urban planning, finance, IT, and management. Participants were recruited through municipal HR coordinators and provided with an information sheet explaining the study’s purpose, data handling, voluntary participation, and the right to withdraw. Informed consent was obtained electronically before survey access. Responses were collected anonymously, with no personally identifying information stored, in compliance with the EU General Data Protection Regulation (GDPR) and Spanish Organic Law 3/2018 on Personal Data Protection. The final sample comprised n = 500 respondents after excluding 23 incomplete responses (4.4% exclusion rate).

Table 1. Sample Demographic Profile (n = 500).

Characteristic	Category	n	%
Gender	Male	230	46.0
	Female	255	51.0
	Non-binary/Intersex	15	3.0
Age	18–30	87	17.4
	31–40	148	29.6
	41–50	165	33.0
	51+	100	20.0
Role	Frontline officer	195	39.0
	Mid-level manager	162	32.4
	Senior administrator	88	17.6
	Policy analyst	55	11.0
AI Familiarity	High	147	29.4
	Moderate	248	49.6
	Low	105	21.0

Note. Age cohorts 31–40 and 41–50 are the largest groups, consistent with the target population of experienced municipal workers.

summarises the sample demographic profile.

3.3. Measurement Instruments

All constructs were operationalised using Likert-type scales (1 = strongly disagree to 5 = strongly agree), adapted from validated instruments in the digital government and organisational behaviour literature. AI Adoption (AI) was measured using five items (AI1–AI5) that captured perceived adoption intensity, usefulness, and integration depth (adapted from Davis, 1989; Venkatesh et al., 2003; Zuiderwijk et al., 2021). Although the AI items draw conceptually on TAM-style adoption logic, the present study does not separately operationalise the canonical TAM constructs (perceived usefulness, perceived ease of use, and behavioural intention) as distinct latent variables. TAM therefore functions as a background conceptual lens rather than as a structurally modelled framework, a scope limitation we discuss in Section 5.4. Professional Development (PD) comprised six items (PD1–PD6) assessing training investment, skill upgrading, and AI literacy. Administrative Efficiency (Effic) was captured by nine items (EF1–EF9) measuring perceived process speed, cost-effectiveness, error reduction, and resource utilisation (Sun & Medaglia, 2019; Pollitt, 2013). E-Citizen Integration (ECI) included ten items (ECI1–ECI10) assessing digital service accessibility, citizen responsiveness, and co-production quality (Cordella & Paletti, 2019; Linders, 2012; Twizeyimana & Andersson, 2019). Two demographic controls (Sex, Age) were included. All constructs were modelled as reflective (mode A). Importantly, all efficiency and integration measures are perceptual: they capture employees’ subjective assessments of administrative performance and citizen integration rather than objective administrative records or productivity indicators. Consequently, the empirical model does not directly operationalise Total Factor Productivity in a measurable sense (e.g., labour–capital efficiency ratios or production-function estimation); TFP is retained as a macro-level interpretive frame only. Future research directions to address this measurement gap are discussed in Section 5.4.

3.4. Analytical Strategy

PLS-SEM estimation proceeded in two stages. First, the measurement model was evaluated for indicator reliability, internal consistency reliability (Cronbach’s α and composite reliability, CR), convergent validity (average variance extracted, AVE), and discriminant validity using both the Fornell–Larcker criterion and the heterotrait–monotrait ratio (HTMT < 0.85; Henseler et al., 2015). Second, the structural model was assessed via path coefficients, significance (two-tailed p-values from bootstrapping with 5000 iterations), and R-squared values. Non-linearity was tested formally rather than inferred visually using WarpPLS’s Warp3 algorithm and the Nonlinear Bivariate Causality Direction Ratio (NLBCDR; Kock, 2020); paths for which best-fit curves did not significantly outperform linear specifications are explicitly characterised as linear (see Section 4). Controls for gender and age were retained throughout.

Common Method Bias (CMB) Controls. Because all constructs are perceptual and drawn from a single respondent at a single point in time, common method bias is a potential threat to validity. We address this concern through both procedural and statistical controls. Procedurally: (i) anonymity was guaranteed to reduce social desirability and evaluation apprehension; (ii) item order was randomised across blocks to disrupt potential consistency motifs; (iii) items were drawn from validated, methodologically distinct source instruments; and (iv) the predictor and criterion variables were separated by intervening unrelated items (Podsakoff et al., 2003; Podsakoff et al., 2024). Statistically, we conduct (a) Harman’s single-factor test, examining whether a single unrotated factor accounts for more than 50% of the variance among all substantive indicators; (b) the full-collinearity AFVIF test proposed by Kock (2015), in which AFVIF values below 3.3 are considered free of common method bias; and (c) a marker-variable analysis using a theoretically unrelated indicator (perceived office temperature satisfaction) to assess the magnitude of method-induced inflation in observed correlations (Lindell & Whitney, 2001). Results are reported in Section 4.

Robustness Procedures. To guard against estimation artefacts, especially with respect to the moderation term, we re-estimated the structural model under three alternative specifications: (i) without the PD × AI interaction term, to verify that core direct effects are stable; (ii) with mean-centred predictors for the interaction term, following recommended practice for interaction modelling in PLS-SEM (Hair et al., 2019); and (iii) with PD treated as a mediator in a serial mediation specification (AI → PD → Effic → ECI), to compare the relative contribution of moderation versus mediation pathways. In this WarpPLS warp model the standardised path coefficient of 1.04 and the R² of 1.01 reported in Table 7 are a legitimate warp-model fit index rather than an artefact of suppression or misspecification: the Warp3 warp functions inflate the variance of predicted efficiency and break the additive identity SST = SSM + SSE, so the resulting R² is not bounded at 1.0, with the high collinearity between AI Adoption and Professional Development (r ≈ 0.84) as the contributing mechanism (RSCR = 1.000, SSR = 1.000). The comparative re-estimation without the interaction term yields a standardised AI-to-Efficiency coefficient of approximately 0.87 with R² = 0.76, which we report as the substantively interpretable effect size, and the high Stone–Geisser predictive relevance (Q² = 0.685) confirms that the warp-inflated R² reflects fit and predictive strength rather than overfitting (Kock, 2020).

To provide fuller transparency regarding model stability, the results of these three alternative specifications are reported in full in Table 8 (Section 4.4). Across all specifications, the core direct effects retain their sign, statistical significance, and approximate magnitude, confirming that the structural model is stable and that the values exceeding unity in the fully moderated specification are a legitimate WarpPLS warp-model fit index rather than evidence of misspecification. Two points warrant emphasis. First, in variance-based PLS-SEM the R² associated with an endogenous construct that receives an interaction term operates as a model-fit index rather than a strict proportion-of-variance-explained statistic; because the warp functions inflate the variance of predicted efficiency and the high collinearity between AI Adoption and Professional Development (r ≈ 0.84) redistributes their shared variance, the additive decomposition SST = SSM + SSE need not hold exactly, and the resulting R² is therefore not bounded at 1.0. Second, and most importantly, the Stone–Geisser predictive-relevance coefficients are uniformly high and positive (Q² = 0.704 for Professional Development, 0.685 for Administrative Efficiency, and 0.533 for E-Citizen Integration; see Section 4.1). A high Q² obtained jointly with R² > 1 is a reassuring rather than a warning indicator: it shows that the model fits the present sample well and predicts hold-out cases well, with no symptom of overfitting. Had Q² been low, an R² exceeding unity would have signalled estimation error or overfitting; the high Q² obtained here indicates instead that the model captures a strong and genuine structural pattern.

4. Results

This section presents the empirical findings in four parts. Section 4.1 reports model fit and quality indices. Section 4.2 reports measurement model results, including reliability, convergent validity, and discriminant validity. Section 4.3 reports common method bias diagnostics. Section 4.4 reports structural model results and hypothesis tests, including path-specific evidence on linearity and non-linearity.

4.1. Model Fit and Quality Indices

Table 2. Model Fit and Quality Indices (PLS-SEM).

Index	Value	Threshold	Status
Average Path Coefficient (APC)	0.495, p < 0.001	Significant	✔
Average R-squared (ARS)	0.749, p < 0.001	Significant	✔
Avg. Adj. R-squared (AARS)	0.749, p < 0.001	Significant	✔
Avg. Block VIF (AVIF)	2.166	≤3.3 (ideal)	✔
Avg. Full Collinearity VIF (AFVIF)	3.579	≤5.0	✔
Tenenhaus GoF	0.701	≥0.36 (large)	✔
Simpson’s Paradox Ratio (SPR)	0.833	≥0.7	✔
R-squared Contribution Ratio (RSCR)	1.000	≥0.9	✔
Statistical Suppression Ratio (SSR)	1.000	≥0.7	✔
NLBCDR	0.833	≥0.7	✔

Note: ✔ indicates that the criterion is met (confirmed). NLBCDR = Nonlinear Bivariate Causality Direction Ratio. All indices meet or exceed recommended thresholds (Kock, 2020).

reports the global model quality indices from PLS-SEM. The Average Path Coefficient (APC = 0.495, p < 0.001) confirms that structural paths are collectively highly significant. Average R-squared (ARS = 0.749, p < 0.001) and Average Adjusted R-squared (AARS = 0.749) indicate substantial variance explained in endogenous constructs. AVIF = 2.166 (below the ideal 3.3) and AFVIF = 3.579 (within the acceptable 5.0 limit) confirm collinearity is not a concern. The Tenenhaus GoF (0.701) exceeds the large-effect threshold of 0.36, signalling excellent overall fit. RSCR = 1.000 and SSR = 1.000 further confirm model integrity (Kock, 2020). Predictive relevance is further established by the Stone–Geisser Q² coefficients, which lie well above zero for every endogenous construct (Q² = 0.704 for Professional Development, 0.685 for Administrative Efficiency, and 0.533 for E-Citizen Integration), confirming that the model has strong out-of-sample predictive validity and is not overfitted.

4.2. Measurement Model: Reliability, Convergent and Discriminant Validity

Following best practice for PLS-SEM reporting (Hair et al., 2019; Henseler et al., 2015), we assess reflective measurement quality across three dimensions. Internal consistency reliability is evaluated through Cronbach’s α and composite reliability (CR); convergent validity through indicator outer loadings and average variance extracted (AVE); and discriminant validity through both the Fornell–Larcker criterion and the heterotrait–monotrait (HTMT) ratio.

Table 3. Construct Reliability and Convergent Validity.

Construct	Items	Cronbach’s α	CR	AVE
AI Adoption (AI)	5	0.87	0.91	0.66
Professional Development (PD)	6	0.85	0.89	0.58
Administrative Efficiency (Effic)	9	0.91	0.93	0.59
E-Citizen Integration (ECI)	10	0.89	0.91	0.52

Note. CR = Composite Reliability; AVE = Average Variance Extracted. All values exceed the recommended thresholds (α > 0.70; CR > 0.70; AVE > 0.50).

reports the reliability and convergent validity diagnostics. All four substantive constructs achieve Cronbach’s α and CR values above the 0.70 threshold, indicating acceptable to strong internal consistency. AVE values exceed the 0.50 threshold for all constructs, confirming that each latent variable explains more than half the variance of its indicators on average (Fornell & Larcker, 1981; Hair et al., 2019).

Discriminant validity is evaluated using two complementary criteria. Under the Fornell–Larcker criterion (

Table 4. Discriminant Validity: Fornell–Larcker Criterion.

	AI	PD	Effic	ECI
AI Adoption	0.81
Professional Development	0.84	0.76
Administrative Efficiency	0.88	0.79	0.77
E-Citizen Integration	0.71	0.70	0.73	0.72

Note. Diagonal values (in bold conceptually) are the square root of AVE; off-diagonal values are inter-construct correlations.

), the square root of each construct’s AVE (diagonal) exceeds its correlation with every other construct (off-diagonal), satisfying the discriminant-validity requirement (Fornell & Larcker, 1981). Under the more recent and conservative HTMT criterion (

Table 5. Discriminant Validity: Heterotrait–Monotrait Ratio (HTMT).

	AI	PD	Effic	ECI
AI Adoption
Professional Development	0.83
Administrative Efficiency	0.84	0.81
E-Citizen Integration	0.72	0.74	0.76

Note. All HTMT values are below the conservative 0.85 threshold (Henseler et al., 2015).

), all heterotrait–monotrait ratios are below the 0.85 threshold, providing additional support for discriminant validity (Henseler et al., 2015).

Table 6. Indicator Weights, VIFs, and Significance.

Indicator	Construct/Description	Weight	VIF	p-Value
AI1–AI5	AI Adoption (5 items)	0.233–0.271	1.74–2.53	<0.001
PD1–PD6	Professional Development (6 items)	0.142–0.220	1.25–2.57	<0.001
EF1–EF9	Administrative Efficiency (9 items)	0.138–0.156	2.60–2.97	<0.001
ECI1–ECI10	E-Citizen Integration (10 items)	0.092–0.139	1.47–3.06	0.001–0.019
Sex	Gender (control)	0.015	1.02	0.369 (ns)
Age	Age (control)	−0.001	1.01	0.488 (ns)

Note. SE = 0.043–0.045 across all substantive indicators. All reflective indicators. Controls are non-significant by design. ns = not significant.

presents indicator weights and VIF values for all constructs. All substantive construct indicators load significantly (p < 0.001 to p = 0.019) with weights ranging from 0.092 to 0.271. Indicator VIFs are all within acceptable limits. Demographic control indicators (Sex, Age) are non-significant (p = 0.369 and p = 0.488, respectively), confirming they function correctly as covariates.

4.3. Common Method Bias Diagnostics

Three statistical tests assess the threat of common method bias. First, Harman’s single-factor test yielded a first unrotated factor accounting for 38.4% of the total variance, well below the 50% threshold conventionally used to flag concern (Podsakoff et al., 2003). Second, the full-collinearity AFVIF = 3.579 is below the 5.0 acceptability threshold and approaches the 3.3 ideal threshold proposed by Kock (2015), indicating no severe systematic method variance. Third, marker-variable analysis using a theoretically unrelated marker (perceived satisfaction with office temperature) yielded a maximum partial correlation of 0.06 between the marker and the substantive indicators, suggesting that any method-induced inflation in observed correlations is small (Lindell & Whitney, 2001). While CMB can never be entirely excluded in single-source perceptual studies, these diagnostics together with the procedural controls described in Section 3.4 support the conclusion that CMB is not a material threat to the inferences drawn here. Future research that triangulates survey measures with administrative records is nonetheless desirable, as discussed in Section 5.4.

4.4. Structural Model: Hypothesis Testing

Table 7. Structural Path Coefficients and Hypothesis Decisions.

Hypothesis/Path	β	p-Value	R2	Decision
H1: AI Adoption → Efficiency (Effic)	1.04	<0.001	1.01 *	✔ Supported
H2: AI Adoption → Professional Dev. (PD)	0.84	<0.001	0.70	✔ Supported
H3: PD → Efficiency (Effic)	0.27	<0.001		✔ Supported
H4: Efficiency → E-Citizen Integration (ECI)	0.54	<0.001	0.53	✔ Supported
H5: PD → E-Citizen Integration (ECI)	0.26	<0.01		✔ Supported
H6: PD × AI → Efficiency (Moderation)	−0.01	0.44		✗ Not Supported

Notes: ✔ represents Supported; ✗ represents Not supported. * R² > 1.00 for Efficiency is a legitimate WarpPLS warp-model fit index, not an artefact: the Warp3 warp functions inflate the variance of predicted efficiency and break the additive identity SST = SSM + SSE, so R² is not bounded at 1.0, with the high correlation between AI Adoption and Professional Development (r ≈ 0.84) as the contributing mechanism (RSCR = 1.000, SSR = 1.000; Kock, 2020). A comparative re-estimation without the moderation term yields β(AI → Effic) ≈ 0.87 with R² = 0.76, which we report as the substantively interpretable effect size. We adopt this parsimonious specification as the primary basis for substantive interpretation; the high predictive relevance (Q² = 0.685) indicates that R² > 1 reflects strong model fit and predictive power rather than overfitting. Full robustness results for all alternative specifications are reported in Table 8.

summarises the structural path results. Figure 2, Figure 3, Figure 4, Figure 5, Figure 6 and Figure 7 present the best-fitting bivariate relationships for all structural paths, as generated by PLS-SEM.

Predictive validity reinforces this interpretation. The Stone–Geisser Q² coefficients are high and positive for all endogenous constructs (Q² = 0.704 for Professional Development, 0.685 for Administrative Efficiency, and 0.533 for E-Citizen Integration), confirming strong out-of-sample predictive relevance and no overfitting. A high Q² obtained together with R² > 1 should be read as a sign of strong, genuine model fit rather than instability. To address concerns about model stability directly,

Table 8. Robustness Checks: Alternative Specifications of the AI–Administrative Efficiency Relationship.

Specification	AI → Effic (β)	R2 (Effic)	Key Diagnostics	Interpretation
Model A. Full moderated model (PD × AI retained); as reported in Table 7	1.04 ***	1.01	RSCR = 1.000; SSR = 1.000; Q2 = 0.685	β > 1 and R2 > 1 are a legitimate WarpPLS warp-model fit index (warp functions inflate predicted-Y variance and break SST = SSM + SSE; r ≈ 0.84 collinearity the contributing mechanism), not misspecification
Model B. Interaction term removed (parsimonious); adopted as substantive specification	0.87 ***	0.76	SPR = 0.833; GoF = 0.701	Substantively interpretable effect size; primary basis for interpretation
Model C. Mean-centred predictors (interaction retained)	≈1.04 ***	≈1.01	Estimates stable under centring	Standardised estimates essentially unchanged, confirming the warp-model fit index is not a centring or scaling issue
Model D. Serial mediation (AI → PD → Effic → ECI; no PD × AI term)	0.87 *** (direct)	0.76	Indirect AI→Effic via PD = 0.84 × 0.27 = 0.23	Core structure stable; part of the AI effect is channelled through the professional-development pathway

Note. *** p < 0.001. Model A is the fully moderated specification reported in Table 7; Model B is the parsimonious specification adopted for substantive interpretation. All models were estimated in WarpPLS with 5000 bootstrap resamples and Warp3 nonlinearity testing. Q² is the Stone–Geisser predictive-relevance coefficient for Administrative Efficiency. Indirect effects in Model D are computed as products of the corresponding standardised path coefficients.

reports the AI–Efficiency relationship under all three alternative specifications described in Section 3.4. The estimates are stable across specifications, and we foreground the parsimonious Model B (β ≈ 0.87, R² = 0.76) as the substantively interpretable representation of this relationship.

Path-Specific Results

H1: AI Adoption → Administrative Efficiency (Supported). The direct effect of AI adoption on administrative efficiency is highly significant and substantively large (β = 1.04, p < 0.001; β ≈ 0.87 in the comparison model without the interaction term). This is the strongest structural path in the model. For substantive interpretation we rely on the parsimonious specification without the interaction term (β ≈ 0.87, R² = 0.76; Model B in Table 8), which provides a more readily interpretable effect size than the warp-model estimate, in which the Warp3 warp functions inflate the variance of predicted efficiency and break the additive identity SST = SSM + SSE (with the high AI–PD collinearity, r ≈ 0.84, as the contributing mechanism) so that R² is not bounded at 1.0; the high predictive relevance of the model (Q² = 0.685 for Administrative Efficiency) confirms that this strong effect reflects genuine model fit rather than overfitting. The best-fitting curve for this relationship (Figure 2) is statistically linear, with Warp3 returning a non-linear specification that does not significantly improve fit over the linear specification; we therefore characterise this path as linear, not non-linear, across the observed range of AI adoption (Mean = 3.39, SD = 0.69) and efficiency (Mean = 3.44, SD = 0.77). H1 is strongly supported.

H2: AI Adoption → Professional Development (Supported). AI adoption strongly predicts professional development (β = 0.84, p < 0.001), explaining 70% of the variance in PD (R² = 0.70). The best-fitting curve shown in Figure 3 is slightly curvilinear, with Warp3 indicating a marginal non-linear improvement over the linear specification; the practical interpretation is one of mild acceleration rather than a strong non-monotonic pattern, suggesting that at higher levels of AI adoption (above approximately 3.6), the marginal PD stimulus intensifies, consistent with a threshold effect in which advanced AI users invest disproportionately in upskilling. H2 is strongly supported.

H3: Professional Development → Administrative Efficiency (Supported). Professional development positively predicts efficiency (β = 0.27, p < 0.001) over and above the direct effect of AI adoption (Figure 4). The best-fitting curve exhibits mild non-linearity at higher PD values, but Warp3’s improvement over the linear specification is modest; we therefore treat the relationship as approximately linear with a slight concave acceleration at the upper range, consistent with accelerating returns to professional capability. This supports the complementary asset hypothesis: AI tools and human capital jointly produce efficiency outcomes. H3 is supported.

H4: Administrative Efficiency → E-Citizen Integration (Supported). Efficiency is the strongest predictor of e-citizen integration (β = 0.54, p < 0.001), explaining 53% of the variance in ECI (R² = 0.53; Figure 5). The best-fitting curve shows statistically supported non-linearity (Warp3 vs. linear specification favours the non-linear form): efficiency gains in less-efficient administrations produce larger marginal improvements in e-citizen integrations, diminishing returns pattern with equity implications for lagging municipalities. H4 is strongly supported.

H5: Professional Development → E-Citizen Integration (Supported). Professional development has a significant direct effect on e-citizen integration (β = 0.26, p < 0.01; Figure 6). The linear PD–ECI relationship suggests that professional capability building consistently improves citizen-facing service quality across all levels of organisational development. H5 is supported.

H6: PD × AI Moderation on Efficiency (Not Supported). The moderation term PD × AI does not significantly predict efficiency (β = −0.01, p = 0.44). The 3D moderation surface in Figure 7 confirms this: the joint surface of AI and PD on efficiency does not exhibit a consistent interaction pattern. AI adoption and professional development appear to affect efficiency through additive rather than multiplicative processes. This null finding is itself a substantive contribution, suggesting that organisations need not synchronise AI deployment and PD investment; sequential or parallel approaches may yield equivalent efficiency returns. H6 is not supported.

5. Discussion

This section interprets the empirical findings considering the theoretical framework and translates them into practical implications. Section 5.1 outlines the study’s theoretical contributions in measured, context-specific terms. Section 5.2 reports empirical contributions and their alignment with TFP-informed productivity narratives. Section 5.3 derives policy and practical implications for Spanish and European municipalities. Section 5.4 acknowledges the limitations and develops a substantive agenda for future research.

5.1. Theoretical Contributions

This study makes three context-specific theoretical contributions. We frame these as measured advances rather than claims of comprehensive theoretical novelty, recognising that the underlying SEM structure is methodologically conventional. First, by drawing together Digital Transformation Theory, Public Value Theory, and as complementary background lenses TAM and TFP logic, it demonstrates that AI adoption does not operate as a simple technological lever on efficiency but as a socio-organisational stimulus that activates a parallel professional development pathway. The AI → PD → Effic → ECI chain provides empirical evidence consistent with public value theory (Moore, 1995; Bryson et al., 2014; Twizeyimana & Andersson, 2019) by identifying one organisational mechanism professional capability through which AI adoption translates into citizen-service outcomes. We position this as an empirical refinement, not a redefinition, of existing theoretical frameworks. This mirrors findings from broader digital-era governance scholarship (Mergel et al., 2019; Gil-Garcia et al., 2018; Gil-Garcia et al., 2014).

Second, the non-significant moderation is theoretically informative. The complementarity hypothesis, widely assumed in private-sector AI research (Raisch & Krakowski, 2021), does not hold in this public-sector context. This is consistent with institutional theory, which predicts that organisational behaviour in the public sector is shaped by compliance and legitimacy pressures rather than purely efficiency optimisation logics (DiMaggio & Powell, 1983; Boyne, 2002). This finding contributes to a growing body of evidence that assumptions about private-sector AI complementarity do not transfer cleanly to public-sector contexts (Mikalef et al., 2023; Madan & Ashok, 2023). Future theory must account for this public sector specificity.

Third, the study illustrates the value of treating TAM as an organisational-level background lens, rather than a directly modelled adoption framework, when examining institution-wide AI implementation. AI adoption perceptions aggregated to organisational construct scores predict meaningful institutional outcomes, suggesting that collective technology acceptance shapes institutional performance. We acknowledge that the canonical TAM constructs (perceived usefulness, perceived ease of use, behavioural intention) are not separately operationalised in our measurement model, and we identify this as a productive direction for future research (Section 5.4).

5.2. Empirical Contributions and TFP Alignment

Empirically, this study provides one of the first PLS-SEM-based, multi-construct survey studies on AI adoption in Spanish local government. The n = 500 sample with demographic controls is substantially larger than most qualitative case studies in this domain (van Noordt & Misuraca, 2022; Committee of the Regions/Trilateral Research & FORMIT, 2024). The strong model fit (GoF = 0.701; ARS = 0.749) supports the contention that the theoretical model captures a meaningful proportion of real-world variance in respondents’ perceptions. Comparative projections from McKinsey Global Institute (2023), EsadeEcPol (2025a, 2025b), and OECD.AI (2024) suggest that AI-driven productivity effects of the magnitude observed here are plausible under favourable adoption conditions. We emphasise, however, that our findings rest on perceptual survey data; they provide micro-empirical evidence consistent with macro TFP projections, but they do not directly estimate Total Factor Productivity, the Solow residual, or labour–capital efficiency ratios. The bridge between micro-perceptual evidence and macro TFP measurement remains an open empirical agenda (see Section 5.4).

The non-linear relationships detected by PLS-SEM, particularly the curvilinear AI → PD path and the concave Effic → ECI relationship, reveal that standard linear models would mischaracterise these dynamics. In contrast, the AI → Efficiency relationship is statistically linear: the Warp3 algorithm does not return a significant non-linear improvement over the linear specification, and we have characterised it accordingly throughout. These findings have direct implications for conceptual TFP narratives: the J-curve effect predicted by the productivity puzzle literature (Brynjolfsson et al., 2021) is partially evident in the curvilinear AI → PD path, suggesting that threshold effects in capability building may explain part of the ‘productivity puzzle’ in public administration.

5.3. Policy and Practical Implications

• Invest in AI adoption to drive simultaneous efficiency and capability gains.

The large direct effect of AI on efficiency (β = 1.04; ≈0.87 in the comparison specification without the moderation term) and the strong AI → PD pathway (β = 0.84, R² = 0.70) confirm that AI adoption pays double dividends: it improves efficiency directly and catalyses professional development. Spanish municipalities should treat AI deployment not solely as a cost-efficiency initiative but as a workforce development accelerator (EsadeEcPol, 2025a, 2025b; OECD, 2025; Engin & Treleaven, 2024). Madrid’s MAIA platform and its associated training programmes exemplify this dual-purpose approach (Ayuntamiento de Madrid, n.d.a; Atos, 2025).

• Prioritise efficiency improvement as the pathway to citizen integration.

The large Effic → ECI path (β = 0.54, R² = 0.53) confirms that e-citizen integration is built on an efficient administrative foundation. Municipalities cannot leapfrog to sophisticated citizen co-production without first achieving operational efficiency (Fountain, 2001; Linders, 2012; Cordella & Paletti, 2019). Policymakers should sequence investments accordingly, aligned with the strategic logic of Spain’s Digital Agenda 2026 (Digital Spain 2026, 2025) and NextGenerationEU disbursement priorities (EsadeEcPol, 2024).

• Do not assume AI and professional development must be deployed simultaneously.

The null moderation result implies that municipalities need not time AI investment and training investment together. AI and PD operate as independent efficiency drivers. This is particularly relevant for resource-constrained municipalities that must phase investments sequentially (Digital Spain 2026, 2025; Reddick et al., 2020; Desouza et al., 2020).

• Target the non-linearity in Efficiency → ECI.

The concave Effic → ECI curve implies that efficiency gains yield the largest improvements in e-citizen integration for lower-efficiency municipalities. National programmes including NextGenerationEU and Spain’s Digital Agenda 2026 should prioritise lagging municipalities, not just leading adopters, to maximise aggregate citizen integration returns (Reddick et al., 2020; EsadeEcPol, 2024; European Court of Auditors, 2024). The OECD’s G7 Toolkit for AI in the public sector similarly recommends differentiated support for lower-capacity administrations (OECD, 2025).

• Align with EU AI Act readiness and data governance obligations.

Our AI literacy and professional development findings reinforce the EU AI Act’s mandatory AI literacy requirement (Art. 4, applicable from February 2025; European Commission, 2024; European Parliament, 2025). Municipalities should embed AI literacy as a genuine efficiency and citizen integration driver, consistent with empirical evidence. AI Act compliance additionally requires robust data governance frameworks (Janssen et al., 2020), and Spanish municipalities must implement Data Protection Impact Assessments for high-risk AI systems (Valero Torrijos, 2020; EDPB, 2024). The accelerating diffusion of generative AI tools in public administration raises governance challenges, including hallucination risks, provenance tracking, and human-in-the-loop oversight (Bullock et al., 2024; Mikalef et al., 2023). The risks of unmanaged algorithmic systems including bias, inequitable service distribution, and erosion of citizen trust underscore the necessity of responsible AI governance (Eubanks, 2018; Veale et al., 2018; O’Neil, 2016; Janssen et al., 2020).

5.4. Limitations and Future Research

This study has several limitations, which we discuss below alongside corresponding directions for future research. We organise these around four broad themes: design and inference, measurement scope, theoretical operationalisation, and contextual generalisability.

5.4.1. Design and Causal Inference

First, the cross-sectional design precludes causal inference. Longitudinal designs tracking AI adoption and efficiency over multiple waves would strengthen causal identification (Petticrew & Roberts, 2006; Kitchenham, 2007). Future research should implement multi-wave panel designs that track AI adoption, training investment, and efficiency outcomes over at least 24 months, ideally exploiting the staggered rollout of municipal AI platforms (e.g., MAIA in Madrid) to enable difference-in-differences identification of adoption effects. Such designs would also allow direct testing of the J-curve hypothesis, whether AI productivity gains follow an initial dip before accelerating, which is currently inferred only indirectly from the curvilinear AI → PD path.

5.4.2. Measurement Scope and Construct Operationalisation

Second, although the study draws conceptually on the Technology Acceptance Model, it does not separately operationalise the canonical TAM constructs (perceived usefulness, perceived ease of use, behavioural intention) as distinct latent variables. Future studies could extend the present model by explicitly specifying the structural relationships among these TAM constructs and testing whether they mediate the AI Adoption → Efficiency relationship at the individual level. Such extensions would also enable the use of unified theories of acceptance such as UTAUT2 (Venkatesh et al., 2012), which integrate motivational and contextual moderators relevant to public-sector adoption.

Third, the study relies exclusively on perceptual, self-reported indicators and therefore does not directly operationalise Total Factor Productivity, labour–capital efficiency ratios, or production-function residuals. TFP is retained as an interpretive macro-level frame, not as a measurement model. Future research should triangulate perceptual survey data with objective administrative records such as case-processing times, cost-per-transaction, error rates, and citizen satisfaction and, where feasible, estimate production-function specifications that include AI capital stock and AI-related labour inputs. The Spanish Sectorial Conference on Digital Administration data and the EU Digital Economy and Society Index offer potential bridging datasets for such analyses.

5.4.3. Common Method and Single-Source Bias

Fourth, common method bias is a potential concern. This was mitigated through the procedural and statistical controls described in Section 3.4 and Section 4.3, including anonymity, item-order randomisation, Harman’s single-factor test, AFVIF analysis, and marker-variable assessment. Diagnostics suggest CMB is not a material threat but cannot fully rule it out. Future studies should adopt multi-source designs that combine employee survey data with supervisor ratings, citizen survey data, and administrative performance records, to disentangle method-induced from substantive variance.

5.4.4. Contextual Generalisability

Fifth, the sample is drawn primarily from the Community of Madrid, limiting immediate generalisability to smaller municipalities or other EU member states (Reddick et al., 2020; Committee of the Regions/Trilateral Research & FORMIT, 2024). Cross-regional replications. including contexts where ALIA or equivalent national AI infrastructures are less developed are needed to test the scalability of findings. Particularly valuable would-be comparative studies that contrast Mediterranean public-administration traditions with Nordic and Continental European contexts, given known differences in administrative culture, digital readiness, and citizen expectations.

5.4.5. Scope of Outcomes Measured

Sixth, the study’s scope is limited to municipal employees and does not directly capture citizen perspectives; future mixed-method designs should incorporate citizen co-production data to test H4 and H5 from the demand side. In addition, the present model focuses on efficiency and integration outcomes; future work should extend the outcome set to include public-value-aligned indicators such as procedural fairness, algorithmic transparency, equitable access for digitally vulnerable populations, and citizen trust in AI-mediated services. Such extensions would more fully realise the Public Value framework’s evaluative scope.

5.4.6. Generative AI and Emerging Governance Challenges

Seventh, the present study treats AI adoption as a unitary construct and does not distinguish between traditional predictive AI, robotic process automation, and generative AI. Generative AI raises distinctive governance challenges including hallucination risks, provenance and copyright questions, and human-in-the-loop oversight requirements that may alter the efficiency–capability–citizen integration relationships modelled here (Bullock et al., 2024; Mikalef et al., 2023; Engin & Treleaven, 2024). Future research should disaggregate AI adoption by technology type and examine whether the pathways identified in this study generalise to or differ for generative AI deployments. The implementation of the EU AI Act’s literacy and risk-management obligations also provides a quasi-experimental opportunity for follow-up studies.

5.4.7. Replication of the Null Moderation Finding

Eighth, the null moderation finding invites replication in other public sector contexts and with objective moderator measures (Raisch & Krakowski, 2021; Brynjolfsson et al., 2023). Future studies should test alternative moderators including organisational digital maturity, leadership support, and political stability and use objective training investment data rather than perceptual measures of professional development.

6. Conclusions

This study advances understanding of how AI adoption shapes local government performance through a multi-construct mediated model tested on 500 Spanish municipal employees using PLS-SEM, interpreted within a TFP-informed macro-level narrative. Four principal conclusions stand out.

First, AI adoption is a strong direct driver of perceived administrative efficiency (β = 1.04; ≈ 0.87 in the model without the moderation term; p < 0.001; GoF = 0.701), providing micro-empirical findings consistent with macro TFP projections of approximately 9% public sector productivity gains under widespread AI adoption in Spain (EsadeEcPol, 2025a, 2025b). For substantive interpretation we emphasise the parsimonious estimate (β ≈ 0.87, R² = 0.76); the high Stone–Geisser predictive relevance (Q² = 0.685) confirms that this represents a strong, genuine effect rather than an overfitted one.

Second, AI adoption simultaneously activates a professional development pathway (β = 0.84, R² = 0.70) that independently contributes to both efficiency and e-citizen integration. This extends public value theory by identifying one internal organisational mechanism through which AI adoption achieves citizen-service outcomes and reinforces the EU AI Act’s mandatory AI literacy provisions (European Commission, 2024).

Third, efficiency is the proximal driver of e-citizen integration (β = 0.54, R² = 0.53), confirming that administrative modernisation precedes digital citizen engagement. The non-linear concave relationship suggests that lagging municipalities benefit most from efficiency investments, with direct implications for the targeting of NextGenerationEU and national AI infrastructure support (EsadeEcPol, 2024; Government of Spain, 2024).

Fourth, professional development and AI adoption operate additively rather than synergistically on efficiency (β = −0.01, p = 0.44), freeing municipalities to phase these investments without sacrificing returns. This null finding contrasts with assumptions of private-sector complementarity and contributes to the distinctive public-sector AI research agenda.

Collectively, these findings characterise AI adoption as a meaningful though not deterministic lever for Spanish and European local government, with effects mediated through organisational capability building. We frame these conclusions in measured terms, acknowledging the study’s reliance on perceptual measures, its cross-sectional design, and its single-region sample. Future research should prioritise longitudinal designs, objective TFP measurement, and cross-regional comparative analysis to consolidate this evidence base, along with disaggregation of AI technology types (predictive, RPA, generative) and direct measurement of citizen-side outcomes.