How Agentic AI Reshapes Work: The Roles of Cognitive Surplus, Psychological Safety, and AI-Human Task Reallocation

Summary

This protocol describes a survey-based workflow for examining how Agentic AI technological characteristics relate to cognitive surplus, psychological safety, and AI-human task reallocation among 331 employees. Partial least squares structural equation modeling with multi-group analysis reveals role-based differences between leaders and staff.

Abstract

This study examines how the technological characteristics of Agentic AI relate to employees' psychological states and behavioral responses in organizations. Drawing on Job Demands-Resources (JD-R) theory and Conservation of Resources (COR) theory, we conceptualize Agentic AI technological characteristics (AITC) as a higher-order construct comprising autonomous planning, multi-tool orchestration, and proactive feedback, and propose that these characteristics function as job resources associated with greater cognitive surplus and psychological safety, which in turn relate to AI-human task reallocation. We further investigate whether these relationships differ between leaders and staff. A cross-sectional, anonymous online survey was administered via an online survey platform to participants recruited through a research panel using purposive nonprobability sampling; eligibility was restricted to adults with current or prior AI-related work experience, and screened, attention-checked responses yielded a final sample of 331 employees (171 leaders, 160 staff). All constructs were measured with validated scales adapted to the Agentic AI context on 7-point Likert scales. Data were analyzed using partial least squares structural equation modeling (PLS-SEM), applying a two-stage approach for the second-order construct, bootstrapping with 5,000 resamples for direct and indirect effects, and multi-group analysis (MGA) for leader-staff comparisons. The measurement model demonstrated acceptable reliability and validity, and the structural results were consistent with the theorized resource-gain model: AITC was positively associated with cognitive surplus, psychological safety, and task reallocation, with significant mediation pathways and stronger AITC-resource associations among leaders. By focusing on psychological mechanisms rather than technical performance, this study extends emerging Agentic AI research and offers practical implications for role-differentiated AI adoption strategies.

Introduction

The proliferation of Agentic AI is redefining the role of artificial intelligence in industries and organizational life^1,2. Agentic AI can be defined as an advanced form of artificial intelligence that perceives changing environments, sets goals independently through reasoning mechanisms, and orchestrates sequences of actions with minimal human oversight^3,4. It is distinguished from traditional AI, including conventional generative AI tools, in three respects: autonomy, because agentic systems pursue objectives with minimal human intervention rather than executing pre-programmed rules^5,6; core function, because their purpose is to automate complex workflows rather than single tasks⁷; and architecture, because they coordinate multiple large language models, application programming interfaces (APIs), and agents rather than relying on a single model⁷. These capabilities rest on the defining technological characteristics examined in this study^8,9. Autonomous planning (AP) means that, given a user's goal, the system independently decomposes it into sub-tasks and revises its plan as constraints change¹⁰; for example, an employee can provide a one-line instruction and receive a workable multi-step project plan. Multi-tool orchestration (MTO) means that the system selects, sequences, and integrates external tools and APIs^1,11,12; for example, combining web search, document editors, and data-analysis tools to produce a single coherent report. Proactive feedback (PF) means that the system flags errors, missing information, and risky assumptions before being asked^10,13; for example, reminding an employee of validation or compliance checkpoints during task execution. We focus on these three characteristics, collectively termed Agentic AI technological characteristics (AITC), because they are the functional capabilities that differentiate agentic from traditional AI in everyday work^7,10; attributes such as transparency, controllability, and explainability matter for AI trust in general but are not specific to the agentic paradigm, and we therefore treat them as boundary conditions for future research rather than focal dimensions.

Despite the rapid diffusion of Agentic AI in workplaces, the extant literature has primarily addressed its technical architecture, operational efficiency, and task-automation capabilities, while paying comparatively little attention to how these characteristics reshape employees' psychological states and behavioral responses^14,15. This omission matters because reallocating tasks between humans and AI is not a purely technical adjustment but a psychologically and organizationally embedded process: employees must challenge existing workflows, redefine role boundaries, and accept new forms of collaboration with intelligent systems^16,17. Such changes do not occur automatically simply because advanced technology is available. Accordingly, the purpose of this study is threefold: to examine whether AITC function as job resources associated with cognitive surplus, psychological safety, and AI-human task reallocation; to investigate the mediating and sequentially mediating roles of cognitive surplus and psychological safety; and to explore whether these structural relationships differ between leaders and staff. The novelty of this study is threefold. First, it conceptualizes and operationalizes Agentic AI technological characteristics—autonomous planning, multi-tool orchestration, and proactive feedback—as a higher-order job resource within the JD-R framework, rather than treating AI adoption as an undifferentiated context. Second, it introduces AI-human task reallocation as a domain-specific, technology-enacted form of task crafting, distinct from both generic job crafting and top-down digital-transformation outcomes. Third, it provides, to our knowledge, the first multi-group evidence that the resource-generating associations of Agentic AI differ systematically between leaders and staff.

The theoretical rationale integrates Job Demands-Resources (JD-R) theory and Conservation of Resources (COR) theory. JD-R theory holds that every job comprises job demands, the physical, emotional, and cognitive aspects of work that require sustained effort, and job resources, the aspects that reduce demands, support goal achievement, and stimulate personal development; their balance shapes engagement, well-being, and performance^18,19. COR theory adds a dynamic mechanism: individuals strive to obtain, retain, and protect valued resources, and accumulated resources initiate gain spirals in which existing resources are reinvested to acquire further resources^20,21,22. From this joint perspective, AITC can be conceptualized not merely as technical features but as a new form of job resource: by absorbing repetitive and cognitively taxing tasks such as planning and diagnostic analysis, agentic systems reduce demands and conserve employees' cognitive energy and time^15,23. We define cognitive surplus (CS) as the reservoir of cognitive energy and discretionary time that becomes available when such routine tasks are delegated to technology, and that can be reinvested in collective participation, collaboration, and the creation of new value²⁴. CS is thus conceptually distinct from work engagement, which is an affective-motivational state, and from mere time availability, because it denotes a convertible stock of personal resources in the COR sense^20,24. CS is likewise distinct from psychological detachment and recovery, which concern the restoration of depleted energy away from work, and from time affluence, which denotes the subjective perception of having enough time; CS instead refers to freed cognitive capacity that remains available for reinvestment during work itself. In COR terms, accumulated CS represents a resource gain that should support occupational well-being by buffering exhaustion and burnout, whereas its absorption by additional routine demands would signal resource loss^20,23,24. Whether conserved time actually converts into surplus may depend on individual characteristics such as AI literacy, on workload norms that can absorb saved time with additional demands, and on organizational culture; we therefore treat the resource-gain pathway as a hypothesis to be tested rather than an assumption.

H1: AITC is positively associated with CS.

Psychological safety (PS) is defined as the belief that one's work environment is safe for interpersonal risk-taking, such as admitting mistakes, asking for help, and challenging the status quo^25,26. Although psychological safety has often been studied as a team-level climate, this study conceptualizes and measures it as an individual-level perception, consistent with our individual-level survey design^27,28. When agentic systems provide learning-oriented feedback and supportive tools, the interpersonal cost of experimenting and exposing errors may decline, because employees can test and correct their work before presenting it to others^26,28. We acknowledge, however, that the opposite dynamic is plausible: where agentic systems are deployed for surveillance or fault-finding accountability, perceived interpersonal risk and mistrust may increase^29,30. Because the characteristics examined here are assistive rather than evaluative in nature, we hypothesize a positive association while treating this two-sided possibility as an important boundary condition.

H2: AITC is positively associated with PS.

We define AI-human task reallocation (AHTR) as employees' self-initiated redistribution of tasks between themselves and AI systems, delegating routine work to AI while expanding their own strategic and creative responsibilities. We position AHTR as a domain-specific form of job crafting. In the JD-R-based model of Tims and Bakker, job crafting comprises increasing job resources, increasing challenge demands, and decreasing hindrance demands^17,31; in the typology of Wrzesniewski and Dutton, it comprises task, relational, and cognitive crafting³². AHTR corresponds most closely to task crafting enacted through technology: offloading routine tasks to AI simultaneously decreases hindrance demands and mobilizes a new job resource^17,23,32. It is narrower than a general digital-transformation outcome because it captures discretionary, employee-initiated changes in task boundaries rather than top-down organizational redesign³³. We privilege task reallocation over relational and cognitive crafting because the redistribution of tasks is the most immediate behavioral interface between employees and agentic systems and is directly observable in early adoption, whereas relational and cognitive crafting typically emerge as downstream adaptations^31,32. Unlike prior studies that treat AI adoption as a contextual antecedent of general job crafting^16,34, we model the reallocation itself as the focal crafting behavior. Because such reallocation disrupts established workflows, employees are more likely to enact it when they possess sufficient cognitive resources to redesign their work and feel psychologically safe to take the associated interpersonal risks^16,34. Within the gain-spiral logic of COR theory, psychological safety also functions as a contextual resource that protects and amplifies the accumulation of cognitive surplus²².

H3: CS is positively associated with AHTR.

H5: PS is positively associated with AHTR.

H6: PS is positively associated with CS.

H9: AITC is positively associated with AHTR.

Integrating these arguments, AITC is expected to relate to AHTR not only directly but also indirectly, through the resource-gain pathways of cognitive surplus and psychological safety.

H4: CS mediates the relationship between AITC and AHTR.

H7: PS mediates the relationship between AITC and AHTR.

H8: PS and CS sequentially mediate the relationship between AITC and AHTR.

Finally, employees do not interpret and respond to Agentic AI uniformly. Gibson, Cooper, and Conger described perceptual distance between leaders and their teams as cognitive discrepancy regarding the same organizational objectives, rooted in asymmetries of information access, accountability, risk exposure, and work experience³⁵. Building on this view, leaders and staff occupy structurally different positions with respect to AI adoption: leaders carry strategic accountability and tend to frame AI as an augmentation-oriented resource for productivity and data-driven decision-making, whereas staff are relatively more exposed to automation-related job insecurity and learning overload^20,30. We treat these patterns as tendencies to be tested empirically rather than fixed dispositions of either group. Such role-based discrepancies may generate heterogeneous patterns in the relationships among AITC, CS, PS, and AHTR, yet empirical research that explicitly captures these group differences remains limited.

H10a: The associations of AITC with CS and PS differ between leaders and staff.

H10b: The associations of PS and CS with AHTR differ between leaders and staff.

The protocol below describes the complete workflow, from participant recruitment and screening through survey administration, data cleaning, measurement validation, structural model estimation, mediation analysis, and multi-group analysis (MGA) (Figure 1), so that the procedure can be independently replicated. The research model is depicted in Figure 2: AITC, modeled as a second-order construct comprising AP, MTO, and PF, relates to AHTR directly and indirectly through CS and PS, with PS also relating to CS; the leader-staff role difference is examined through MGA.

Figure 1. Overall study-design workflow. Schematic of the complete research protocol. Participants were recruited via the online panel using purposive nonprobability sampling and screened for AI-related work experience; eligible respondents completed an anonymous survey measuring all constructs on 7-point Likert scales. After data cleaning (screening failures, attention-check failures, and implausibly short completion times removed; final N = 331), the measurement model was validated (reliability: Cronbach's α, composite reliability; convergent validity: average variance extracted; discriminant validity: HTMT and Fornell-Larcker), the second-order construct was validated via the two-stage approach, and the structural model, mediation pathways, and leader-staff multi-group analysis were estimated with 5,000 bootstrap resamples. Please click here to view a larger version of this figure.

Figure 2. Proposed research model. Conceptual model of hypothesized relationships. Agentic AI technological characteristics (AITC), a second-order construct comprising autonomous planning (AP; the system's capacity to decompose goals into sub-tasks and revise plans), multi-tool orchestration (MTO; the capacity to select, sequence, and integrate multiple tools and APIs), and proactive feedback (PF; the capacity to flag errors and risks unprompted), is hypothesized to be positively associated with cognitive surplus (CS; the reservoir of cognitive energy and discretionary time freed when routine tasks are delegated to AI), psychological safety (PS; the perception that the work environment is safe for interpersonal risk-taking), and AI-human task reallocation (AHTR; employees' self-initiated redistribution of tasks between themselves and AI) (H1, H2, H9). CS and PS are hypothesized to be positively associated with AHTR (H3, H5), PS with CS (H6), and CS and PS to mediate, individually and sequentially, the AITC-AHTR relationship (H4, H7, H8). Arrows represent hypothesized structural paths; H10a/H10b denote hypothesized leader-staff differences tested via multi-group analysis. Please click here to view a larger version of this figure.

Protocol

All procedures were performed in accordance with the ethical standards for research involving human participants at aSSIST University and Dong-A University. Since all responses were anonymized upon collection and no personal identifiable information was collected at any stage, the Institutional Review Board of aSSIST University determined that the study was exempt from full review pursuant to Article 15 of the Bioethics and Safety Act of the Republic of Korea and Article 13 of its Enforcement Rule. The survey opened with an information and consent page stating that participation was restricted to individuals currently employed and involved in an AI project, describing the study purpose, and explaining the voluntary nature of participation, the anonymity of responses, and the intended use of the data; only respondents who explicitly agreed proceeded to the questionnaire. The consent page is provided in Supplementary File 1.

Sample size determination

An a priori power analysis was conducted before data collection using G*Power 3.1.9.7 (RRID: SCR_013726) for the linear multiple regression family, specifying the largest number of predictors directed at a single endogenous construct in the model; in this study, AHTR is predicted by three constructs (AITC, CS, and PS). With a medium effect size (f² = 0.15), α = 0.05, and statistical power of 0.80, the required minimum sample was 77 cases; the G*Power protocol output is provided in Supplementary File 2. The minimum sample size was cross-checked using the inverse square root method for PLS-SEM³⁶, under which approximately 155 cases are required to detect a minimum standardized path coefficient of approximately 0.20 at the 5% significance level with 80% power. Finally, the recruitment target was set to exceed both minima substantially, in order to support bootstrapping of indirect effects and split-sample multi-group analysis; the final sample (N = 331; leaders n = 171, staff n = 160) satisfied the minimum sample requirement within each subgroup.

Participant recruitment and screening

Participants were recruited through the online research panel using purposive nonprobability sampling; that is, the sample was deliberately restricted to respondents satisfying an a priori substantive criterion—current or prior work experience involving AI—rather than drawn at random from the general population. Eligibility was restricted to respondents aged 18 years or older. AI-related work experience was operationalized through the screening item “Which of the following best describes your current situation?”, presented before the main questionnaire; only respondents who indicated that they were currently employed and currently involved in AI-related work proceeded, whereas the remaining options (employed but not involved in AI work; not employed but previously involved; not employed and not involved) triggered immediate screen-out. Data collection was run in a single session on 8 February 2026 (10:28–13:30), and participants were compensated at GBP 6.98 per hour in accordance with the platform’s fair-pay policy. Duplicate participation was prevented through the platform’s unique participant identifiers and the survey platform’s duplicate-response prevention settings. The number of respondents at each stage was recorded to document the inclusion/exclusion workflow: 537 responses were recorded; 151 were screened out by the screening questions (386 remaining); 40 were removed for failing the attention checks, including straight lining (346 remaining); and 15 were removed for implausibly short completion times (under approximately 180 s), yielding the final sample of N = 331.

Standardized definition and survey instructions

Participants were informed that the survey examined how Agentic AI may affect organizations and their members. At the start of the survey, all participants were shown a brief standardized definition of Agentic AI, consistent with the definition given in the Introduction—an advanced AI system characterized by autonomous planning, multi-tool orchestration, and proactive feedback that performs goal-directed actions with limited human oversight (Figure 1, step 2). Eligibility was restricted to respondents who were currently employed and involved in an AI project, which provided a relevant work-related basis for responding to the questionnaire. No separate comprehension-check item was included; the standardized definition therefore served as a contextual orientation rather than a verified comprehension check. Participants were instructed to answer every item based on their actual experience of using or working with such Agentic AI systems in organizational contexts.

Survey instrument development and administration

The questionnaire was developed from validated measurement scales adapted to the Agentic AI organizational context. Table 1 lists every item, its source scale, and descriptive statistics: AP, MTO, and PF items were adapted from recent Agentic AI conceptualizations^7,8; CS items from generative-AI cognitive-resource research³⁷; PS items from Plomp et al.³⁸; and AHTR items from JD-R-based job crafting scales^17,39 adapted to task reallocation. All constructs were measured on 7-point Likert scales ranging from 1 (“strongly disagree”) to 7 (“strongly agree”). The survey was organized into blocks covering the consent page, screening, AITC items (AP, MTO, PF), CS, PS, AHTR, and demographic/control characteristics; the instrument also contained additional scales (AI literacy and role-breadth self-efficacy) that are beyond the scope of the present model and were collected but not analyzed here, reported for transparency. The survey was hosted on the online survey platform; the back button was disabled so that respondents could not revise earlier answers, and items were presented in a fixed order within blocks. The median completion time was 7 min 44 s. Two instructed-response attention-check items were embedded as survey-flow branches that closed the survey when answered incorrectly: (i) a factual check, “Is AI an abbreviation for Artichoke Information?” (correct answer: No), and (ii) a scale-comprehension check confirming that responses were recorded on a 7-point Likert scale ranging from “strongly disagree” to “strongly agree”. Before the main administration, a pilot test was conducted with 124 respondents to evaluate the clarity of the survey items, the overall survey flow, completion duration, and the feasibility of the questionnaire; completion duration was reviewed to identify unusually short responses and to confirm that the survey length was appropriate for the target respondents. The pilot test did not indicate any wording problems requiring item revision, and no wording changes were made prior to the main survey. In addition, the pilot data were analyzed using PLS-SEM as a preliminary validation procedure to assess indicator performance, construct reliability, convergent validity, discriminant validity, collinearity, and the plausibility of the proposed structural relationships; the results supported the suitability of the instrument for the main survey.

Construct		Item	Question			Mean	STDEV	VIF	Source
AITC	AP	AP1	When I provide a goal, the AI helps create a multi-step plan to reach it.			5.81	0.9	1.44	Abou Ali, M., et al. (2025) ; Sapkota, R., et al. (2025)
		AP2	The AI can revise its plan when constraints or requirements change.			5.75	0.99	1.37
		AP3	The AI breaks complex work into manageable sub-tasks without needing step-by-step instructions.			5.33	1.37	1.33
	MTO	MTO1	The AI can combine multiple tools or apps (e.g., search, documents, data tools) to complete a task.			5.69	1.11	1.5
		MTO2	The AI suggests the right tool(s) and uses them in the correct sequence to produce a deliverable.			5.36	1.04	1.64
		MTO3	The AI helps integrate outputs from different tools into one coherent result.			5.61	1.01	1.63
	PF	PF1	The AI proactively points out errors, missing information, or risky assumptions in my work.			4.9	1.44	1.68
		PF2	The AI suggests improvements before I explicitly ask for feedback.			5.07	1.38	1.85
		PF3	The AI reminds me of checkpoints (e.g., validation, compliance, quality) to prevent rework.			5.02	1.4	1.6
CS		CS1	Using AI has freed up my mental energy to focus on more complex and meaningful work.			5.59	1.28	2.07	Liu, Y., Sheng, F., & Liu, R. (2025)
		CS2	I have more time available for strategic thinking and problem-solving since using AI.			5.61	1.23	2.38
		CS3	The workload from routine tasks has decreased significantly, giving me headspace for new initiatives.			5.42	1.28	2.23
		CS4	I feel less mentally exhausted at the end of the workday because AI handles repetitive tasks.			5.15	1.5	1.91
PS		PS1	In my workplace, it is safe to take a risk related to using AI in new ways.			4.74	1.37	1.52	Plomp et al. (2019)
		PS2	My manager encourages experimentation with AI and is tolerant of mistakes.			4.52	1.61	1.68
		PS3	I feel comfortable speaking up if I have concerns about how AI is being used in my team.			5.48	1.24	1.32
		PS4	My colleagues are supportive if I try new ways of working with AI.			5.26	1.32	1.48
		PS5 (excluded)	I am not afraid of negative consequences if I propose innovative uses of AI in my work.			4.47	1.59	1.23
AHTR		AHTR1	I deliberately transfer routine tasks (e.g., data entry, report formatting) to AI so I can focus on higher-value work.			5.12	1.35	1.68	Tims, M., & Bakker, A. B. (2010); Li, W. et al. (2024)
		AHTR2	I actively seek opportunities to use AI to handle tasks I previously did manually.			5.58	1.28	1.84
		AHTR3	Since using AI, I have taken on new or expanded responsibilities that are more strategic or creative.			5.18	1.3	1.58
		AHTR4 (excluded)	I negotiate with my manager about which tasks AI should handle and which I should do.			4.09	1.65	1.23
		AHTR5	I have redesigned my job by reallocating routine work to AI, freeing myself for more meaningful tasks.			4.92	1.35	1.67

Table 1: Measurement items, sources, descriptive statistics, and full collinearity assessment. Complete survey instrument. Each item is listed with its construct (AITC dimensions AP, MTO, PF; CS; PS; AHTR), item code, wording, mean, standard deviation (STDEV), full collinearity variance inflation factor (VIF; values <3.3 indicate the absence of common method variance contamination), and the source scale from which it was adapted. All items were measured on 7-point Likert scales (1 = strongly disagree, 7 = strongly agree). PS5 and AHTR4 were removed during measurement model evaluation due to low outer loadings. AITC = Agentic AI technological characteristics; AP = autonomous planning; MTO = multi-tool orchestration; PF = proactive feedback; CS = cognitive surplus; PS = psychological safety; AHTR = AI-human task reallocation. Please click here to download this Table.

Data preparation and cleaning

The raw responses were exported in CSV format. Cases failing the screening criterion or either attention-check item were excluded, as were responses with substantial missing data (forced-response settings minimized within-survey missingness); the resulting exclusion counts at each stage are documented in protocol section 2 (537 recorded → 331 retained). The retained data were inspected for consistency and screened for inattentive response patterns: straight-lining responses were removed via the attention checks, and responses with implausibly short completion times (under approximately 180 s) were excluded. Variables were then coded for analysis (e.g., Role: Leader = 1, Staff = 0; Gender: Male = 1, Female = 0) as documented in Table 2, categorical controls such as industry were dummy-coded with k-1 indicators, and the finalized dataset was imported into the PLS-SEM software. A codebook mapping every variable to its item text, coding, and construct is provided as Supplementary File 2.

Measurement model evaluation

Using the PLS-SEM software, each observed item was assigned to its latent construct, and all first-order constructs (AP, MTO, PF, CS, PS, AHTR) were specified as reflective. AITC was specified as a second-order construct comprising AP, MTO, and PF and was estimated with the two-stage approach^40,41,42: latent variable scores were first obtained for the first-order constructs and then used (AP_LVS, MTO_LVS, PF_LVS) as formative indicators of AITC. The PLS algorithm was run with the path weighting scheme, a maximum of 3,000 iterations, and a stop criterion of 10⁻⁷. Internal consistency reliability was evaluated against Cronbach’s α and composite reliability (CR) thresholds of 0.70⁴³, treating α values of 0.5–0.6 as marginally acceptable for short scales in emerging research contexts⁴³; convergent validity was evaluated against the average variance extracted (AVE) threshold of 0.50⁴⁴; and discriminant validity was evaluated using the heterotrait-monotrait ratio (HTMT <0.85, conservative threshold)⁴⁵ and the Fornell-Larcker criterion⁴⁴. Where reliability or convergent validity criteria were not met, outer loadings were inspected, and poorly performing indicators were removed while preserving content validity, with every removal and its rationale documented; in this study, PS5 and AHTR4 were removed due to low outer loadings, which improved the convergent validity of PS and AHTR.

Structural model and mediation analysis

The structural paths were specified in accordance with the research model (Figure 2), and collinearity among predictor constructs was examined using inner-model variance inflation factors (VIF < 5.0)³⁶ and full collinearity VIFs (<3.3)^46,47. Path significance was estimated by bootstrapping with 5,000 resamples using percentile confidence intervals and two-tailed testing, and standardized path coefficients (β), sample means (M), standard deviations (STDEV), t-values, p-values, 95% confidence intervals, R², and effect sizes (f²) were reported. Out-of-sample predictive power was assessed using PLSpredict with 10 folds and 10 repetitions; for each endogenous indicator, Q²predict (>0 indicates predictive relevance) was inspected, and the PLS-SEM prediction errors (RMSE and MAE) were compared against the linear-model (LM) benchmark⁴⁸. For the mediation analysis, the indirect effects of AITC on AHTR through CS, through PS, and through the sequential PS → CS pathway were specified, and each specific indirect effect was evaluated using the bootstrap distribution (5,000 resamples), with an indirect effect judged significant when its 95% confidence interval excluded zero.

Multi-group analysis

Respondents were classified into leader and staff groups based on self-reported organizational role (Role: Leader = 1, Staff = 0; Table 2); the leader group comprised team leads/supervisors, managers/department heads, and executives/senior leadership, and the staff group comprised staff-level employees. Before path coefficients were compared, measurement invariance across groups was established using the measurement invariance of composite models (MICOM) procedure⁴⁵. Step 1 (configural invariance) was satisfied by identical model specification across groups, and step 2 (compositional invariance) was supported for all constructs, with original correlations at or above their 5% quantiles and non-significant permutation p-values (AHTR c = 0.999, p = 0.524; AITC c = 0.954, p = 0.205; CS c = 1.000, p = 0.800; PS c = 0.976, p = 0.073), establishing partial measurement invariance and justifying the multi-group comparison (full MICOM results in Supplementary File 3). The model was then estimated separately for each group, and path differences were tested using bootstrap-based PLS-MGA, with one-tailed and two-tailed p-values reported for the leader-staff difference (Δ) of each path; group-specific path coefficients, t-values, and p-values were reported for all direct and indirect paths, together with the magnitude of each between-group difference.

Reporting, verification, and troubleshooting

All reliability, validity, structural, mediation, and MGA outputs were compiled; the result tables were exported; and every reported statistic was verified against the software-generated output to ensure consistency across tables, figures, and text. For troubleshooting, the following guidelines apply: if many responses fail screening or attention checks, refine the eligibility requirements and clarify the survey instructions; if reliability (CR < 0.70) or convergent validity (AVE < 0.50) criteria are unmet, review outer loadings and remove weak indicators while preserving theoretical coverage; if collinearity is identified (VIF above thresholds), re-evaluate construct specifications and eliminate redundant indicators; if mediation or group effects are non-significant, re-examine sample adequacy and construct operationalization before interpretation; and use an adequate number of bootstrap resamples (e.g., 5,000) to stabilize estimates.

Results

Participant characteristics

Table 2 presents the demographic characteristics of the final sample (N = 331). The sample comprised 65.0% male and 35.0% female respondents, with a relatively balanced distribution between leaders (51.7%, n = 171) and staff (48.3%, n = 160). The largest age group was under 30 years (46.5%), followed by 31–40 years (32.0%); 76.4% held undergraduate or postgraduate degrees. Respondents were drawn from a range of industries, led by technology and telecommunications (46.2%), and from organizations of varied sizes. This profile is consistent with the study's purposive focus on employees actively engaged in AI-related work.

Characteristics	Categories	Frequency (n)	%	Cumulative %
Gender	Male (1)	215	65	65
	Female (0)	116	35	100
Age	-30	154	46.5	46.5
	31-40	106	32	78.5
	41-50	44	13.3	91.8
	51-	27	8.2	100
Role	Leader (1)	171	51.7	51.7
	Staff (0)	160	48.3	100
Education	High School	34	10.3	10.3
	College	44	13.3	23.6
	Undergraduate	142	42.9	66.5
	Postgraduate	111	33.5	100
Industry	Technology Telecommunications ICT	153	46.2	46.2
	Professional Services (Consulting, Legal, Accounting)	38	11.5	57.7
	Manufacturing Industrial	37	11.2	68.9
	Financial Services	30	9.1	77.9
	Healthcare	30	9.1	87
	Public Sector	24	7.3	94.3
	Retail	19	5.7	100
Firm Size	Fewer than 50	72	21.8	21.8
	50-99	36	10.9	32.6
	100-499	85	25.7	58.3
	500-999	46	13.9	72.2
	1000-4999	50	15.1	87.3
	5000 or more	42	12.7	100
Total		331	100	100

Table 2: Demographic characteristics of the final sample. Distribution of the 331 respondents by gender, age, organizational role, education, industry, and firm size. n = frequency; % = percentage of the total sample; Cumulative % = cumulative percentage within each characteristic. Parenthesized values in the Categories column indicate the binary coding used in analysis (e.g., Leader = 1, Staff = 0).

Measurement model evaluation

Because AITC is modeled as a second-order construct, a two-step approach was applied⁴⁰. First, the reliability and validity of the first-order constructs were evaluated using Cronbach's α, composite reliability (CR), and average variance extracted (AVE), as shown in Table 3. PS5 and AHTR4 were excluded due to low outer loadings, which improved the convergent validity of PS (AVE = 0.563) and AHTR (AVE = 0.619)³⁶. All CR values exceeded the 0.70 threshold, and all AVE values surpassed 0.50. Although Cronbach's α for AP was 0.585, its CR (0.781) and AVE (0.547) remained within acceptable limits, and α values of 0.5–0.6 can be considered acceptable for short scales⁴³. Discriminant validity was supported: all HTMT ratios (0.25–0.77) were below the conservative 0.85 threshold (Table 4)⁴⁵, and the square root of each construct's AVE exceeded its highest inter-construct correlation under the Fornell-Larcker criterion (Table 5)⁴⁴.

Construct		α	Composite reliability	AVE
AITC	AP	0.585	0.781	0.547
	MTO	0.731	0.848	0.65
	PF	0.758	0.858	0.669
CS		0.858	0.904	0.701
PS		0.741	0.837	0.563
AHTR		0.795	0.867	0.619

Table 3: Reliability and convergent validity of the first-order constructs. α = Cronbach's alpha (internal consistency; threshold 0.70, with 0.5–0.6 marginally acceptable for short scales); Composite reliability (CR; threshold 0.70); AVE = average variance extracted (convergent validity; threshold 0.50). Values reported after removal of PS5 and AHTR4. AP = autonomous planning; MTO = multi-tool orchestration; PF = proactive feedback; CS = cognitive surplus; PS = psychological safety; AHTR = AI-human task reallocation.

Constructs		Heterotrait-monotrait ration (HTMT)
		AP	MTO	PF	CS	PS	AHTR
AITC	AP
	MTO	0.75
	PF	0.64	0.59
CS		0.63	0.59	0.48
PS		0.45	0.36	0.25	0.48
AHTR		0.62	0.59	0.45	0.77	0.54

Table 4: Discriminant validity of the first-order constructs (HTMT). Heterotrait-monotrait ratio of correlations between construct pairs. Values below the conservative threshold of 0.85 support discriminant validity. AP = autonomous planning; MTO = multi-tool orchestration; PF = proactive feedback; CS = cognitive surplus; PS = psychological safety; AHTR = AI-human task reallocation.

Constructs		Fornell-Larcker criterion
		AP	MTO	PF	CS	PS	AHTR
AITC	AP	0.74
	MTO	0.49	0.81
	PF	0.42	0.44	0.82
CS		0.46	0.47	0.4	0.84
PS		0.31	0.27	0.19	0.39	0.75
AHTR		0.43	0.45	0.37	0.64	0.43	0.79

Table 5: Discriminant validity of the first-order constructs (Fornell-Larcker criterion). Diagonal elements (bold) are the square roots of each construct's AVE; off-diagonal elements are inter-construct correlations. Discriminant validity is supported when each diagonal element exceeds all correlations in its row and column.

Because all variables were self-reported by the same respondents, both procedural and statistical remedies were applied^49,50. Procedurally, respondents were guaranteed anonymity, the consent statement emphasized voluntary participation and the absence of right or wrong answers, predictor and criterion items were placed in separate survey sections, validated short scales were used to reduce fatigue, and attention checks screened out careless responding. Statistically, Harman's single-factor test⁵¹ showed that a single unrotated factor explained 30.99% of variance, below the 50% criterion^49,52. As a more stringent check, a full collinearity assessment⁴⁶ showed all VIFs below 3.3 (Table 1), suggesting no evidence of severe common method bias, although such diagnostics cannot conclusively rule out method variance^46,47.

The significance of the upward dimension effects of the second-order formative construct (AITC) was assessed using bootstrapping with 5,000 resamples^36,41. As shown in Table 6, the formative weights of AP_LVS (0.493, p < 0.001), MTO_LVS (0.492, p < 0.001), and PF_LVS (0.249, p < 0.001) were all significant, and all VIFs were below 3.3, indicating that each lower-order dimension contributes substantively to AITC without collinearity concerns. The final structural model was then estimated using the two-stage procedure^41,42 (Figure 3).

Formative indicator	VIF	Original β	Mean β	STDEV	t-value	p-value
AP_LVS → AITC	1.418	0.493	0.489	0.094	5.232	0
MTO_LVS → AITC	1.457	0.492	0.489	0.092	5.333	0
PF_LVS → AITC	1.335	0.249	0.248	0.09	2.754	0.006

Table 6: Upward dimensional effects for the second-order formative construct. Bootstrap estimates (5,000 resamples) of the formative weights linking the latent variable scores of AP, MTO, and PF (AP_LVS, MTO_LVS, PF_LVS) to AITC in the two-stage approach. β = standardized weight (Original) and bootstrap mean; STDEV = bootstrap standard deviation; VIF = variance inflation factor (<3.3 indicates no collinearity concern); t-value = \|β/STDEV\|; p-value from the bootstrap distribution. AITC = Agentic AI technological characteristics; AP = autonomous planning; MTO = multi-tool orchestration; PF = proactive feedback.

Figure 3. Estimated structural model. Final model with standardized path coefficients (β) for the full sample (N = 331). All abbreviations as in Figure 2; AP_LVS, MTO_LVS, and PF_LVS denote the latent variable scores of AP, MTO, and PF used as formative indicators of AITC in the two-stage approach. Solid arrows indicate significant paths; values in parentheses are R² (the proportion of variance explained in each endogenous construct: AHTR = 0.471, CS = 0.363, PS = 0.114). Significance was assessed by bootstrapping with 5,000 resamples; ***p < 0.001. AP = autonomous planning; MTO = multi-tool orchestration; PF = proactive feedback; AITC = Agentic AI technological characteristics; CS = cognitive surplus; PS = psychological safety; AHTR = AI-human task reallocation. Please click here to view a larger version of this figure.

Structural model and mediation analysis

Table 7 reports the path coefficients with p-values, sample means (M), standard deviations (STDEV), t-values, and 95% confidence intervals. AITC was positively associated with CS (β = 0.485, p < 0.001), supporting H1, and with PS (β = 0.338, p < 0.001), supporting H2. CS was positively associated with AHTR (β = 0.444, p < 0.001), supporting H3, as was PS (β = 0.179, p < 0.001), supporting H5. PS was positively associated with CS (β = 0.230, p < 0.001), supporting H6, and AITC with AHTR (β = 0.215, p < 0.001), supporting H9. The model explained 47.1% of the variance in AHTR (R² = 0.471), 36.3% in CS, and 11.4% in PS.

Constructs	β	Sample mean (M)	Standard deviation (STDEV)	T statistics (|O/STDEV|)	p-values	0.025	0.975	R²
AITC → AHTR	0.215	0.212	0.059	3.64	0	0.099	0.332	0.471
CS → AHTR	0.444	0.445	0.054	8.155	0	0.335	0.549
PS → AHTR	0.179	0.181	0.05	3.596	0	0.085	0.279
AITC → CS	0.485	0.485	0.051	9.587	0	0.384	0.579	0.363
PS → CS	0.23	0.231	0.053	4.299	0	0.129	0.336
AITC → PS	0.338	0.347	0.055	6.094	0	0.236	0.451	0.114
PS → CS → AHTR	0.102	0.103	0.027	3.786	0	0.054	0.159
AITC → PS → AHTR	0.06	0.062	0.019	3.225	0.001	0.027	0.102
AITC → CS → AHTR	0.215	0.216	0.036	6.007	0	0.149	0.29
AITC → PS → CS → AHTR	0.034	0.036	0.011	3.122	0.002	0.017	0.06
AITC → PS → CS	0.078	0.08	0.022	3.485	0	0.041	0.127

Table 7: Structural model results and specific indirect effects (full sample). β = standardized path coefficient; M = bootstrap sample mean; STDEV = bootstrap standard deviation; t = \|β/STDEV\|; p = two-tailed p-value; 0.025/0.975 = lower/upper bounds of the 95% bootstrap confidence interval (an interval excluding zero indicates significance); R² = variance explained in the endogenous construct. Rows with arrows through multiple constructs (e.g., AITC → PS → CS → AHTR) report specific indirect (mediation) effects. Bootstrapping used 5,000 resamples.

The bootstrap results for specific indirect effects (Table 7) show that CS significantly mediated the AITC-AHTR relationship (β = 0.215, 95% CI [0.149, 0.290], p < 0.001), supporting H4, and PS likewise (β = 0.060, 95% CI [0.027, 0.102], p = 0.001), supporting H7. The sequential pathway AITC → PS → CS → AHTR was also significant (β = 0.034, 95% CI [0.017, 0.060], p = 0.002), supporting H8. Table 8 summarizes the hypothesis-testing results.

	Hypothesis			Results
H1	Agentic AI technological characteristics are positively associated with cognitive surplus.			Accepted
H2	Agentic AI technological characteristics are positively associated with psychological safety.			Accepted
H3	Cognitive surplus is positively associated with AI-human task reallocation.			Accepted
H4	Cognitive surplus mediates the relationship between Agentic AI technological characteristics and AI-human task reallocation.			Accepted
H5	Psychological safety is positively associated with AI-human task reallocation.			Accepted
H6	Psychological safety is positively associated with cognitive surplus.			Accepted
H7	Psychological safety mediates the relationship between Agentic AI technological characteristics and AI-human task reallocation.			Accepted
H8	Psychological safety and cognitive surplus sequentially mediate the relationship between Agentic AI technological characteristics and AI-human task reallocation.			Accepted
H9	Agentic AI technological characteristics are positively associated with AI-human task reallocation.			Accepted
H10a	The associations of Agentic AI technological characteristics with cognitive surplus and psychological safety differ between leaders and staff.			Accepted
H10b	The associations of psychological safety and cognitive surplus with AI-human task reallocation differ between leaders and staff.			Partially accepted

Table 8: Summary of hypothesis tests. Each hypothesis with its corresponding structural path and decision (accepted/rejected) based on the bootstrap results in Table 7 (significance threshold p < 0.05). AITC = Agentic AI technological characteristics; CS = cognitive surplus; PS = psychological safety; AHTR = AI-human task reallocation.

Out-of-sample predictive power was assessed with PLSpredict (10 folds, 10 repetitions; Table 9). All endogenous indicators returned positive Q²predict values, indicating predictive relevance, and the PLS-SEM model produced lower RMSE and MAE than the linear-model (LM) benchmark for nearly all indicators, the only exception being PS_4. These results indicate medium-to-high out-of-sample predictive performance for psychological safety, cognitive surplus, and AI-human task reallocation⁴⁸.

Indicator	Q²predict	PLS-SEM RMSE	PLS-SEM MAE	LM RMSE	LM MAE	IA RMSE	IA MAE
PS_1	0.067	1.321	1.042	1.325	1.042	1.368	1.098
PS_2	0.018	1.596	1.314	1.602	1.333	1.610	1.363
PS_3	0.048	1.209	0.937	1.218	0.941	1.239	0.983
PS_4	0.084	1.262	0.977	1.258	0.965	1.318	1.046
CS_1	0.218	1.138	0.876	1.144	0.881	1.287	1.030
CS_2	0.276	1.045	0.792	1.048	0.793	1.229	0.971
CS_3	0.181	1.157	0.884	1.161	0.889	1.278	1.011
CS_4	0.170	1.371	1.088	1.377	1.093	1.506	1.209
AHTR_1	0.149	1.251	0.970	1.259	0.977	1.357	1.028
AHTR_2	0.188	1.151	0.895	1.158	0.900	1.277	1.003
AHTR_3	0.176	1.185	0.901	1.191	0.908	1.305	0.990
AHTR_5	0.130	1.261	0.982	1.269	0.991	1.352	1.039

Table 9: Out-of-sample predictive assessment (PLSpredict). Predictive metrics for each endogenous indicator (PS, CS, AHTR) from PLSpredict (10 folds, 10 repetitions). Q²predict = predictive relevance (>0 indicates the PLS-SEM model outperforms the most naive benchmark); PLS-SEM RMSE/MAE = root-mean-square error and mean absolute error of the PLS-SEM model; LM RMSE/MAE = corresponding errors for the linear-model benchmark; IA RMSE/MAE = indicator-average benchmark. PLS-SEM errors below the LM benchmark indicate satisfactory predictive performance. PS = psychological safety; CS = cognitive surplus; AHTR = AI-human task reallocation. Note: All Q²predict > 0. PLS-SEM RMSE/MAE < LM benchmark for all indicators except PS_4, indicating medium-to-high predictive performance.

Multi-group analysis

MGA examined whether the structural relationships differ between leaders and staff, who occupy different positions of accountability, job security, and task composition. As shown in Table 10, the associations of AITC with the two resource variables were significantly stronger among leaders: AITC → CS (leader β = 0.599 vs. staff β = 0.363; Δ = 0.236, two-tailed p = 0.014) and AITC → PS (leader β = 0.435 vs. staff β = 0.199; Δ = 0.236, two-tailed p = 0.043). These differences are practically meaningful: the AITC-CS association among leaders was roughly 1.6 times that among staff, and the AITC-PS association was more than twice as strong. The remaining direct paths (AITC → AHTR, CS → AHTR, PS → AHTR, PS → CS) did not differ significantly between groups (all two-tailed p > 0.19), indicating that once resources are formed, their associations with task reallocation operate similarly across roles. The group-specific bootstrap estimates (Table 11) further reveal asymmetric significance patterns in four pathways: PS → AHTR was significant for leaders (β = 0.234, p = 0.001) but not staff (β = 0.133, p = 0.064); the indirect path AITC → PS → AHTR was significant for leaders (β = 0.102, p = 0.005) but not staff (β = 0.026, p = 0.178); conversely, PS → CS → AHTR was significant for staff (β = 0.140, p = 0.001) but only marginal for leaders (β = 0.059, p = 0.054); and AITC → PS → CS was significant for staff (β = 0.057, p = 0.043) but marginal for leaders (β = 0.067, p = 0.065). Taken together, the psychological-safety route to task reallocation appears more direct for leaders, whereas for staff, psychological safety relates to reallocation primarily by enabling cognitive surplus, a pattern consistent with role differences in accountability and job-security exposure (H10a supported; H10b partially supported).

Constructs	Δ (Leader - Staff)	1-tailed (Leader vs Staff) p value	2-tailed (Leader vs Staff) p value
AITC → AHTR	0.007	0.482	0.963
AITC → CS	0.236	0.007	0.014
AITC → PS	0.236	0.022	0.043
CS → AHTR	-0.1	0.814	0.371
PS → AHTR	0.101	0.161	0.322
PS → CS	-0.135	0.901	0.199

Table 10: Multi-group analysis (PLS-MGA) of leader-staff path differences. Δ = difference between leader and staff standardized path coefficients (leader minus staff); p-values from bootstrap-based PLS-MGA (1-tailed and 2-tailed). Significant differences (2-tailed p < 0.05) indicate that the path differs between leaders (n = 171) and staff (n = 160). AP = autonomous planning; MTO = multi-tool orchestration; PF = proactive feedback; CS = cognitive surplus; PS = psychological safety; AHTR = AI-human task reallocation; AITC = Agentic AI technological characteristics.

	Original (Leader)	Original (Staff)	Mean (Leader)	Mean (Staff)	STDEV (Leader)	STDEV (Staff)	t value (Leader)	t value (Staff)	p value (Leader)	p value (Staff)
AITC → AHTR	0.233	0.227	0.237	0.224	0.098	0.077	2.381	2.945	0.017	0.003
AITC → CS	0.599	0.363	0.599	0.368	0.057	0.077	10.426	4.734	0	0
AITC → PS	0.435	0.199	0.448	0.226	0.075	0.089	5.82	2.24	0	0.025
CS → AHTR	0.383	0.484	0.381	0.486	0.091	0.065	4.19	7.406	0	0
PS → AHTR	0.234	0.133	0.234	0.134	0.073	0.072	3.185	1.856	0.001	0.064
PS → CS	0.154	0.289	0.157	0.289	0.07	0.077	2.212	3.741	0.027	0
PS → CS → AHTR	0.059	0.14	0.06	0.141	0.031	0.043	1.93	3.269	0.054	0.001
AITC → PS → AHTR	0.102	0.026	0.104	0.029	0.036	0.02	2.808	1.347	0.005	0.178
AITC → CS → AHTR	0.229	0.176	0.228	0.18	0.06	0.047	3.834	3.757	0	0
AITC → PS → CS → AHTR	0.026	0.028	0.027	0.031	0.016	0.015	1.65	1.852	0.099	0.064
AITC → PS → CS	0.067	0.057	0.071	0.064	0.036	0.028	1.844	2.02	0.065	0.043

Table 11: Group-specific bootstrap estimates for leaders and staff. Standardized path coefficients (Original), bootstrap means, standard deviations (STDEV), t-values, and two-tailed p-values were estimated separately within the leader (n = 171) and staff (n = 160) groups, for all direct and specific indirect paths. Bootstrapping used 5,000 resamples within each group. AP = autonomous planning; MTO = multi-tool orchestration; PF = proactive feedback; CS = cognitive surplus; PS = psychological safety; AHTR = AI-human task reallocation; AITC = Agentic AI technological characteristics.

In sum, the measurement model met the reliability and validity criteria; AITC was positively associated with CS, PS, and AHTR; CS and PS mediated the AITC-AHTR association individually and sequentially; the model showed satisfactory out-of-sample predictive performance; and the associations of AITC with CS and PS were significantly stronger among leaders, whereas the paths from resources to task reallocation did not differ significantly between roles. Taken together, the results support the hypothesized resource-gain model and motivate the role-differentiated interpretation developed in the Discussion.

DATA AVAILABILITY:

The data and analysis files are uploaded as Supplementary File 1, Supplementary File 2, Supplementary File 3, Supplementary File 4, Supplementary File 5, Supplementary File 6, and Supplementary File 7.

Supplementary File 1: Questionnaire.Please click here to download this file.

Supplementary File 2. G*Power protocol output.Please click here to download this file.

Supplementary File 3. MICOM Step 2 — Compositional invariance (permutation, 1,000). Compositional invariance established for all constructs (original c ≥ 5% quantile; permutation p > 0.05). Partial measurement invariance is sufficient to justify PLS-MGA.Please click here to download this file.

Supplementary File 4. Dataset.Please click here to download this file.

Supplementary File 5. MGA bootstrap SmartPLS report.Please click here to download this file.

Supplementary File 6. MGA permutation SmartPLS report. Please click here to download this file.

Supplementary File 7. Codebook — variable, item, construct, scale, coding, and source.Please click here to download this file.

Discussion

This study examined how the technological characteristics of Agentic AI relate to employees' psychological resources and task-reallocation behavior, and whether these relationships differ across organizational roles. Using a cross-sectional survey of 331 AI-experienced employees and PLS-SEM with mediation and multi-group analysis, the results were consistent with the theorized resource-gain model: AITC was positively associated with cognitive surplus, psychological safety, and AI-human task reallocation; cognitive surplus and psychological safety mediated the AITC-AHTR relationship individually and sequentially; and the associations of AITC with both resource variables were significantly stronger among leaders than staff.

The findings carry three theoretical implications. First, they extend JD-R theory into the Agentic AI context by treating technological characteristics not as tools but as job resources: the strong AITC-CS association (β = 0.485) is consistent with the view that autonomous planning, multi-tool orchestration, and proactive feedback conserve cognitive energy and discretionary time that employees can reinvest^15,23,24. Second, they connect COR theory's gain-spiral logic to AI-enabled work redesign: the significant sequential pathway (AITC → PS → CS → AHTR) suggests that psychological safety operates as a contextual resource that supports the accumulation of cognitive surplus, which in turn relates to reallocation behavior^20,22. Third, by positioning AHTR as a technology-specific form of task crafting within the job crafting literature^17,31,32, the study moves beyond the dominant focus on technical performance and automation outcomes and offers a psychological account of how agentic systems become embedded in work redesign. Because the design is cross-sectional, these pathways should be read as associations consistent with the theorized causal ordering rather than demonstrations of causality.

The mediation results clarify why advanced AI availability alone may not translate into changed work practices. The direct AITC-AHTR association (β = 0.215) was complemented by indirect pathways of comparable combined magnitude through cognitive surplus (β = 0.215) and psychological safety (β = 0.060): employees appear to reallocate tasks to AI when the technology demonstrably frees cognitive resources and when the interpersonal climate makes experimentation feel safe^16,25,38. This dual mechanism implies that task reallocation is as much a resource and climate phenomenon as a technology phenomenon, which may explain heterogeneous adoption outcomes across otherwise similar deployments.

The multi-group results add a role-based boundary condition. The AITC-CS and AITC-PS associations were roughly 1.6 and 2.2 times stronger, respectively, for leaders than for staff, whereas the downstream paths from resources to reallocation did not differ significantly between groups. One interpretation, consistent with perceptual-distance arguments^30,35, is that leaders, who hold greater decision authority and strategic accountability, more readily convert agentic capabilities into perceived resources, while staff, who are more exposed to automation-related job insecurity and learning demands, realize psychological safety benefits less directly: for staff, psychological safety related to reallocation primarily through cognitive surplus (PS → CS → AHTR: β = 0.140, p = 0.001), whereas for leaders the direct PS → AHTR path was significant (β = 0.234, p = 0.001). These are tendencies observed in this sample rather than fixed dispositions, and the confidence intervals around group differences warrant cautious interpretation.

Practically, the findings suggest role-differentiated deployment strategies. For all employees, organizations can protect the conversion of saved time into cognitive surplus by adjusting workload norms and key performance indicators so that AI-freed hours are reinvested in strategic, creative, and collaborative tasks rather than absorbed by additional routine demands; concrete levers include AI-use governance that clarifies accountability when AI errs, structured training in delegation-oriented AI skills, and recognition systems that reward documented task redesign. For staff specifically, the results imply that psychological safety interventions, such as no-blame review of AI experiments, manager tolerance of early errors, and explicit job-security communication, are prerequisites for cognitive surplus to form at all; for leaders, who convert AITC into resources more readily, the priority is channeling that surplus into governance and contingency planning that narrows the leader-staff perception gap rather than widening it. These implications extend to high-stakes settings such as fraud detection and financial-security operations, in which agentic systems increasingly triage alerts, orchestrate investigation workflows, and flag anomalous transactions: in such environments, psychological safety is a precondition for employees to question, override, and report AI-generated judgments without fear of blame, and reallocating routine monitoring to AI can free analysts’ cognitive resources for complex, judgment-intensive cases, while governance must ensure that automation does not erode accountability for high-consequence decisions. More broadly, the observed leader-staff differences argue for differentiated interventions: training curricula that emphasize delegation and oversight skills for leaders and experimentation and reskilling support for staff, governance roles that reflect each group’s risk exposure, and communication strategies that make the resource gains realized by leaders visible and attainable for staff.

Several limitations qualify these conclusions and define future directions. All variables were self-reported by the same respondents at one time point; despite procedural remedies and statistical checks suggesting that common method bias was not severe, the cross-sectional design precludes causal inference and cannot capture time-lagged resource spirals. In addition, AITC was measured through employees’ self-reported perceptions of agentic system capabilities, which may not correspond to the systems’ actual technical capabilities; future designs should pair perceptual measures with system logs or capability audits. Although participants were shown a brief standardized definition of Agentic AI at the start of the survey, no comprehension-check item verified their understanding of the focal technology, and future studies should include such a check. The purposive, platform-recruited sample of AI-experienced employees, weighted toward technology industries and younger respondents, is appropriate for an early model test but limits generalizability to organizations with low AI maturity, smaller firms, and non-digital sectors; the single-panel recruitment in one collection period likewise implies cultural and contextual specificity, so the observed associations may not generalize across labor markets, AI-governance regimes, or occupational cultures. Psychological safety was also measured as an individual perception rather than a shared team climate. Future research should employ longitudinal or experience-sampling designs to trace gain spirals, multilevel designs that separate individual and team-level psychological safety, probability or organization-based samples across AI-maturity levels, and extensions of the AITC construct to attributes such as transparency, controllability, and explainability. Notwithstanding these limitations, the study provides a replicable protocol and initial evidence that the associations of Agentic AI with work outcomes operate through, and are bounded by, employees’ psychological resources, knowledge that should help researchers, practitioners, and policymakers deploy agentic systems in ways that augment rather than erode the human side of work.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Prolific	Prolific	https://www.prolific.com/	Online survey pannel provider
Qualtrics	Qualtrics, LLC	https://www.qualtrics.com/	Survey coding
SmartPLS	SmartPLS GmbH	version 4.0.1, https://www.smartpls.com/	PLS-SEM analysis