An Exploratory Pilot Study of Cognitive Factors Influencing the Adoption of AI Coding Agents in Agile Development Teams

Abstract

This study considers the cognitive factors influencing the adoption of artificial intelligence (AI) coding agents among software engineers in Agile development environments. In this context, Transparency refers to the degree to which an AI system’s decision processes are clear or understandable to its users, while Trust denotes the confidence developers place in the reliability and correctness of AI-generated code. Specifically, it analyzes the roles of Transparency and Trust as predictors of developers’ Behavioral Intention to deploy AI-generated code. Using a quantitative, cross-sectional survey design administered to 19 Agile software professionals, this study uses descriptive statistics, Pearson correlation analysis, and simple and multiple linear regression to test a mediation model based on the Technology Acceptance Model (TAM). In this exploratory pilot sample, Trust emerged as the strongest predictor of Behavioral Intention (β = 0.64, p = 0.007), though findings should be interpreted cautiously given the limited sample size. Trust demonstrated the strongest association with Behavioral Intention, while Transparency showed a non-significant positive association with Trust. The observed pattern is consistent with a possible indirect relationship, although mediation was not confirmed in this sample. This evidence suggests that organizations seeking to improve AI adoption in development teams should prioritize the reliability and accuracy of AI tools, conditions that encourage trust, over transparency features alone. This study adds to the cognitive engineering literature on human–AI interaction in high-cognitive-demand work environments.

1. Introduction

The recent emergence of AI coding agents such as GitHub Copilot, Cursor, and Devin is transforming professional software development workflows. By autonomously generating and reviewing code, these tools are shifting developers’ roles from sole authors to overseers of AI outputs [1]. Despite increasing adoption, usage remains inconsistent [2]. This inconsistency highlights the need to understand the variables that drive adoption, especially as AI integration accelerates in the coming decade [3], and sets the stage for examining the cognitive aspects influencing developers’ engagement with these technologies.

Building on this context, it is important to note that while prior studies of technology adoption focus on usability and productivity, they often overlook the cognitive and psychological aspects that play a key role in Agile environments [4]. Developers managing AI-generated code must assess correctness and fit under substantial cognitive load, relying on calibrated trust in AI outputs [5], which points to a deeper cognitive dimension in adoption.

Thus, adoption in software engineering depends not only on the proven benefits of AI tools but also on whether developers form the mental models and behavioral intentions needed to incorporate them. Cognitive engineering highlights that this adoption is not automatic, even when performance gains are clear [6], suggesting that other influencing factors must be considered.

Motivated by these considerations, this study addresses a gap in cognitive engineering research by empirically examining, for the first time within the context of a real-world enterprise pilot, how two key cognitive constructs, Transparency and Trust, predict software developers’ Behavioral Intention to deploy AI-generated code within Agile teams. In contrast to previous research, this study uses data from a structured Qualtrics survey administered to developers involved in deploying Devin AI, an AI coding agent, within a live enterprise software organization, thereby grounding its analysis in practical experience.

The following research questions guide this study:

RQ1: What is the relationship between AI tool transparency and developer trust in AI-generated code?

RQ2: Does Trust significantly predict software developers’ Behavioral Intention to deploy AI-generated code?

RQ3: Does Trust mediate the relationship between Transparency and Behavioral Intention to deploy AI-generated code?

Because of the emerging nature of enterprise AI coding agent deployment, this study is positioned as an exploratory pilot investigation intended to identify preliminary patterns and generate hypotheses for future research rather than provide definitive causal evidence.

2. Literature Review

2.1. Technology Acceptance in Software Engineering

The Technology Acceptance Model (TAM), introduced by Davis [7], is the dominant framework for studying adoption behavior in information systems research. In its original form, TAM proposes five core constructs: Perceived Usefulness (PU), the degree to which a user believes a technology will enhance their performance; Perceived Ease of Use (PEOU), the degree to which use of the technology is expected to require minimal effort; Attitude Toward Use, reflecting the user’s overall positive or negative evaluation of the technology; Behavioral Intention to Use (BI), the user’s stated plan to use the technology; and Actual Use, the behavioral outcome. Within the model, PEOU influences both PU and Attitude; PU and Attitude jointly shape Behavioral Intention; and Behavioral Intention drives Actual Use. External variables, such as individual characteristics, system features, and task context, serve as antecedents that shape PU and PEOU, making TAM sufficiently flexible to be extended to domain-specific applications. Since Davis’s original formulation, researchers have extended TAM by adding external variables to improve its explanatory power in specific contexts. Extensions of TAM, particularly the Unified Theory of Acceptance and Use of Technology (UTAUT), introduce additional constructs, including Social Influence and Facilitating Conditions [8]. In the context of AI-assisted software development, researchers have applied extended TAM frameworks to understand adoption patterns among developers [9], finding that domain-specific cognitive constructs, such as trust in AI accuracy and cognitive demand, are necessary additions to traditional TAM variables [10]. The present study extends TAM by treating Transparency and Trust as external cognitive variables that influence Behavioral Intention to deploy AI-generated code. Transparency is theorized to function as an antecedent that shapes Trust, which in turn predicts Behavioral Intention, following the external-variables logic of extended TAM frameworks. Ali et al. [1] validated a similar approach in a student programming context using structural equation modeling, confirming that an extended TAM incorporating trust and learning motivation constructs accounted for 47% of variance in actual AI tool use.

The present study does not attempt to test the full TAM nomological network. Rather, it investigates Transparency and Trust as external cognitive antecedents operating within a TAM-informed perspective. This approach is appropriate because recent AI adoption literature suggests that trust-related constructs may be especially salient in high-consequence human–AI decision contexts where users must evaluate autonomous outputs before deployment. The objective therefore is not to replace traditional TAM constructs such as Perceived Usefulness or Perceived Ease of Use, but to explore whether trust-oriented cognitive variables provide additional explanatory value in AI-assisted software development settings.

2.2. Cognitive Factors: Trust, Transparency, and Cognitive Load

Trust is recognized as a primary determinant of human behavior toward automated and AI systems [11]. In the context of AI code generation, Choudhuri et al. [12] found that developers’ trust is shaped by the perceived reliability and contextual precision of AI outputs, and that trust formation precedes stable adoption behavior. Transparency, defined as the degree to which an AI system makes its decision processes legible to users, has been theorized to facilitate trust [13]. However, empirical evidence for a direct link between transparency and adoption behavior is mixed, as some studies show that transparency may increase mental effort without proportionally increasing trust [5].

Khan and Suhluli [5] examined generative AI use in research contexts, finding that cognitive load, fatigue, and trust interrelate in complex ways, affecting sustained AI use. In the context of programming education, Ali et al. [1] similarly found that trust positively affected perceived ease of use of AI coding assistants among undergraduate computer science students, though it did not directly predict perceived usefulness, suggesting that the role of trust varies meaningfully across user populations and task contexts. High transparency may paradoxically increase cognitive demand by requiring developers to evaluate more information about AI behavior. This finding motivates the mediation model tested in the present study: Transparency may influence adoption indirectly, with Trust serving as the cognitive bridge.

2.3. AI Adoption in Agile Development Contexts

Agile methodologies, with their time-sensitive and team-based focus, shape how AI is adopted. Li et al. [3] project that generative AI will fundamentally reshape software development workflows within the next decade, with autonomous code generation and review becoming standard practice, a trajectory that amplifies the urgency of understanding current adoption barriers at the cognitive level. Sprint-driven delivery cycles create pressure to rapidly audit AI-generated code, possibly increasing or reducing cognitive load depending on the tool’s reliability [14]. Radulovic and Klobucar [15] identified adoption friction points specific to Agile, including role ambiguity when AI agents take on tasks traditionally owned by developers, and sprint velocity disruptions during AI integration periods.

Ozkaya [16] framed AI-augmented software development as a frontier shift in which developers must learn to collaborate with AI systems rather than simply use them as tools, introducing new cognitive demands around oversight, verification, and trust. Russo [17] examined the complexity of generative AI adoption across software engineering roles, finding that engineers with higher autonomy and task complexity showed different adoption curves than those in more routine development roles. Similarly, Ram et al. [14] found that adoption patterns differentiated by role within Agile teams, with architects and senior engineers adopting earlier than junior developers, a pattern explained by differences in self-efficacy and task mental model clarity.

2.4. Human–AI Collaboration in Software Development

Treude and Gerosa [6] proposed a taxonomy of human–AI interaction in software engineering, distinguishing between delegation, co-creation, and supervision modes. Each mode imposes different cognitive demands: delegation requires trust sufficient to reduce verification behavior; co-creation requires alignment of mental models; supervision requires calibrated skepticism and the cognitive capacity to detect AI errors reliably.

Sauvola et al. [2] conducted a grounded theory study of AI tool adoption in software development, identifying trust calibration, the ongoing process by which developers adjust their reliance on AI outputs based on experience, as a key mechanism distinguishing sustained adopters from those who abandoned AI tools after initial use. Cynthia et al. [18] further found differential usage patterns between core and peripheral developers within the same team, suggesting that organizational role and cognitive familiarity with AI outputs are as important as the tool’s characteristics in predicting adoption depth.

3. Methodology

3.1. Research Design, Variables, and Metrics

This study uses a quantitative, cross-sectional exploratory survey study to examine the relationships between Transparency, Trust, and Behavioral Intention in AI coding agent adoption. The study tests a mediation model in which Transparency (X) predicts Trust (M), which predicts Behavioral Intention to Deploy (Y). A quantitative approach is appropriate given the study’s objective: to test theoretically derived relationships among latent cognitive constructs using statistical inference. The cross-sectional design permits efficient measurement of cognitive variables across a diverse sample of software developers at a single point in time, allowing for comparisons at the individual level.

The variables Transparency (independent), Trust (mediator), and Behavioral Intention (dependent) were each measured with Likert items (1 = Strongly Disagree, 5 = Strongly Agree). A construct score is the mean of its items (3, 5, and 3 items, respectively), giving one continuous score from 1 to 5 per respondent. Participant descriptors (role, years of experience, Agile practice) were collected to describe the sample and were not used as covariates.

Item scores were averaged into construct scores and summarized with the mean, median, mode, and standard deviation. Bivariate associations were quantified with Pearson’s r and interpreted as weak (0.1–0.3), moderate (0.3–0.5), or strong (>0.5). Four ordinary least squares regression models were then estimated. For each model, coefficients with standard errors, the standardized β, R², and 95% confidence intervals were reported. Statistical significance was set at α = 0.05 (two-tailed). Given n = 19, regression assumptions were checked with residual, line plot fit, and normality plots.

3.2. Target Population, Sample, and Data Collection

The target population consists of Agile software developers with experience using or interacting with AI coding agents, including but not limited to GitHub Copilot, ChatGPT, Cursor, and Devin AI. A purposive sample of 19 respondents was recruited from professional Agile development teams within enterprise software organizations, including participants from an active Devin AI pilot program. While the sample is small relative to traditional quantitative standards, it is appropriate for an exploratory study in a nascent research area and consistent with prior pilot studies examining AI adoption in software contexts [10]. Given the exploratory nature of this investigation and limited accessibility to enterprise AI pilot participants, this study should be considered a pilot study intended to establish preliminary evidence regarding cognitive factors influencing AI coding agent adoption.

Data were collected via a structured online questionnaire administered through Qualtrics. The questionnaire comprised closed ended Likert scale items (1 = Strongly Disagree, 5 = Strongly Agree) measuring three primary constructs. Participation was voluntary, and respondent anonymity and confidentiality were assured. The complete de-identified item-level dataset is available in the Supplementary Materials to support reproducibility.

3.3. Operationalization of Constructs

Three constructs were measured in this study. Transparency (independent variable, X) was operationalized as respondents’ perceived ability to understand how the AI coding agent generates and justifies its outputs, including the explainability of code suggestions and the clarity of the agent’s decision logic. Trust (mediating variable, M) was operationalized as developers’ perceived confidence in the AI agent’s accuracy, reliability, and contextual appropriateness across development tasks. Behavioral Intention to Deploy (dependent variable, Y) was operationalized as respondents’ self-reported likelihood of integrating AI-generated code into production systems, capturing the behavioral endpoint of cognitive adoption processes consistent with TAM frameworks [7].

Transparency was measured with three items, Trust with five items, and Behavioral Intention with three items (11 items total). All items used a 5-point Likert scale (1 = Strongly Disagree, 5 = Strongly Agree), and each construct score was the mean of its items.

Item content was informed by established technology acceptance and trust-in-automation frameworks rather than adopted verbatim from a single validated instrument. Trust items drew on Davis’s TAM perceived usefulness and ease-of-use constructs [7] and on trust-in-automation research related to explainable AI, including the trust framework used by Choudhuri et al. [12]. Transparency items were author-developed, conceptually grounded in the explainability/trust signal distinction described by Çetinkaya and Krämer [13]. Behavioral Intention items were author-developed, informed by technology adoption research in software-engineering contexts [12,19]. All items were written or adapted specifically to reference AI coding agents.

Representative items include “I trust the accuracy of the AI coding assistant’s code suggestion” and “I trust an AI coding assistant will not introduce bugs or vulnerabilities” (Trust); “AI coding assistant is transparent about its limitations” and “I have a clear understanding of what an AI coding assistant can and cannot do” (Transparency); and “I have effectively integrated an AI coding agent into my development workflow” and “I feel comfortable deploying code suggested by an AI coding assistant” (Behavioral Intention). The complete survey instrument and item wording are provided in Appendix A (Table A1). Q24 was reverse-scored prior to scale aggregation. Q10 was included because effective workflow integration reflects a user’s readiness to incorporate AI-generated outputs into routine development activities. Q23 was included because perceived fit between AI capabilities and team tasks represents an important antecedent of future deployment intentions.

The internal consistency was low for two of the three constructs (Transparency α = 0.14; Behavioral Intention α = 0.31), likely reflecting the small sample size (n = 19) and, for Transparency, the absence of a validated source scale; Trust showed acceptable reliability (α = 0.73). This limits confidence in the Transparency and Intention composites and is addressed further in the Limitations. Because the Transparency scale was author-developed rather than adapted from a validated instrument, and its internal consistency in this sample was low (α = 0.14), we treat the Transparency construct as exploratory; validating a dedicated Transparency scale is a priority for future work.

3.4. Data Analysis Approach

Data analysis was conducted in Microsoft Excel. Descriptive statistics, including means, medians, modes, and standard deviations, were calculated for each construct to characterize the sample’s cognitive profile. Pearson’s correlation coefficient (r) was computed to assess the direction and strength of bivariate relationships among constructs prior to regression analysis. Correlation strength was interpreted as weak (0.1–0.3), moderate (0.3–0.5), or strong (>0.5).

Regression analysis was conducted to test the relationships between variables and to evaluate their statistical significance. Three study variables were defined: the independent variable (X) is Transparency, the mediating variable (M) is Trust, and the dependent variable (Y) is Intention to Deploy. The expressed mediation relationship is Transparency (X) → Trust (M) → Intention to Deploy (Y), meaning Transparency influences Trust, Trust influences Intention to Deploy, and Trust mediates the effect of Transparency on Intention to Deploy.

Four regression models were formulated to test these relationships and answer the research questions.

• Model 1: Transparency → Trust

Dependent variable (Y): Trust. Independent variable (X): Transparency.

Trust = β₀ + β₁(Transparency) + ε

Test: Does Transparency influence Trust?

• Model 2: Trust → Intention to Deploy

Dependent variable (Y): Intention to Deploy. Independent variable (X): Trust.

Intention to Deploy = β₀ + β₁(Trust) + ε

Test: Does Trust influence Intention to Deploy?

• Model 3: Transparency → Intention to Deploy

Dependent variable (Y): Intention to Deploy. Independent variable (X): Transparency.

Intention to Deploy = β₀ + β₁(Transparency) + ε

Test: Does Transparency directly influence Intention to Deploy?

• Model 4: Mediation Model

Dependent variable (Y): Intention to Deploy. Independent variable (X): Transparency. Mediating variable (M): Trust.

Intention to Deploy = β₀ + β₁(Transparency) + β₂(Trust) + ε

Test: Does Trust mediate the relationship between Transparency and Intention to Deploy?

Regression analysis results provided the path coefficients (β), intercept (β₀), F-statistics, and p-values. Statistical significance was evaluated at α = 0.05. Each model also reports the standard error of estimate (ε), reflecting the degree of unexplained variation in the dependent variable.

Regression assumptions were checked graphically for all four models. Linearity and homoscedasticity were assessed with residual plots (residuals against the predictor): the absence of curvature supports linearity, and an even spread of points across the predictor range supports homoscedasticity. Normality of residuals was assessed with histograms, where a single, roughly symmetric peak centered near zero supports the assumption. Given the small sample (n = 19), these graphical checks are indicative, not conclusive.

Regression modeling follows standard ordinary least squares procedures described by Montgomery et al. [20], while Pearson correlation follows procedures described by Cohen et al. [21].

4. Results

4.1. Descriptive Statistics

The final sample consisted of 19 Agile software developers, including software engineers, senior engineers, and program managers. All participants reported prior experience with at least one AI coding agent tool. Descriptive statistics for the three study elements are presented in Table 1.

Table 1. Descriptive Statistics for Study Constructs.

Construct	Mean	SD	Median	Min	Max
Transparency (X)	3.09	0.65	3.00	2.33	4.67
Trust (M)	3.23	0.58	3.40	2.20	4.60
Intention to Deploy (Y)	3.47	0.71	3.33	2.33	5.00

All three constructs scored above the Likert midpoint (3.0), indicating moderately favorable perceptions of AI coding agents across the sample. Behavioral Intention to Deploy exhibited the highest mean (M = 3.47), with a right-skewed distribution, indicating that a subset of highly enthusiastic adopters elevated the overall average. Trust demonstrated the most consistent perceptions, reflected in the lowest standard deviation (SD = 0.58). Transparency showed the highest scores across all three measures of central tendency (M = 3.09, Mdn = 3.00), suggesting that developers view AI transparency as present but not particularly influential in their everyday experience.

This pattern of moderate but differentiated perceptions across constructs is consistent with early to mid-adoption stages in extended TAM frameworks involving AI systems [6], where behavioral intention can be relatively strong even when foundational belief constructs such as transparency remain only moderately favorable.

4.2. Inferential Analyses

Pearson’s correlation coefficients were computed among all three study constructs. The results are presented in Table 2.

Table 2. Pearson Correlation Coefficients Among Study Constructs.

Variable Pair	Correlation
Transparency ↔ Trust	r(17) = 0.42, p = 0.070
Trust ↔ Intention to Deploy	r(17) = 0.60, p = 0.007
Transparency ↔ Intention to Deploy	r(17) = 0.30, p = 0.208

All three variable pairs demonstrated positive correlations, though only one correlation was statistically significant. Trust and Behavioral Intention exhibited the strongest correlation (r = 0.60), indicating that developers with higher trust in AI accuracy are substantially more likely to intend to deploy AI-generated code. Transparency and Trust showed a moderate positive correlation (r = 0.42), suggesting that greater perceived AI transparency is associated with higher trust, though the relationship is incomplete; transparency contributes to trust but is not sufficient on its own. The weakest correlation was observed between Transparency and Behavioral Intention (r = 0.30), indicating that transparency alone has limited direct influence on deployment decisions. These findings suggest a pattern in which Trust is more strongly associated with Behavioral Intention than Transparency; however, mediation cannot be inferred from the current data.

Four regression models were estimated to formally test the hypothesized relationships and evaluate potential mediation. The results are summarized in Table 3.

Table 3. Regression Model Output.

Predictor	β	SE	95% CI	t	p
Model 1: Transparency → Trust
Intercept (β0)	1.99	0.65	(0.62, 3.37)	3.06	0.007
Transparency (β1)	0.40	0.21	(−0.04, 0.84)	1.94	0.070
Model 2: Trust → Intention to Deploy
Intercept (β0)	1.40	0.68	(−0.03, 2.83)	2.06	0.055
Trust (β1)	0.64	0.21	(0.21, 1.08)	3.10	0.007
Model 3: Transparency → Intention to Deploy (Direct Effect)
Intercept (β0)	2.53	0.73	(0.98, 4.07)	3.45	0.003
Transparency (β1)	0.31	0.23	(−0.18, 0.80)	1.32	0.206
Model 4: Mediation Model (Transparency + Trust → Intention to Deploy)
Intercept (β0)	1.30	0.79	(−0.37, 2.97)	1.65	0.119
Trust (β1)	0.61	0.24	(0.12, 1.11)	2.61	0.019
Transparency (β2)	0.06	0.22	(−0.41, 0.53)	0.27	0.791

Model 1 examined the direct effect of Transparency on Trust. The positive coefficient (β₁ = 0.40) indicates that higher transparency is associated with increased developer trust. However, this relationship did not reach statistical significance at the 0.05 level (F(1, 17) = 3.74, p = 0.070). The standard error (ε = 0.21) of the predictor reflects moderate unexplained variation in Trust, consistent with the expectation that additional cognitive factors, such as prior AI experience and task self-efficacy, likewise influence trust formation.

Model 2 tested the effect of Trust on Behavioral Intention to deploy AI-generated code. Trust was a statistically significant predictor (β₁ = 0.64, F(1,17) = 9.61, p = 0.007), indicating that each unit increase in Trust is associated with a 0.64 unit increase in Behavioral Intention. This finding provides preliminary evidence consistent with Trust as a reliable and meaningful cognitive determinant of AI adoption behavior (RQ2). The moderate standard error (ε = 0.21) suggests that other variables not captured in this model, such as peer influence or organizational readiness, likewise contribute to deployment intention.

Model 3 estimated the total direct effect of Transparency on Behavioral Intention, controlling for Trust. The coefficient (β₁ = 0.31) was positive but not statistically significant (F(1, 17) = 1.73, p = 0.206), confirming that Transparency alone does not meaningfully explain variation in developers’ deployment decisions (RQ1, RQ3 precursor). The higher standard error (ε = 0.23) relative to Model 2 further supports the conclusion that Trust captures more explanatory variance in Behavioral Intention than Transparency does directly.

Model 4 examined the joint effects of Transparency and Trust on Behavioral Intention, enabling assessment of mediation. The overall model was statistically significant (F(2,16) = 4.58, p = 0.027). Within the model, Trust remained a significant predictor (β₁ = 0.61, p = 0.019), while Transparency’s coefficient was substantially attenuated and non-significant (β₂ = 0.06, p = 0.79). Although Trust remained significant when both predictors were included, the Transparency-to-Trust relationship did not reach statistical significance. Therefore, mediation was not established, and the observed coefficient pattern should be interpreted only as exploratory. Although the pattern of coefficients was consistent with an indirect-effect interpretation, the Transparency-to-Trust path failed to reach conventional significance levels. Consequently, mediation cannot be definitively established within this sample.

Table 4 shows that Trust exhibited a large association with Behavioral Intention (Model 2: β_std = 0.60, 95% CI [0.19, 1.01], R² = 0.36), which remained large after controlling for Transparency (Model 4: β_std = 0.0.61, 95% CI [0.2, 1.02]). Transparency showed a medium-to-large association with Trust (Model 1: β_std = 0.42, 95% CI [−0.04, 0.88], R² = 0.18) and a medium total association with Intention (Model 3: β_std = 0.30, 95% CI [− 0.19, 0.79], R² = 0.09); both intervals included zero. When Trust was included, the association between Transparency and Intention was negligible (Model 4: β_std = 0.06, 95% CI [−0.45, 0.57]). Trust is the only reliably supported predictor of Behavioral Intention, explaining about 36% of its variance. Transparency’s associations are moderate but imprecise, and they become negligible once Trust is included, suggesting that any influence of Transparency on Intention operates through Trust. Because the confidence intervals are wide and several include zero, these findings are preliminary and require replication in larger samples.

Table 4. Effect sizes for regression models.

Models	Predictor	β	95% CI (β)	Effect Size (β_std)	95% CI (β_std)	R2
Transparency → Trust	Transparency	0.40	(−0.04, 0.84)	0.42	(−0.04, 0.88)	0.18
Trust → Intention	Trust	0.64	(0.21, 1.08)	0.60	(0.19, 1.01)	0.36
Transparency → Intention	Transparency	0.31	(−0.18, 0.80)	0.30	(−0.19, 0.79)	0.09
Transparency + Trust → Intention	Transparency	0.06	(−0.41, 0.53)	0.06	(−0.45, 0.57)	0.36
	Trust	0.61	(0.12, 1.11)	0.58	(0.2, 1.02)

For model 1, the effect size is a medium-to-large association. Transparency explains about 18% of the variance in Trust, but the CI includes zero, so the effect is imprecise and not statistically confirmed. Model 2: Large, clearly positive effect. The CI excludes zero, and Trust explains about 36% of the variance in Intention. This is the strongest finding. Model 3: Medium in size, but the CI includes zero, so Transparency alone does not reliably predict Intention. It explains about 9% of the variance. Model 4: With Transparency controlled, Trust keeps a large effect that excludes zero. Transparency’s effect is negligible, and its CI is centered near zero. Together, they explain about 36% of Intention.

5. Discussion

The primary finding of this study is that Trust is a statistically significant predictor of software developers’ Behavioral Intention to deploy AI-generated code, while Transparency does not exert a significant direct effect. This pattern is theoretically important: it suggests that developers do not adopt AI coding agents primarily because they understand how those agents work, but because they have developed confidence in what those agents produce. This distinction, between cognitive comprehension and affective confidence, has practical implications for AI adoption strategy in Agile settings.

The moderate, positive, but non-significant relationship between Transparency and Trust (r = 0.42, p = 0.070) indicates that transparency contributes to trust formation but is insufficient on its own. Understanding an AI agent’s logic may be a necessary but not sufficient condition for trusting its outputs. Developers may require demonstrated performance, correct code, reliable suggestions, minimal hallucination, before transparency translates into actionable trust [13].

The results are broadly consistent with the technology acceptance literature. Choudhuri et al. [12] similarly found that trust in AI-generated code is more strongly affected by perceived system reliability than by explainability features. Wang et al. [22] found that developers’ trust in AI code generation tools was largely determined by the accuracy and contextual fit of suggestions, rather than transparency mechanisms such as explanation panels or confidence scores.

The observed pattern may indicate a possible indirect relationship through Trust; however, mediation remains unconfirmed and requires validation using larger samples and more rigorous indirect-effect testing procedures. The potential mediation pattern is consistent with Wanner et al. [19], who found that system transparency influences user acceptance through trust as an intervening variable. However, Çetinkaya and Krämer [13] found that the relationship between transparency and trust is moderated by user expertise, suggesting that more experienced developers may require less transparency to form stable trust, a boundary condition this study could not assess at n = 19. While the coefficient pattern aligns directionally with prior studies proposing trust as an intervening construct, the present study did not provide sufficient evidence to establish mediation. Larger studies employing formal indirect-effect testing are needed. However, as discussed in Section 4.2, our data cannot confirm the required first-stage path (Transparency-to-Trust, p = 0.070) at conventional significance with n = 19, so our study is best read as consistent with, rather than an independent confirmation of, Wanner et al.’s finding. Our sample, drawn from working Agile software professionals rather than novice users, may sit at a point along this moderation continuum that our design was underpowered to isolate at n = 19; this remains a plausible, testable explanation for why our own Transparency-to-Trust path fell short of conventional significance.

Russo [17] offers a particularly close point of comparison, having applied a mixed-methods design combining TAM, Diffusion of Innovation theory, and Social Cognitive Theory to an initial structured sample of 100 software engineers. Two of Russo’s findings map closely onto ours. First, Russo’s Hypothesis 1, that developers’ general perceptions about generative AI technology directly predict their intention to use it, was not supported; instead, Compatibility Factors, the fit of the tool within developers’ existing workflows, emerged as the dominant driver of adoption at this early stage of AI integration, outperforming conventional technology acceptance predictors. Our own finding that Transparency, arguably a specific facet of developers’ “perceptions about the technology,” likewise fails to directly predict Behavioral Intention is consistent with this pattern: broad cognitive appraisals of an AI tool do not appear to translate directly into adoption decisions in either study. Second, the overall adoption profile we observe, all three constructs above the scale midpoint but below 4.0 on a 5-point scale, matches Russo’s broader characterization of the field as currently in a generative but cautious growth phase, in which developers are engaging with AI tools without yet exhibiting confident, unreserved reliance on their outputs. Where our study diverges from Russo’s is in scope: we did not measure Compatibility Factors, and Russo’s finding that workflow fit dominates individual-level adoption decisions suggests a concrete construct for extending our own model in future replications. The overall adoption scores, all constructs above the scale midpoint but below 4.0, are consistent with Russo’s [17] characterization of AI adoption in software engineering as currently in a generative but cautious growth phase, where developers are engaging with but have not yet reached a high level of confident reliance on AI outputs.

Li et al. [3] provide a different, complementary point of reference. Unlike the empirical trust studies discussed above, theirs is a conceptual analysis proposing four scenario trajectories for how generative AI may reshape software development over the coming decade, rather than a quantitative test of adoption predictors; we treat it as motivating context rather than a directly comparable empirical benchmark. Read against their central claim, that generative AI is positioned to become a standard, load-bearing component of software development practice, our finding that trust rather than transparency is the proximate driver of developer adoption carries a practical implication: if the trajectory described by Sauvola et al. materializes, organizations should prioritize investment in AI reliability and output quality, the antecedents of trust identified here and by Choudhuri et al. [12], over investment in explainability tooling alone.

A notable contextual contrast emerges when comparing these findings with those of Ali et al. [1], who studied AI coding assistant adoption among undergraduate computer science students using a similar extended TAM framework. In their study, Trust significantly predicted Perceived Ease of Use (β = 0.238, p = 0.003) but did not significantly predict Perceived Usefulness or Behavioral Intention directly. The present study finds a stronger and more direct path from trust to intention (β = 0.64, p = 0.007), suggesting that professional developers form a tighter cognitive link between trust and deployment decisions than student learners do. This difference is theoretically consistent with the higher-stakes nature of professional software development, where students may tolerate uncertainty in AI outputs as a learning exercise, but professional developers require a reliable trust foundation before committing AI-generated code to production systems. Taken together, these two studies suggest that the role of trust in AI adoption is not uniform across contexts, but scales with the consequentiality of the adoption decision.

The Agile development context introduces important moderating conditions that the current model does not fully capture. Delivery cycles structured around sprints create time pressure that may suppress developers’ capacity to thoroughly audit AI-generated code, compressing the trust calibration process. When trust is insufficiently developed, and transparency is high, the result may be cognitive overload rather than confident adoption, a dynamic noted by Radulovic and Klobucar [15] in their examination of friction points in Agile and AI integration.

Conversely, developers in sustained AI pilot programs, such as the Devin AI context from which participants in this study were partially drawn, may develop faster trust calibration due to repeated exposure and structured feedback loops. The moderate standard errors across models suggest that individual and contextual factors not captured in the current three construct model meaningfully affect adoption outcomes.

5.1. Limitations

Several limitations constrain the generalizability of these findings. First, the sample size of 19 represents a recognized challenge for quantitative research, given its small N. The borderline significance of the Transparency to Trust path (p = 0.070) is a direct consequence of limited statistical power, and the true effect size may be larger than observed. Future studies should target a minimum of 50 respondents to adequately power correlation and mediation analyses [23].

Second, reliance on self-reported Likert-scale data introduces social desirability and introspection biases. Developers may not accurately report their cognitive states regarding trust or transparency, and subjective experience of cognitive load may diverge from objective performance indicators [5]. A mixed-methods extension, including behavioral trace data such as acceptance rates of AI suggestions, time on task, and error rates, would strengthen construct validity.

Third, this study did not measure objective AI performance characteristics or code quality outcomes. The variables capture developer perceptions of the AI agent, not the agent’s actual accuracy or transparency. This distinction matters: developers may trust an AI agent that produces objectively flawed code, or distrust one that is objectively accurate, depending on their prior experience and mental models [6]. Additionally, future studies should incorporate core TAM constructs (perceived usability and perceived ease of use) alongside Transparency and Trust to examine their relative explanatory power.

An additional limitation concerns the internal consistency of the Transparency (α = 0.14) and Behavioral Intention (α = 0.31) measures. These values fall below commonly accepted reliability thresholds and suggest that the items may not consistently capture a single underlying construct. Consequently, relationships involving these variables should be interpreted cautiously, as measurement error may have attenuated observed correlations and regression coefficients. The findings should therefore be viewed as exploratory indications rather than evidence derived from fully validated measurement scales. Future work should refine the survey instrument and re-evaluate reliability using larger samples.

5.2. Alternative Interpretations

A skeptical reading of these results warrants consideration. The correlation between Trust and Behavioral Intention (r = 0.60) is strong but not deterministic, indicating that 36% of variance in Behavioral Intention is explained by Trust. The remaining variance suggests that organizational factors, management endorsement, team norms, and infrastructure readiness may play roles as important as individual cognitive constructs. Future research should incorporate multilevel modeling to disentangle cognitive effects at the individual level from adoption dynamics at the team and organizational levels.

Additionally, the cross-sectional design cannot establish causal directionality. It is possible that developers who already intended to deploy AI-generated code reported higher trust retrospectively, rather than trust preceding and causing adoption intention. Longitudinal designs tracking cognitive variables and adoption behavior over time are needed to establish a causal sequence. Ali et al. [1] similarly identified longitudinal design and broader construct inclusion, such as cognitive load and over-reliance, as important directions for future TAM-based AI adoption research, reinforcing the convergent need for more temporally sensitive study designs in this domain.

6. Conclusions and Recommendations

This study examined cognitive factors influencing AI coding agent adoption among Agile software developers. Trust in AI accuracy emerged as the primary cognitive determinant of Behavioral Intention to deploy AI-generated code. Transparency demonstrated a positive but non-significant association with Trust, while Trust emerged as the strongest predictor of Behavioral Intention. Future studies should investigate whether trust functions as an intervening mechanism using larger samples and formal mediation procedures. These findings yield three evidence-based recommendations for organizations seeking to improve AI adoption in development teams.

First, organizations should prioritize improving the accuracy and reliability of AI coding agents over implementing transparency features. When developers experience consistent, correct AI suggestions, trust forms naturally through experience. Trust, not transparency dashboards or explanation panels, is the proximate driver of deployment behavior.

Second, structured AI exposure programs should be designed to accelerate trust calibration. This includes monitored pilot periods with defined evaluation checkpoints, allowing developers to develop accurate mental models of AI agent capabilities and limitations through repeated experience. The Devin AI pilot context from which participants were drawn represents one such model.

Third, cognitive engineering interventions targeting Agile team dynamics should address the conditions under which Agile sprint pressure may suppress trust development. When AI-generated code must be reviewed rapidly within sprint cycles, insufficient trust calibration may lead to either uncritical acceptance or outright rejection of AI outputs, both of which are maladaptive adoption behaviors. Teams should explicitly budget cognitive overhead for AI review as a distinct task category in sprint planning.

Future research should replicate this study with larger samples, incorporate objective performance measures, and employ longitudinal designs to track trust and intention over time. Expanding the model to include constructs such as Self-Efficacy, Cognitive Load, and Team Trust would provide a more comprehensive cognitive engineering account of AI adoption in professional software development environments. Comparative studies across populations, contrasting professional developers with student learners, building on foundational work such as Ali et al. [1], would further clarify how stakes, experience, and organizational context moderate the trust adoption relationship.