Verification Behaviour as a Link Between Trust in Augmented Analytics and Perceived Decision Quality

The increasing integration of artificial intelligence (AI) into business intelligence (BI) platforms has transformed how organizational users access and interpret analytical information. AI-generated insights, such as revenue trends, anomaly alerts, and forecasts, can provide decision-relevant information; however, users must still determine whether these outputs accurately reflect the underlying data and are appropriate for organizational use. This issue has become increasingly important as augmented analytics capabilities are embedded in BI platforms. These capabilities can respond to natural-language queries, detect anomalies, generate narrative summaries, and produce forecasts with limited support from specialist analysts [1]. Consequently, employees in sales, finance, operations, and human resources can perform analytical tasks that previously required substantial technical expertise or assistance from data analysts [2].

Recent research has increasingly positioned BI and AI as strategic organizational capabilities. BI capabilities have been linked to AI-enabled logistics innovation through improved decision support [3], and AI-enabled BI to digital marketing performance through data-driven decision-making [4]. Related studies have examined AI as a resource for improving managerial decision-making [5] and have considered its risks at institutional and governance levels [6]. Less is known, however, about whether individual users verify AI-generated insights before acting on them, or how such verification behaviour relates to their perceived decision quality.

From a management perspective, the question concerns how organizations divide analytical work between people and AI. Raisch and Krakowski [7] distinguish automation, in which machines take over a task, from augmentation, in which humans and machines work on it together, and argue that the two are interdependent: tasks that begin as human--AI collaboration can shift towards automation over time, and an overemphasis on either approach can set off self-reinforcing cycles with harmful organizational consequences. Augmented analytics is typically introduced as augmentation, with the AI preparing insights and the employee remaining responsible for judging them. Verification is the point at which this division of labour is enacted in daily work. When users stop checking, an arrangement designed as augmentation may operate in practice as unacknowledged automation. Examining verification behaviour therefore links an individual-level mechanism to a managerial choice about how far AI-generated insights should be relied upon without review.

Existing research offers two contrasting perspectives on the role of trust in this verification process.\linebreak Adoption-oriented research generally treats trust as a condition facilitating the acceptance and use of AI-generated recommendations, and the authors’ earlier, non-peer-reviewed preprint applied this view to non-technical BI users [8]. More broadly, trust is considered an important prerequisite for integrating AI into organizational routines and realizing its potential value [9]. In contrast, human-factors research suggests that high trust in automation can reduce users’ motivation to monitor and verify system outputs, a pattern commonly described as automation complacency [10-11]. Recent evidence has extended this concern to AI-based recommendations [12] and generative AI use at work, where confidence in AI tools has been associated with less critical evaluation of their outputs [13]. Accordingly, trust may relate to decision quality through two countervailing pathways: it may facilitate the adoption and use of AI-generated insights while simultaneously reducing verification and increasing the likelihood that inaccurate or contextually unsuitable outputs are accepted without scrutiny. This distinction is consistent with Schoeffer et al.’s [14] argument that reliance on AI should be conceptually separated from the quality of the decisions that result from that reliance.

Two research gaps limit current understanding of this relationship in organizational settings. First, existing evidence on AI overreliance has largely relied on experiments in which reliance is inferred from participants’ agreement with intentionally correct or incorrect recommendations [15-16]. Although such designs provide strong experimental control, they do not adequately capture how employees verify AI-generated outputs during routine BI use. Second, automation complacency has primarily been studied in controlled settings, including route-planning and medical-diagnosis tasks, rather than among non-technical employees using commercial BI platforms in their everyday work. It therefore remains unclear whether trust is associated with reduced verification in routine augmented-analytics use and whether verification behaviour helps explain the relationship between trust and perceived decision quality.

Accordingly, this study examines whether trust in augmented analytics is associated with less verification of AI-generated insights among non-technical users, whether verification behaviour mediates the relationship between trust and perceived decision quality, and how perceived explainability and AI literacy relate to verification behaviour, including whether AI literacy moderates the trust--verification relationship. These expectations are formalized in seven hypotheses (H1--H7), which are presented in Section 2.4.

To this end, this study analyses survey data from 350 non-technical BI users in Vietnam. Trust in the AI capabilities of BI software is negatively associated with self-reported verification behaviour ($\beta = -0.196$), whereas verification behaviour is positively associated with perceived decision quality ($\beta = 0.220$). Within the specified model, the resulting negative indirect association partially offsets the positive direct association between trust and perceived decision quality, so that the total association is approximately 14\% smaller than the direct association. The direct association between perceived system explainability and verification is not statistically significant, and the hypothesized moderating role of AI literacy received no statistical support.

This study makes three contributions. First, it conceptualizes verification as a distinct form of user behaviour and develops a ten-item verification behaviour scale (VER) with initial validation evidence, extending experimental research on AI overreliance to organizational survey research. Second, it examines whether trust is associated with perceived decision quality through two pathways with opposing signs, thereby clarifying that the positive trust--outcome relationship assumed in BI-adoption research may represent a net association. Third, it examines the roles of perceived system explainability and AI literacy in verification behaviour, including whether AI literacy attenuates the negative association between trust and verification.

The present study differs from the authors’ earlier non-peer-reviewed preprint [8] in its theoretical framing, analytical model and dataset. Whereas the preprint conceptualized trust as a uniformly beneficial component of an AI-adoption process grounded in cognitive delegation and analysed a sample of 250 respondents, the present study draws on automation-complacency theory to investigate whether trust may also be associated with reduced verification of AI-generated insights, and it analyses a separately collected sample of 350 respondents. The present study also introduces verification behaviour as a mediator between trust in augmented analytics and perceived decision quality.

Augmented analytics applies machine learning and natural language processing to automate data preparation, insight discovery and the explanation of results within BI platforms [1]. The organizational promise is self-service at scale: employees without analytical training can obtain answers without waiting for a specialist team. Research on self-service BI, however, indicates that such users frequently lack the expertise required to assess data scope, definitions and quality [2], and augmented features intensify this asymmetry. A narrative statement such as the one described in the introduction reaches the user already interpreted, so that the user’s role shifts from producing an analysis to accepting or questioning one. Logg et al. [17] found that non-experts weighted algorithmic advice more heavily than equivalent human advice, a tendency the authors termed algorithm appreciation, whereas experts discounted it. In a BI context, this implies the possibility that users who rely heavily on augmented analytics may be less inclined to critically examine its outputs.

Lee and See [10] define trust in automation as the attitude that an agent will assist in achieving one’s goals under conditions of uncertainty and vulnerability, and argue that favourable outcomes require calibrated trust, that is, trust commensurate with the actual capability of the system; overtrust can lead to misuse, whereas distrust can result in disuse. Parasuraman and Manzey [11] characterize the behavioural signature of overtrust as complacency, whereby operators reduce their monitoring and cross-checking of automated outputs, particularly when a system has demonstrated reliability and their attention is divided across competing demands. Two features of this account are relevant to the present study. Complacency is a matter of attention allocation rather than of ignorance, and has therefore been observed among experienced operators as well as novices. It may also remain undetected until the automated system fails, because a reliable system rarely penalizes omitted checks. Routine BI work satisfies both preconditions: most AI-generated summaries are correct on most occasions, and users consult them while engaged in other tasks. Evidence from generative AI points in the same direction. In a survey of knowledge workers, Lee et al. [13] found that higher confidence in AI tools was associated with lower self-reported critical thinking, whereas higher confidence in one’s own abilities was associated with higher levels of critical thinking. Harbarth et al. [12] further demonstrated that attitudinal and behavioural dimensions of complacency may move in opposite directions: a transparent system reduced users’ expectations of delegating tasks to it while increasing complacent behaviour. This finding cautions against treating trust and reduced checking as equivalent constructs.

Verification behaviour is defined here as the frequency with which users check AI-generated insights against underlying data, the scope of the analysis, alternative sources, or input from other people before using them in reports, communications, or decisions. The construct is behavioural and is therefore distinct from trust, which is an attitude towards the system, and from perceived explainability, which is a perception of what the system discloses. It is also distinct from reliance, since a user may rely on an insight after verifying it, the pattern that Schoeffer et al. [14] identify as a condition for humans to improve upon AI recommendations. Following the scale-development procedure proposed by MacKenzie et al. [18], the construct domain was specified to cover four forms of checking: inspecting the data and scope behind an insight, recomputing or cross-checking its figures, consulting colleagues or data specialists, and assessing its plausibility against one’s own knowledge.

Figure 1 summarizes the research model. Self-confidence in one’s own analytical skill, frequency of AI feature use, tenure with AI features and organizational function are included as controls on verification behaviour and perceived decision quality.

Explainability and trust. Lee and See [10] identify an understanding of how an automated system operates and the processes underlying its outputs as one of the three foundations of trust. A review of empirical work on trust in AI lists explainability among its most consistent antecedents [9], and a meta-analysis of 90 studies reports a significant but modest positive correlation between explainability and trust [19]. A competing argument holds that explanations may expose flaws in an AI’s reasoning and thereby reduce trust. This argument is set aside in the present context because BI explanations largely describe the data scope and calculation behind routine summaries and rarely reveal reasoning errors to non-technical users.

H1: Perceived explainability is positively associated with trust in augmented analytics.

Explainability and verification. Explanations may also reduce the effort required to check an insight. Vasconcelos et al. [16] showed that explanations can reduce overreliance by lowering the cost of verifying AI-generated answers, and Schemmer et al. [20], who define appropriate reliance as the ability to distinguish correct from incorrect AI advice, demonstrated experimentally that explanations shape it. In a BI context, an explanation that names the period, region or data source behind a figure provides users with the material needed to check it. The competing mechanism is substitution, whereby users read the explanation in place of checking. Bansal et al. [15] found that explanations increased acceptance of AI recommendations irrespective of their correctness, and Harbarth et al. [12] observed more complacent behaviour with a transparent system. The positive prediction is retained because BI explanations direct attention to verifiable data rather than to model internals, although the substitution account is treated as the principal alternative.

H2: Perceived explainability is positively associated with verification behaviour.

Trust and verification. Complacency theory predicts that trust reduces the perceived need to monitor [11], and survey evidence links confidence in AI tools with less critical evaluation of their outputs [13]. A competing view holds that well-calibrated trust need not reduce checking, since users who trust a system based on its demonstrated reliability may still verify high-stakes outputs [10]. In routine BI use, however, trust is built largely from accumulated experience of a system that is usually correct, which is precisely the condition under which monitoring declines.

H3: Trust in augmented analytics is negatively associated with verification behaviour.

Verification and perceived decision quality. Verification can help identify errors in data scope, definitions or calculations before they enter a decision. Schoeffer et al. [14] show formally that human--AI complementarity depends on the ability to override incorrect advice, which requires users to evaluate the advice before relying on it. Visinescu et al. [21] also report that the quality of the information acted upon is associated with perceived decision quality in BI settings. Verification also carries costs, since it consumes time and may delay decisions. These costs are set aside here because the outcome of interest concerns how good and how successful decisions were judged to be rather than how quickly they were reached.

H4: Verification behaviour is positively associated with perceived decision quality.

Trust and perceived decision quality. This hypothesis tests an association that is widely assumed in adoption-oriented research and is included as scaffolding for the mediation test rather than as a novel claim. Users who trust a system act on its insights instead of disregarding them, thereby avoiding the losses associated with algorithm aversion [22], and the authors’ earlier preprint proposed the same association for non-technical BI users [8], which the present study tests in an independent sample.

H5: Trust in augmented analytics is positively associated with perceived decision quality.

The mediating role of verification. If H3 and H4 hold, trust should also relate to perceived decision quality indirectly, through reduced verification, and that indirect association should be negative. Zhao et al. [23] classify a mediation pattern in which the direct and indirect effects have opposite signs as competitive mediation, whereby the indirect effect offsets part of the direct effect in the total association. This is the specific form that the complacency argument takes for decision outcomes [11].

H6: Verification behaviour mediates the relationship between trust and perceived decision quality, such that the indirect association is negative.

The moderating role of AI literacy. AI literacy refers to the competencies required to evaluate AI critically, to work with it and to use it as a tool [24]. The elaboration likelihood model holds that individuals scrutinize information when they possess both the motivation and the ability to do so [25]. Users with higher AI literacy may be better able to recognize where augmented analytics tends to fail and they may therefore be less likely to let trust reduce their verification efforts. The opposite expectation is also plausible, since users who feel competent with AI may become more comfortable delegating to it, and Parasuraman and Manzey [11] note that expertise does not necessarily protect against complacency. Because AI literacy is measured rather than manipulated, the hypothesis concerns a conditional association rather than a causal moderation.

H7: AI literacy weakens the negative association between trust in augmented analytics and verification behaviour.

The study employed a single cross-sectional survey of non-technical BI users working in organizations in Vietnam. Recruitment proceeded through four channels: occupational groups on Facebook (129 respondents, 36.9\%), direct contact through a customer list (95, 27.1\%), LinkedIn (67, 19.1\%) and referral by earlier respondents (59, 16.9\%). Responses were collected in a single survey round between 3 and 26 September 2026. The questionnaire defined BI software, AI features and “information produced by AI” using workplace examples, and four screening questions restricted participation to employed respondents working outside data, BI and information technology (IT) functions (A1--A2), who did not write code, build dashboards or data models, or operate statistical software themselves (A3), and who had used at least one built-in AI feature of their BI software in the preceding three months (A4). Respondents who had not made a decision or recommendation based on AI-generated information in that period (H0) were not required to complete the decision-quality section. Of the 428 questionnaires returned, 4 incomplete responses were removed, together with 8 completed in less than 120 seconds, 13 responses that failed an attention-check item embedded in the questionnaire (6 of which also contained identical answers to all ten verification items), 3 duplicate records and 50 respondents who answered “No” to H0. This left 350 valid responses, representing 81.8\% of the returned questionnaires; the sample profile is summarized in Table 1.

Item-level missing values were rare (10 of the 11,900 responses across the 34 scale items, or 0.08\%) and were replaced with the corresponding item means; nine were left blank and one was an out-of-range entry on EXP3.

The inverse square root method indicates that detecting a path coefficient of 0.15 with 80\% power at a one-tailed 5\% significance level requires 275 observations [26], so the sample of 350 exceeds the requirement for the main paths. It does not meet the requirement for interaction effects of the magnitude typically observed in survey research, a point to which Section 5.3 returns.

Unless stated otherwise, items used five-point agreement scales; full item wording and response anchors are provided in Table A1. Perceived explainability (EXP, five items) was adapted from the explanation satisfaction measures reviewed by Hoffman et al. [27] and reworded for AI-generated insights in BI software. Trust in augmented analytics (TRUST, seven items) was adapted from McKnight et al. [28], covering reliability (four items) and functionality (three items). AI literacy (LIT, three items) and perceived decision quality (DQ, four items) were adapted from Pinski and Benlian [29] and Visinescu et al. [21], respectively. DQ assessed respondents’ evaluations of AI-informed decisions with known outcomes. VER was measured using a newly developed ten-item frequency scale.

The items were translated from English into Vietnamese and back-translated [30], then reviewed by two experts for clarity and relevance to BI work. Expert review prompted minor adjustments to BI terminology. The questionnaire was subsequently pre-tested with 30 users, following the sample-size recommendation of Perneger et al. [31]. All pre-test participants reported clear comprehension, and no further wording changes were made.

Control variables included self-confidence in analytical skill (single item), AI feature use frequency (five categories), AI feature tenure and organizational function. The latter was represented by four dummy variables, with sales and other functions as the reference group.

The VER scale covered the four checking domains defined in Section 2.3, with item assignments provided in Table A1. VER8, which concerns reading an insight’s explanation, was assigned to data and scope checking because BI explanations describe the underlying data scope and calculations (Section 2.4).

VER was specified as reflective because its items were treated as interchangeable, covarying expressions of a general tendency to verify insights before use [18]. This specification is consistent with the complacency account underlying H3 [11]. A formative specification remains defensible because the items describe distinct checking acts (Section 5.3).

To address common method bias from the single questionnaire [32], constructs appeared in separate blocks with individual instructions, verification used frequency rather than agreement anchors, decision quality was measured last, and responses were anonymous. Statistical checks included full collinearity variance inflation factors (VIFs) for the five focal constructs, with values above 3.3 indicating bias [33], and a marker variable model [34]. The four-item product retention tendency marker [35], a consumer trait theoretically unrelated to the study constructs, was placed between the trust and verification blocks.

Partial least squares structural equation modelling (PLS-SEM) in SmartPLS 4 [36] was chosen for the prediction-oriented model, new scale, single-item controls and interaction term [37]. Estimation used path weighting, standardized data and mean replacement for missing values. Item skewness ($-0.23$ to 0.05) and excess kurtosis ($-0.67$ to $-0.20$) showed no marked non-normality; non-normal prediction errors motivated mean absolute error (MAE) comparisons with a linear model (LM).

Following confirmatory composite analysis [38], assessment covered loadings, internal consistency (Cronbach’s $\alpha$, $\rho_A$, $\rho_C$), convergent validity (average variance extracted, AVE) and discriminant validity (heterotrait--monotrait ratio, HTMT). Criteria were loadings $\geq 0.708$, reliability 0.70--0.95, AVE $\geq 0.50$, inner VIFs $< 3$ [37], and upper bounds of one-tailed HTMT bootstrap confidence intervals (CIs) below 0.85 [39].

Path tests used 10,000 bootstrap resamples, two-tailed percentile CIs and a fixed seed; mediation followed Zhao et al. [23]. H7 used the two-stage approach [40], with $f^2$ benchmarks of 0.005, 0.01 and 0.025 for small, medium and large effects [41]; conditional effects were estimated at AI literacy’s mean $\pm 1$ standard deviation.

Out-of-sample prediction was assessed with PLSpredict (10 folds, 10 repetitions, fixed seed, early-antecedent prediction) [42] and the cross-validated predictive ability test (CVPAT) [43]. $Q^2_{\mathrm{predict}}$ used the training-sample mean benchmark; item-level MAE comparisons used LM, and CVPAT used indicator-average (IA) and LM benchmarks. Two robustness checks re-estimated the model without controls or with the trust--verification path reversed; sensitivity analysis excluded VER8 and VER10.

Table 2 summarizes the measurement model results.

The measurement model meets the established criteria, with one qualification concerning the new verification scale. Item loadings ranged from 0.731 to 0.866 for all items except two verification items, VER8 (0.499) and VER10 (0.410). Following Hair et al. [44], items loading between 0.40 and 0.70 are removed only where removal is required to bring reliability or AVE above threshold. Because VER already reached $\rho_C = 0.909$ and AVE = 0.506 with both items included, they were retained in order to preserve the content domain of the new scale, and they are examined further below and in Section 5.1. Cronbach’s $\alpha$ ranged from 0.815 to 0.886, $\rho_A$ from 0.821 to 0.903 and $\rho_C$ from 0.890 to 0.911, and outer VIFs from 1.160 to 2.364.

Discriminant validity was established. All HTMT values among the focal constructs were at or below 0.328, and the upper bounds of the corresponding one-tailed 95\% confidence intervals did not exceed 0.418 ( Table 3). The highest HTMT value involving a control variable was 0.341, between VER and self-confidence. The Fornell--Larcker criterion was also satisfied (Table A4).

There is no evidence that common method variance poses a substantial threat to the results. Full collinearity VIFs for the five focal constructs, computed from the latent variable scores, ranged from 1.083 to 1.238. Correlations between the marker and the focal constructs ranged in absolute value from 0.008 to 0.092. In the marker model reported in Table A2, none of the marker’s paths to TRUST, VER or DQ was significant (all $p \geq 0.186$), and the reported main-model path coefficients changed by at most 0.005 without altering any inference.

Turning to the structural model, inner VIFs ranged from 1.000 to 1.439 and the standardized root mean square residual (SRMR) was 0.059. The main-model $R^2$ values were 0.091 for TRUST, 0.200 for VER and 0.117 for DQ ( Figure 2); the corresponding adjusted $R^2$ values were 0.089, 0.177 and 0.093, respectively.

Table 4 reports the hypothesis tests. Perceived explainability was positively associated with trust ($\beta = 0.302$, $p < 0.001$), supporting H1. Its direct association with verification was small, negative and not significant ($\beta = -0.071$, $p = 0.191$), so that H2 is not supported. Trust was negatively associated with verification ($\beta = -0.196$, $p < 0.001$) and verification was positively associated with perceived decision quality ($\beta = 0.220$, $p < 0.001$), supporting H3 and H4. Trust was also directly and positively associated with perceived decision quality ($\beta = 0.304$, $p < 0.001$), supporting H5.

In the moderation model using the full ten-item VER scale, $R^2$ for VER was 0.205. The TRUST $\rightarrow$ VER coefficient in this model was $-0.200$. Inner VIFs ranged from 1.000 to 1.444, and the interaction term had a VIF of 1.031. The SRMR for the saturated model was 0.050.

In the main model, verification mediates a negative indirect association between trust and perceived decision quality ($\beta = -0.043$, 95\% CI [$-0.083$, $-0.015$]). With the direct path positive and the indirect path negative, the pattern corresponds to competitive mediation [23], supporting H6.

Table 5 presents the additional paths, controls and effects. Among the controls in the main model, self-confidence in one’s own analytical skill showed the strongest association with verification ($\beta = 0.287$, $p < 0.001$, $f^2 = 0.094$), whereas the associations of use frequency and AI feature tenure were not significant at the 5\% level. Seven of the eight organizational-function coefficients had p-values above 0.05 and 95\% confidence intervals that included zero (Table A6). The exception was marketing, which, relative to the reference group of sales and other functions, was associated with higher perceived decision quality ($\beta = 0.285$, $p = 0.055$); its percentile bootstrap confidence interval lay entirely above zero (95\% CI [0.005, 0.581]). Because the $p$-value, which is based on the bootstrap standard error, and the percentile interval lead to different conclusions at the 5\% level, this coefficient is reported as marginally significant.

The total association of trust with perceived decision quality ($\beta = 0.261$, 95\% CI [0.159, 0.363]) is therefore approximately 14\% smaller than the direct association, so that, within the specified model, the negative indirect association through verification partially offsets the positive direct association. Explainability also related to verification through trust ($\beta = -0.059$, 95\% CI [$-0.102$, $-0.027$]), and its total association with verification was negative and significant ($\beta = -0.130$, $p = 0.016$). The serial path from explainability through trust and verification to perceived decision quality was small but excluded zero ($\beta = -0.013$, 95\% CI [$-0.027$, $-0.004$]).

Figure 2 summarizes the main-model estimates for H1--H6 and the moderation-model estimate for H7; controls are omitted for clarity. The interaction between AI literacy and trust on verification was not significant ($\beta = -0.076$, $p = 0.142$, 95\% CI [$-0.173$, 0.030]) and its effect size was small ($f^2 = 0.007$), so that H7 is not supported. AI literacy itself was positively associated with verification ($\beta = 0.165$, $p = 0.001$ in the moderation model). The simple slopes ran counter to the hypothesized direction: the trust--verification association was $-0.124$ ($p = 0.076$) at one standard deviation below the mean of AI literacy and $-0.276$ ($p < 0.001$) at one standard deviation above. The difference between the slopes was not statistically reliable, however, and the index of moderated mediation also included zero ($\beta = -0.017$, 95\% CI [$-0.046$, 0.006]).

Table 6 summarizes the predictive assessment, and Table A3 reports the item-level PLSpredict results. $Q^2_{\mathrm{predict}}$ was positive for TRUST (0.082) and VER (0.098) but negative for DQ ($-0.031$), so the model predicted DQ less accurately than the training-sample means. Because the prediction errors departed from normality, the LM comparison used MAE [42]: all seven trust items and seven of the ten verification items had lower prediction errors than the benchmark, indicating high and medium predictive power respectively. All four DQ items also had lower prediction errors than the LM benchmark, although their $Q^2_{\mathrm{predict}}$ values were negative. In CVPAT, the model outperformed the IA benchmark for TRUST and VER but not for DQ, and the LM benchmark for TRUST and DQ but not for VER. The two benchmarks answer different questions: the training-sample mean shows whether the antecedents carry out-of-sample information about an outcome at all, whereas the LM shows whether the model’s structure predicts as well as an unconstrained regression on the same items. Predictive support for DQ is therefore limited and depends on the benchmark: the model predicted DQ better than the LM benchmark but not better than the item means.

The robustness checks answer different questions (Table A2). In the model without controls, the hypothesized paths retained their direction and statistical significance classification. The indirect association remained negative ($\beta = -0.042$, 95\% CI [$-0.080$, $-0.015$]). The reversed model offers an alternative account of the same associations and does not confirm the direction of the proposed mediation: verification was negatively associated with trust ($\beta = -0.176$, $p = 0.001$). The direct explainability--verification path is sensitive to this choice. It was not significant in the main model ($\beta = -0.071$, $p = 0.191$) but was significant in the reversed model ($\beta = -0.136$, $p = 0.012$), close to its total association in the main model ($\beta = -0.130$). Among the other paths estimated in both the main and reversed models, none changed direction or statistical significance.

Table A7 compares the full ten-item scale with an eight-item scale excluding VER8 and VER10 in the main model with controls; H7 is estimated in the moderation model. Without the two items, the composite reliability of VER rose from 0.909 to 0.919 and its AVE from 0.506 to 0.587, while the trust--verification path, the verification--decision quality path and the specific indirect association kept their sign and significance and changed by at most 0.013. The direct explainability--verification path changed from $-0.071$ ($p = 0.191$) to $-0.095$ ($p = 0.050$).

Additional robustness checks with the eight-item scale (not tabulated) yielded a reversed VER $\rightarrow$ TRUST path of $-0.175$ ($p = 0.001$), an indirect TRUST $\rightarrow$ VER $\rightarrow$ DQ association of $-0.039$ ($p = 0.015$) without controls, and marker paths with $p \geq 0.200$. Both items are retained to preserve the content domain of the scale.

Overall, trust in augmented analytics is associated with less verification of AI-generated insights (H3).\linebreak Verification mediates a negative indirect association between trust and perceived decision quality that operates alongside a larger positive direct association (H4--H6). In the main model, perceived explainability had a negative indirect association with verification through trust and a negative total association, while the direct association was not statistically significant. AI literacy was positively associated with verification, but its hypothesized moderating role received no statistical support (H7).

First, the negative trust--verification association is consistent with an automation-complacency explanation in knowledge work involving commercial BI software. Parasuraman and Manzey [11] associated complacency with reduced monitoring under conditions of high reliability and divided attention. The observed association is consistent with this explanation among users who were interpreting a narrative rather than monitoring an ongoing process. However, this association does not constitute direct evidence of complacency, as complacency was not measured. Moreover, because the reverse path was also statistically significant, the finding could equally reflect the possibility that users who engage in more verification tend to develop lower levels of trust. The cross-sectional design does not allow these alternative interpretations to be distinguished.

Second, the competitive mediation suggests that the positive trust--outcome association assumed in adoption-oriented research, including the authors’ earlier preprint [8], may be interpreted as a net figure: within the specified model, the total association is approximately one seventh smaller than the direct association. Given the cross-sectional design and the use of perceived decision quality, these findings do not demonstrate that trust reduces decision quality by lowering verification. Nevertheless, this finding is consistent with the argument of Schoeffer et al. [14] that reliance and decision quality should be modelled separately, and with the distinction between calibrated trust and overtrust drawn by Lee and See [10]. From a management perspective, the pattern is consistent with the automation--augmentation paradox [7], in which tasks designed for human--AI collaboration can drift towards automation, but whether greater trust leads users to check less over time, and thereby weakens the human contribution to augmentation, cannot be determined from these data.

Third, in the main model, the direct association between perceived explainability and verification was not statistically significant, whereas both the indirect association through trust and the total association were negative. The negative indirect association is more consistent with the substitution account proposed by Bansal et al. [15] and Harbarth et al. [12] than with the cost-reduction account proposed by Vasconcelos et al. [16]. Because explainability was measured as a perceived construct rather than experimentally manipulated, and the direct path was sensitive to model specification ($-0.071$ to $-0.136$ across the models in Section 4), this does not show that explanations cannot encourage checking; whether the explanations in respondents’ BI tools directed users to the data needed for checking was not examined here. AI literacy was associated with more verification, but its hypothesized moderation received no statistical support; the simple slopes pointed the other way, in line with the observation that expertise does not necessarily protect against complacency [11], but their difference was not statistically reliable. Self-confidence in one’s own analytical skill was the strongest correlate of verification, echoing Lee et al. [13].

Fourth, the newly developed VER scale performed adequately but not uniformly, and its validation evidence comes from a single sample. VER8, which records reading the explanation displayed with an insight, loaded weakly (0.499), and its cross-loading on perceived explainability (0.273; Table A5) left the narrowest margin of any item, which suggests that, although the item was assigned to inspecting data and scope (Section 3.2), reading an explanation is closer to consuming the system’s account than to checking it. VER10, which records assessing plausibility against one’s own knowledge, describes a mental rather than an observable act. Excluding both items changed the H3, H4 and H6 estimates by at most 0.013 (Table A7). Both items are candidates for revision and validation in an independent sample.

For managers, the results suggest that augmentation may need to be maintained by design rather than simply assumed [7]. The findings motivate two proposed actions and one test for organizations implementing augmented analytics. Their effectiveness in improving decision-making was not examined in this study and therefore requires evaluation in organizational practice. Figure 3 positions these steps along the pathway from an AI-generated insight to a decision and specifies, for each, the responsible actor, the trigger, the measure, and the supporting results.

Action 1: put the evidence one click away. Because perceived explainability showed no significant direct association with verification, the first proposal focuses on what explanations display rather than how many are provided. Whenever an AI feature is introduced or rolled out, BI governance teams can display the filter scope and data source alongside each generated insight, with a one-click route to the underlying table, and monitor the proportion of insights accessed at source before being used in a report or decision. Part of the time that AI saves in preparing analyses could be reallocated to this step.

Action 2: require a second check above a value threshold. Because trust was associated with less verification, while verification was associated with higher perceived decision quality, department heads can require confirmation from a second source or a second person for decisions based entirely on AI-generated insights that exceed a value threshold set by the organization. The proportion of such decisions with a recorded check shows whether the requirement is applied (uptake), and the number of errors detected before use shows whether the check identifies problems (evaluation). Such a requirement would act as a cognitive forcing function, which has been shown to reduce overreliance on AI in experimental settings [45].

Test: rehearse verification routines in addition to general AI literacy. AI literacy was positively associated with verification, but its hypothesized moderation (H7) received no statistical support, so the results neither support nor rule out general AI-literacy training. Learning and development teams could therefore pilot training that rehearses concrete verification routines using the organization’s own dashboards with a comparison group, and assess its effectiveness through the pre--post change in VER scores relative to the comparison group rather than through test results.

The study adopts a cross-sectional design; therefore, the identified paths represent associations, and the reversed model indicates that the direction of the trust--verification relationship cannot be established from these data; automation complacency and objective decision outcomes were not measured. Decision quality is perceived and self-reported, and was collected in the same questionnaire as its predictors, so respondents with limited analytical expertise may misjudge their own decisions, while those who verify, or trust AI features, may rate their decisions more favourably. Predictive support for perceived decision quality is limited and depends on the benchmark used. A formative specification of VER is defensible because its items describe distinct acts, two items require revision, and the scale requires validation in an independent sample. The convenience sample of non-technical BI users in Vietnam, recruited through social-media groups, a customer list and referral, may differ from users in other settings in organizational culture and BI maturity, and excluding the 50 respondents who had not acted on AI-generated information may over-represent users who rely on AI output. Decision stakes, industry, system quality and organizational data culture were not measured and may partly account for the observed associations. AI literacy was measured rather than manipulated, and the study was not powered to detect a small interaction: applying the inverse square root method [26] to an interaction coefficient of the observed magnitude (0.076), with 80\% power at a one-tailed 5\% level, indicates that about 1,070 respondents would be required. Future research could use a two-wave design with later decision outcomes, scenario experiments with planted errors, system logs of verification and field evaluations of the proposed steps.

This study examined how trust in augmented analytics relates to non-technical BI users’ verification of AI-generated insights and to their perceived decision quality. Trust was negatively associated with verification, while verification was positively associated with perceived decision quality, resulting in a negative indirect association alongside a larger positive direct association. In the main model, perceived explainability was positively associated with trust, whereas its direct association with verification was not statistically significant. AI literacy was associated with more frequent verification, but its hypothesized moderating role received no statistical support. These associations are consistent with automation complacency, but may also reflect users who develop lower trust as a result of more frequent verification. Predictive support for perceived decision quality depends on the benchmark used, and the verification scale has initial validation evidence only. The proposed steps remain to be evaluated in organizational practice, and the findings, based on non-technical BI users in Vietnam, may vary across organizational and technological contexts.