Business and Social Sciences

1. Introduction

Walk into almost any large bank today and, somewhere behind the teller counter or the mobile app interface, an algorithm is quietly deciding who gets a loan, whose transaction looks suspicious, and whose profile reads as risky. This is not a futuristic scenario; it is, by now, fairly ordinary banking practice. Artificial intelligence has moved from a peripheral experiment to a working part of how financial institutions detect fraud, score credit, and personalize services, often processing volumes of structured and unstructured data that no human team could realistically review in real time (Mohamed & Yildirim, 2021). And to be fair, the appeal is obvious — faster decisions, lower costs, and in many cases, better accuracy than the manual processes they replace (Truby et al., 2020).

Yet something curious has happened alongside this efficiency gain. As AI systems have crept into every corner of banking — commercial lending, digital-only banks, investment platforms, fintech partnerships — the question of who is actually accountable for what these systems do has not kept pace (Rana et al., 2019). It is one thing to build a model that flags fraud with impressive accuracy; it is quite another to explain, months later, why it denied a particular customer's loan application, or to prove to a regulator that the model was not quietly discriminating along lines it was never supposed to consider. This is where governance, as distinct from mere technical performance, becomes the harder problem (Villar & Khan, 2021).

Part of the difficulty is structural. AI-driven banking systems are not static rule sets that can be audited once and left alone; they are dynamic, constantly retrained on new data, and therefore perpetually in motion in ways that make ongoing monitoring, validation, and regulatory oversight genuinely difficult (Nicoletti, 2021). Add to this a now-familiar list of concerns — limited model explainability, algorithmic bias creeping in through historical data, thinning human oversight, uneven data governance, cybersecurity exposure, and regulatory uncertainty that has not fully caught up with the technology — and it becomes clear why so many institutions are still finding their footing (Ebinger & Omondi, 2020). Left unaddressed, these gaps do more than create technical debt; they erode customer trust and can expose banks to real legal and reputational consequences when governance fails to keep pace with what the technology is actually doing (Sarker et al., 2021).

To their credit, regulators and international standard-setting bodies have not been silent on this. There has been a growing, if still uneven, push toward transparency, accountability, fairness, and continuous monitoring as the pillars of responsible AI governance across the financial sector (Hemphill & Kelley, 2021). What effective governance actually requires — coordinated data management, rigorous model validation, active risk monitoring, internal controls, regulatory compliance, and a genuine ethical framework for decision-making — is reasonably well articulated in principle (Gill, 2020; Yigitcanlar et al., 2020). The trouble, however, is translation: most banking institutions still lean on risk management approaches built for financial and operational risk, and these frameworks were simply never designed with algorithmic decision-making in mind (Ala-Pietilä & Smuha, 2021).

What is somewhat surprising, given how much has been written about AI adoption in finance, explainable AI, and regulatory compliance individually, is how little empirical work has actually asked banking professionals themselves whether their existing risk management frameworks feel adequate to the governance challenges AI introduces (Beccalli et al., 2020). Much of the literature remains conceptual or normative — describing what governance ought to look like — rather than measuring how governance gaps are actually experienced by the people managing these systems day to day. That gap is where this study sits.

This study, then, sets out to do something fairly specific: survey 155 professionals working across commercial banking, fintech, regulation, AI development, and academia, and ask them directly how six governance-related factors — model explainability, regulatory compliance, data governance, algorithmic bias, model risk, and cybersecurity risk — relate to their perception of governance challenges in AI-driven banking. The intention is not merely descriptive. By quantifying which of these factors weighs most heavily on perceived governance adequacy, the study aims to offer banking executives, regulators, and technology developers something more actionable than a general call for 'better governance' — a sense of where, specifically, traditional risk management frameworks are straining hardest, and where reinforcement might matter most.

2. Materials and Methods

2.1 Study Design

This study used a quantitative, cross-sectional survey design to examine perceived governance challenges associated with AI-driven banking systems and the adequacy of conventional risk management approaches in addressing them. A cross-sectional design was chosen deliberately: because the aim was to capture a single, comparable snapshot of professional perception across several industry sectors rather than to track change over time, this approach offered the most practical balance of feasibility and statistical power for the analyses planned (Tagde et al., 2021). It is worth being upfront, though, that this design constrains what can be claimed — associations, not causal sequences — a point we return to in the limitations.

2.2 Participants and Sampling Procedure

Eligible participants were professionals with direct working knowledge of artificial intelligence, banking operations, financial regulation, or risk management, recruited from commercial banks, financial technology companies, regulatory authorities, AI technology firms, and academic or research institutions. Recruitment combined purposive and convenience sampling: purposive, in that individuals were targeted specifically for their domain expertise, and convenience, in that they were reached through professional networking platforms, institutional email lists, and social media channels (Magnan & Markarian, 2011). We recognize that this non-probability approach limits statistical generalizability to the broader population of banking professionals, and we treat the findings accordingly as indicative of a knowledgeable, engaged subgroup rather than a representative cross-section of the industry.

An initial batch of 170 questionnaires was distributed. After removing incomplete responses and internally inconsistent answer patterns (for example, straight-lining across reverse-coded items), 155 questionnaires remained for analysis, yielding a completion rate of 91.2%. No formal a priori power analysis was conducted; the achieved sample nonetheless exceeds commonly cited minimum thresholds (roughly ten observations per predictor) for a six-predictor regression model, giving reasonable confidence in the stability of the coefficient estimates obtained.

2.3 Ethical Considerations

Participation was voluntary, and respondents were informed of the study's purpose, the confidentiality of their responses, and their right to withdraw at any point before submission. No personally identifying information was collected beyond broad demographic categories (Table 1). Consistent with standard practice for anonymous, minimal-risk survey research involving professional adults, informed consent was implied through voluntary completion of the questionnaire.

2.4 Instrument and Measures

The survey instrument comprised two sections. The first collected demographic information — gender, age bracket, educational attainment, years of professional experience, and employment sector (Howes et al., 2014). The second measured the study's six constructs using five-point Likert-type items, anchored from 1 (strongly disagree) to 5 (strongly agree) (Kröger, 2020). Item wording for the independent constructs — Model Explainability, Regulatory Compliance, Data Governance, Algorithmic Bias, Model Risk, and Cybersecurity Risk — was adapted from established conceptualizations in the AI governance and financial risk management literature (Wong, 2020), while the dependent construct, Governance Challenges, captured respondents' overall perception of how well traditional risk frameworks addressed governance gaps in algorithmic banking systems (Fenwick et al., 2018).

To make this reproducible for other researchers: the conceptual framework treated Model Explainability, Regulatory Compliance, Data Governance, Algorithmic Bias, Model Risk, and Cybersecurity Risk as independent variables and Governance Challenges as the dependent variable (Pagallo et al., 2019). Prior to full deployment, the instrument was reviewed for face and content validity by subject-matter reviewers familiar with AI governance in financial services, and minor wording adjustments were made to improve clarity. We note, as a limitation addressed further below, that formal psychometric validation (e.g., confirmatory factor analysis, reliability coefficients such as Cronbach's alpha) was not conducted for this dataset, and we recommend that future replications report these statistics explicitly.

2.5 Data Collection Procedure

Data were collected through a structured online questionnaire distributed via professional networking platforms, institutional email contacts, and relevant social media groups over the study period. This mode of distribution was selected to reach a geographically and professionally diverse pool of respondents within resource constraints typical of an independent academic study (Magnan & Markarian, 2011).

2.6 Statistical Analysis

All analyses were performed in IBM SPSS Statistics, Version 29. Descriptive statistics — frequencies, percentages, means, standard deviations, skewness, and

Table 1 | Demographic profile of surveyed banking, fintech, regulatory, and academic professionals (n = 155).
Respondents were recruited via purposive and convenience sampling across five professional sectors engaged with AI-driven banking systems. Percentages are calculated within each demographic category and may not sum to 100% owing to rounding. Data were collected via a structured online questionnaire (Section 2.2).

Table 2 | Descriptive statistics for governance-related constructs show uniformly high endorsement across all six predictors and the outcome variable. Values represent means and standard deviations for responses on a five-point Likert scale (1 = strongly disagree, 5 = strongly agree). Skewness and kurtosis values fall within accepted thresholds (±1) for approximate normality, supporting the use of parametric analyses (Pearson correlation, multiple linear regression) reported in Tables 3 and Figure 3. SD, standard deviation.

kurtosis — were computed first to characterize respondent demographics and study variables (Kar et al., 2021). Bivariate relationships between the six independent constructs and Governance Challenges were then examined using Pearson product-moment correlation (Taherdoost & Drazenovic, 2024). Before proceeding to regression, multicollinearity diagnostics were run using tolerance and the Variance Inflation Factor (VIF), supplemented by inspection of the condition index (Wong, 2021). Finally, multiple linear regression was used to estimate the independent contribution of each governance factor to perceived governance challenges, with statistical significance evaluated at the 95% confidence level (p < .05) (Wittwehr et al., 2019). Model assumptions — linearity, homoscedasticity, and independence of residuals (assessed via the Durbin–Watson statistic) — were checked prior to interpreting coefficients.

3. Results

3.1 Demographic Characteristics of Respondents

Of the 155 respondents, a modest majority were male (61.3%) relative to female (38.7%), and most fell within the 31–40 (26.5%) and 41–50 (24.5%) age brackets — suggesting, unsurprisingly perhaps, that this is a sample with real industry tenure rather than fresh entrants to the field (Table 1). Educational attainment skewed high: 40.0% held a master's degree, 24.5% a bachelor's degree, and 23.2% a doctoral degree. Professional experience was fairly evenly split between the 5–10 year (29.7%) and 15-plus year (29.0%) categories. By sector, commercial banking accounted for the largest share of respondents (36.1%), followed by financial technology (20.0%) and AI technology companies (16.8%) (Table 1).

3.2 Descriptive Statistics of Study Variables

Every study variable returned a mean above 3.90 on the five-point scale, pointing to a generally high level of agreement among respondents that these governance dimensions matter (Table 2). Model Explainability recorded the highest mean (M = 4.31, SD = 0.61), narrowly ahead of Governance Challenges itself (M = 4.27, SD = 0.58) and Regulatory Compliance (M = 4.22, SD = 0.66). Cybersecurity Risk, while still solidly endorsed, registered the lowest mean of the set (M = 3.96, SD = 0.76). It is worth flagging that this uniformly high agreement — with no variable dipping below the midpoint — likely reflects some ceiling effect typical of expert-sampled attitude surveys, a point we revisit in the discussion.

3.3 Perceived Governance Challenges

When asked directly to rank the governance challenges they encountered, respondents identified Model Explainability as the most pressing, accounting for 21.9% of responses (Figure 1), followed by Regulatory Compliance (18.7%) and Data Governance (16.8%). Algorithmic Bias followed at 14.2%, then Model Risk (11.6%) and Cybersecurity Risk (9.7%), with Operational Resilience trailing at 7.1% (Figure 1). The consistency between this ranked distribution and the mean-score ordering in Table 2 lends the finding some internal coherence rather than appearing as an artifact of a single measurement approach.

3.4 Perceived Benefits of AI Governance Frameworks

Respondents were similarly asked what they saw as the principal benefits of stronger AI governance (Figure 2). Improved transparency led the list (22.6%), followed by better risk control (19.4%) and improved regulatory compliance (16.8%). Fair decision-making (14.2%) and data security (11.6%) followed, with greater accountability (9.0%) and enhanced customer trust (6.4%) rounding out the responses (Figure 2). Read alongside Section 3.3, this suggests a fairly coherent picture: the same explainability and compliance concerns that respondents flagged as the biggest challenges are also, reassuringly, the areas where they expect the most benefit from better governance.

3.5 Correlational Analysis

Pearson correlations between the six predictors and Governance Challenges were uniformly positive (Figure 3). Model Explainability showed the strongest association (r = .756), followed by Regulatory Compliance (r = .704) and Data Governance (r = .671); Algorithmic Bias (r = .638), Model Risk (r = .612), and Cybersecurity Risk (r = .581) showed moderate associations. Inter-predictor correlations ranged from .488 to .641 — a moderate band that, while indicating some shared variance among constructs, remained comfortably below the .80 threshold typically taken as a warning sign for multicollinearity (Figure 3).

3.6 Multiple Linear Regression Analysis

The regression model was statistically significant and explained a substantial share of variance in perceived governance challenges (R² = .714, adjusted R² = .702, F(6,

Figure 1 | Model Explainability is the most frequently cited governance challenge in algorithmic banking systems, followed by regulatory and data-related concerns. Bars show the proportion of respondents (n = 155) identifying each factor as a primary governance challenge. Percentages reflect relative frequency of selection among seven governance categories and sum to 100%. Source data in Table 2.

Figure 2 | Improved transparency and risk control are perceived as the leading benefits of adopting stronger AI governance frameworks in banking. Bars show the proportion of respondents (n = 155) ranking each benefit as a primary advantage of enhanced AI governance. Categories are mutually exclusive per respondent selection.

Figure 3 | Pearson correlation matrix reveals moderate-to-strong positive associations among governance constructs, with no evidence of problematic multicollinearity. Colour intensity indicates correlation strength (Pearson's r); all coefficients are positive and statistically significant at p < 0.01. Inter-predictor correlations (r = 0.488–0.641) remain below the 0.80 threshold conventionally used to flag multicollinearity concerns, consistent with VIF diagnostics in Table 3.

Table 3 | Model Explainability, Regulatory Compliance, and Data Governance are the strongest independent predictors of perceived governance challenges in AI-driven banking. Standardized regression coefficients (β) derive from a multiple linear regression with Governance Challenges as the outcome variable and six governance constructs entered simultaneously as predictors (n = 155). Variance Inflation Factor (VIF) values below 2.0 indicate no meaningful multicollinearity; the Durbin–Watson statistic (1.95) indicates residual independence. Model fit: R² = 0.714, adjusted R² = 0.702, F(6,148) = 61.84, p < 0.001. β, standardized regression coefficient; SE, standard error; VIF, variance inflation factor.

148) = 61.84, p < .001) (Table 3). Model Explainability emerged as the strongest predictor (β = 0.318, t = 5.58, p < .001), followed by Regulatory Compliance (β = 0.264, t = 4.80, p < .001) and Data Governance (β = 0.213, t = 4.02, p = .001). Algorithmic Bias (β = 0.174, p = .001), Model Risk (β = 0.151, p = .002), and Cybersecurity Risk (β = 0.129, p = .006) contributed smaller, but still statistically significant, positive effects. Diagnostic checks supported the model's validity: VIF values ranged from 1.45 to 1.92 (Table 3), well under conventional thresholds for concern, and the Durbin–Watson statistic (1.95) indicated no meaningful autocorrelation in the residuals.

4. Discussion

Taken together, these results make a fairly persuasive case that governance in AI-driven banking cannot be reduced to a question of technological readiness alone — it is, at its core, a question of transparency, regulatory alignment, and disciplined data stewardship (D'Orazio, 2021). Among the six factors examined, Model Explainability stood out consistently: it carried the highest mean score, was named the single largest governance challenge, showed the strongest correlation with overall governance difficulty, and produced the largest standardized regression weight (β = 0.318, p < .001) (Table 2, Table 3). If there is one message this dataset delivers unambiguously, it is this one — explainability is not a peripheral concern for AI governance in banking; it appears to be close to its foundation (Akyüz & Mavnacıoğlu, 2021).

Regulatory Compliance followed closely behind, both in correlation (r = .704) and regression weight (β = 0.264, p < .001) (Table 3), which fits comfortably with the broader narrative in the literature that regulatory frameworks have simply not kept pace with how quickly AI capabilities have advanced in financial services (Hemphill & Kelley, 2021; Wong, 2021). Data Governance rounded out the top three predictors (β = 0.213, r = .671) (Table 3), a finding that, on reflection, makes intuitive sense: AI systems are only as trustworthy as the data feeding them, and weaknesses in data quality, security, or standardization tend to surface downstream as prediction errors and elevated operational risk (Sarker et al., 2021).

Algorithmic Bias, Model Risk, and Cybersecurity Risk, while statistically significant, carried comparatively smaller weights (Table 3) — not unimportant, but perhaps better understood as second-order concerns that compound once the more foundational issues of explainability, compliance, and data governance are addressed, rather than as independent governance failures in their own right (Akyüz & Mavnacıoğlu, 2021). This ordering is echoed in what respondents identified as the chief benefits of stronger governance — improved transparency (22.6%), better risk control (19.4%), and stronger compliance (16.8%) (Figure 2) — reinforcing the sense that AI governance in banking is not merely a compliance checkbox but a multidimensional undertaking involving ethical design principles, continuous model validation, sustained human oversight, and proactive cybersecurity posture (Beccalli et al., 2020).

The explanatory strength of the overall model (R² = .714, adjusted R² = .702) (Table 3) suggests these six factors capture a meaningful share of what drives perceived governance difficulty, though naturally not all of it — nearly 30% of the variance remains unexplained, which is a reminder that other factors (organizational culture, leadership commitment, or jurisdiction-specific regulatory maturity, for instance) likely matter too and were not captured here. It should also be said plainly that, because this is cross-sectional survey data drawn from a non-probability sample, these findings describe association, not causation, and should be interpreted as reflecting the perceptions of a knowledgeable but self-selected group of professionals rather than an objective audit of governance outcomes. The uniformly high mean scores across all constructs (all above 3.90) may additionally reflect a degree of social desirability or ceiling effect common in expert-sampled attitude research, tempering how far these results should be generalized.

With those caveats in mind, the overall pattern still points toward a fairly clear practical implication: conventional risk management frameworks, built primarily around financial and operational risk categories, appear stretched thin when asked to govern algorithmic decision-making, and the reinforcement most urgently needed seems to concentrate around explainability, regulatory alignment, and data governance rather than being evenly distributed across all six factors examined here.

5. Conclusion

This study offers empirical, if necessarily preliminary, evidence that traditional risk management frameworks are not well matched to the governance demands of AI-driven banking. Across 155 professionals surveyed, model explainability, regulatory compliance, and data governance emerged as the strongest drivers of perceived governance challenges, together with smaller but significant contributions from algorithmic bias, model risk, and cybersecurity concerns. Because the design is cross-sectional and the sample non-probability in nature, these findings should be read as suggestive rather than definitive, pointing toward association rather than causal proof. Even so, the consistency across descriptive, correlational, and regression analyses lends the pattern some credibility. Banks, regulators, and technology developers may find it useful to prioritize explainability and regulatory alignment when designing integrated AI governance frameworks, while continuing to build out bias mitigation, model risk management, and cybersecurity resilience as complementary, rather than substitute, priorities.

Acknowledgements

The author A.T.S. et al. thanks the banking, fintech, regulatory, and academic professionals who volunteered their time to complete the survey underlying this study. No dedicated funding was received for this research.

Author Contributions

A.T.S.: conceptualization, methodology, data curation, formal analysis, investigation, writing – original draft, writing – review and editing.

Competing Financial Interests

The author A.T.S. et al. declares no competing financial interests related to this work.

References