Artificial intelligence and bias in banking and financial services: a comprehensive review and future research agenda

Artificial intelligence (AI) enables machines to perform business tasks and activities traditionally executed by humans, including learning, reasoning and decision-making (Cao and Zhang, 2025; Mostafa, 2025). By mimicking human intelligence, AI systems can analyze large datasets, recognize complex patterns and generate predictions that enhance automation, personalization and analytic capabilities (Vlačić et al., 2021). Within the banking and financial services sectors, AI has evolved into an ecosystem that integrates big data analytics, machine learning (ML), natural language processing (NLP) and robotic process automation (RPA), collectively transforming value creation and customer experiences (Carbó-Valverde et al., 2025; Diniz et al., 2024; Lu and Calabrese, 2023).

Recent research emphasizes AI's role as a transformative infrastructure driving digital innovation and strategic renewal in banking (Amaral et al., 2023; Chai et al., 2025; Ho and Chow, 2024). As Ho and Chow (2024) suggest, financial institutions deploy AI-powered chatbots, robo-advisors and voice assistants to deliver continuous, personalized support and to facilitate seamless transactions. Through deep learning and natural language generation, these systems interpret customer intent, simulate conversational empathy and cocreate digital experiences that combine functional efficiency with emotional engagement (Bin-Nashwan et al., 2025; Matthews et al., 2024; Rancati and Maggioni, 2023). Facial and voice recognition technologies have simplified onboarding and authentication processes, improving accessibility for diverse consumer segments. For example, institutions such as the Hong Kong Monetary Authority's FinTech 2025 initiative (Ho and Chow, 2024) and the European Financial Data Space (Chomczyk Penedo and Trigo Kramcsák, 2023) highlight how AI ecosystems increasingly depend on human–machine collaboration to enhance transparency, responsiveness and competitiveness. Despite these advances, AI adoption remains uneven across geographies and organizational types, reflecting asymmetries in technological readiness (Diniz et al., 2024; Tigges et al., 2024), managerial vision (Guenther et al., 2023) and regulatory maturity (Amaral et al., 2023; Bockel-Rickermann et al., 2025; Nguyen, 2024). This uneven diffusion indicates the need for further research on how organizational, cultural and institutional contexts shape AI adoption across global banking systems.

The recent shift from generative to more agentic systems, defined as autonomous models capable of independent reasoning and contextual decision-making, has led AI to perform better in terms of speed, accuracy and customer personalization (Bornet et al., 2025; Eastwood, 2023). However, the costs of potential mistakes are very high, and firms are evaluating combining AI with human augmentation depending on the task complexity (Grewal et al., 2025). The literature collectively portrays AI as both a technological and organizational innovation that enhances efficiency and customer engagement but also introduces new vulnerabilities related to opacity, bias and accountability (Carbó-Valverde et al., 2025; Guenther et al., 2023). Therefore, this study aims to synthesize the current state of research to understand the evolution of AI-driven systems and human-AI interactions in banking and financial services.

Bias in AI has long been recognized as both a technical and social construct, traditionally referring to the assumptions made by a specific model (Mitchell, 1997). Recent research interprets it as “the inclination or prejudice of a decision made by an AI system which is for or against one person or group, especially in a way considered to be unfair” (Ntoutsi et al., 2020, p. 3). For example, within banking and financial services, algorithmic bias manifests when automated credit-scoring or fraud-detection systems disadvantage certain groups based on gender, race, or socioeconomic status (Akter et al., 2021; Fu et al., 2021; Hurlin et al., 2026; Yang et al., 2025). Scholars have proposed several frameworks to classify and mitigate such biases (Ntoutsi et al., 2020) Here, with the goal of focusing on the banking sector, we follow the categorization of Mehrabi et al. (2021), which conceptualize bias within a data–algorithm–user feedback loop, where pre-existing prejudices in datasets are amplified through algorithmic design and user interaction, perpetuating self-reinforcing cycles of discrimination.

Scholars have widely recognized the connection between cognitive and algorithmic bias, showing that bias originating in human reasoning can migrate into machine-learning processes (Akter et al., 2021; Gurdgiev and O'Loughlin, 2020; Hasan et al., 2022). The full pipeline of algorithm-driven design and the potential biases associated with each step is depicted in Figure 1. Originally introduced by Tversky and Kahneman (1974), the concept of cognitive bias explains how individuals make biased-based assumptions and rely on mental shortcuts, or heuristics, in their decision-making. Johnson (2021) argues that AI systems often reproduce the same implicit biases found in human reasoning because algorithms generalize from historically biased data much as humans generalize from experience. Indeed, studies in the banking and financial sectors reveal that auditors' judgments (Hsieh et al., 2020) and investment decisions (Matthews et al., 2024) can be shaped by machine-derived facial or behavioral cues.

On the other hand, AI is often leveraged to detect and, in some instances, mitigate these biases (Hsieh et al., 2020; Annu and Tripathi, 2025). For example, Król and Król (2021) utilize the concept of cognitive bias to investigate how ML techniques can be used to detect anomalies in eye movements related to financial decision-making, which may indicate bias in the decision-making process. Huang (2026) extends cognitive-bias theory to marketing research to explain how bounded rationality influences consumers' trust in automated financial tools. Annu and Tripathi (2025) suggest that AI can detect and mitigate social and ethical biases, thereby supporting more informed investment decision-making. These findings suggest that AI can both reflect and help identify cognitive distortions in banking and financial services, providing behavioral corrections (Athota et al., 2023). Although organizations are developing bias-awareness strategies, the persistence of AI biases raises questions among scholars about whether technical remedies alone can eliminate discrimination and whether additional regulation is required for nondiscriminatory AI-based decision-making (Lauretta and Santamaria, 2025).

To reduce bias and the consequent discrimination that leads to unfairness, it is important first to define fairness. Different attitudes, perspectives and cultural differences lead to various ways of understanding fairness (Saxena, 2019), including parity, equalized odds and counterfactual fairness (Mehrabi et al., 2021). In marketing, fairness means making decisions without any prejudice or favoritism toward a person or a group based on their intrinsic or acquired traits in the context of decision-making (Saxena et al., 2019). Existing literature shows progression toward more specific and quantitative approaches to fairness, including percentile-equivalence modeling (Varley and Belle, 2021), interpretability-driven fairness (Repetto, 2025) and reinforcement-learning adjustments (Zhang et al., 2024). However, fairness remains a context-dependent construct influenced by cultural values and institutional norms (Hurlin et al., 2026; Millo et al., 2024; Momtaz, 2021). While prior reviews (Singh et al., 2025) have examined the broad applications of AI across banking and financial operations, there is still a lack of research focused on algorithmic bias. Therefore, more research on this topic is needed.

In the remaining sections, we review and summarize the current state of AI and bias research within the area of banking and financial services. The bibliometric analysis examines publication/citation trends over time, most influential publications and journals, authorship trends and different keyword assessments. The systematic literature review assesses various methods, theories, variables, research designs and types of bias that have been applied in the domain. Lastly, we identify four overarching research themes to provide a conceptual synthesis of how AI and bias are understood, operationalized and governed in banking and financial services. For these themes, we explore future research directions to further enhance the knowledge on AI and bias in the banking and financial industry.

We examine AI and biases in banking and financial services research by adopting a two-phase, mixed-method approach combining bibliometric analysis with a systematic literature review (Bahmani et al., 2025; Hentzen et al., 2022; Shi et al., 2025). The first phase includes various bibliometric analyses to quantitatively assess the intellectual landscape of the research domain. Statistical tools and metrics, including the number of publications, total citations and keyword occurrence, offer further insights into the social and structural relationships (Donthu et al., 2021; Krey et al., 2022). The second phase qualitatively synthesizes theoretical foundations, methodological approaches and empirical designs, allowing for macro- and micro-level findings.

Consistent with previous research, we completed a comprehensive literature search using Scopus and Web of Science (WoS) databases, applying predefined inclusion and exclusion criteria (Ahamed and Limbu, 2024; Bahmani et al., 2025; Hentzen et al., 2022; Lim et al., 2024). The search was restricted to English-language peer-reviewed articles published in A* and A-ranked journals based on the Australian Business Deans Council (ABDC) list to ensure high quality in publications. Moreover, we included all research fields of the ABDC in the search to guarantee a comprehensive overview of the research domain.

The search focused on titles, abstracts and keywords using terms related to AI and bias in banking and financial services. Our search string was as follows: TITLE-ABS-KEY (algorithm OR “artificial intellige*” OR AI OR “digital transform” OR “cloud comput*” OR “chatbot” OR “virtual assist*” OR analytic* OR “big data” OR blockchain OR cryptocurrency OR “voice assist*”) AND (bias OR prejudice OR discrimination OR racism) AND (bank* OR financ* OR mortgage OR “payday lending” OR “buy now pay later” OR BNPL OR fraud OR credit* OR loan). The search included all articles published up to August 2025. As summarized in Figure 2, we followed the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines for study identification, screening, eligibility and inclusion (Ahamed and Limbu, 2024; Page et al., 2021).

The initial search identified a total of 5,686 articles across both databases. After removing 217 duplicate records, 454 unique articles remained for screening. During the abstract review stage, 280 papers were excluded for falling outside the study's scope (e.g. articles that mentioned banking or credit without directly examining AI bias). Further assessment of the full-text articles led to the exclusion of 109 additional papers, resulting in a final sample of 65 articles.

Prior to analyzing the data, we reviewed titles, authors, countries and keywords to ensure complete, consistent and accurate data (Bahmani et al., 2025; Lim et al., 2024). First, article titles and author names were manually cleaned when necessary. Second, 264 author-provided keywords were reviewed to identify keywords to be combined or split into multiple terms. For example, keywords such as “artificial intelligence” and “AI” were combined. Moreover, missing keywords were added as needed, with original articles consulted whenever possible. For around 5% of the articles, the Orange topic modeling software was used to generate keywords. To verify the accuracy of these software-generated keywords, two authors reviewed those keywords. These changes resulted in a final keyword list of 233 terms. Third, the spelling of some countries was updated, such as “United States” was changed to “USA” and “United Kingdom” was changed to “UK”.

As previously discussed, this study applies a multi-method approach incorporating bibliometric analysis and a systematic review. The bibliometric analysis integrates mathematical and statistical methods to assess quantitative indicators (e.g. journals and publication year frequency) and structural interconnections (e.g. co-occurrence) (Hentzen et al., 2022; Shi et al., 2025). VOSviewer was utilized to map these relational structures. Within the network visualizations, each link represents the strength of association between items, with co-occurrence analysis identifying underlying thematic relationships and conceptual trends (Donthu et al., 2021; Krey et al., 2022). Moreover, colored nodes denote individual keywords linked together within an integrated network (van Eck and Waltman, 2020), where node size corresponds with keyword frequency (i.e. larger nodes represent greater frequency). This approach facilitates the identification of knowledge clusters and the intellectual structure of the research domain (Donthu et al., 2021).

Complementing the bibliometric analysis, a systematic literature review synthesized information across the articles and detected common themes (Ahamed and Limbu, 2024; Shi et al., 2025). Individual studies were first examined vertically and then cross-compared horizontally to identify patterns across research studies (Krey et al., 2022). A detailed coding framework was implemented, encompassing publication details (e.g. title, year, authors), theoretical foundations, constructs, measures, methodological approaches, type of bias, context of AI use and study design.

The publication and citation trends from 2011 to 2025 reveal important developments in research productivity and impact (see Table 1). During this period, the data reflect a growth in publication quantity, fluctuating citation influence and an expected decrease in citations in the most recent two years despite continued increases in the number of articles published per year. These patterns highlight the dynamic nature of research productivity and influence over time.

While the first article on AI and bias in banking and financial services was published in 2011, the early years reflect a low and inconsistent publication output with some gaps (e.g. 2015–2016) and only a total of three publications in the first six years. Starting in 2017, the number of publications steadily increased until reaching 15 articles in 2024 and 2025. Overall, this upward trend reflects the growing research activity and interest in this domain.

Considering the number of citations, inconsistent citation patterns are apparent throughout the time period with citation increasing in 2012 (275 citations), 2018 (383 citations), 2020 (639 citations) and 2023 (413 citations), leading to an overall peak with 652 citations in 2021 (see Figure 3). Regarding average citations, 2018 is the most influential year with an average citation count of 191.5. After 2020, annual citations declined, reaching 132 citations in 2022 and 76 citations by 2025. Furthermore, average citations per article also begin to decline, suggesting an increase in published articles and a decrease in citations per article. This decline could have multiple reasons, including an increased number of publications focusing on similar topics, changes in citation behavior and the recency effect (i.e. lagged recognition for recent articles).

Overall, 3,360 citations across 65 publications in 15 years reflect a maturing research domain characterized by increasing publication output and consistent citation performance. Future research should examine whether citation counts increase again over time or whether the domain moves toward a broader and richer research stream that remains less cited.

The top 20 journals publishing AI and bias research within banking and finance are summarized in Table 2. These journals reflect a wide range of areas, including banking, finance, management, information systems, marketing and economics. The top journal based on publication output is the Journal of Behavioral and Experimental Finance (4 publications; e.g. Gurdgiev and O'Loughlin, 2020), followed by Management Science (3 publications; e.g. Hurlin et al., 2026), Journal of Business Venturing (3 publications; e.g. Matthews et al., 2024) and Annals of Operations Research (3 publications; e.g. Repetto, 2025). The remaining journals either published one or two articles. Interestingly, 45% of the top 20 journals and 50% of the top four journals are ranked A* based on the ABDC list, indicating the high-quality research published in this domain.

In terms of total citations, the Journal of Behavioral and Experimental Finance (499 citations) outperforms all other journals, including Management Science (441 citations) and Journal of Business Venturing (297 citations). Interestingly, the top three journals with the highest average citation (AC) count, i.e. highest impact, are International Journal of Forecasting (AC = 275), University of Chicago Law Review (AC = 251) and Information Systems Research (AC = 180). All three journals have only published one article, yet these articles have already accumulated a substantial number of citations. Overall, the domain of AI and bias in banking and finance is widely dispersed across various journals and research fields, indicative of a broad disciplinary reach and multidisciplinary research interest.

The analysis of the top 20 most influential articles considers total citations (TC) and citations per year (CPY). Table 3 depicts these top 20 articles and corresponding metrics.

The most influential article based on citation count is Ban et al. (2018)'s assessment of ML and portfolio optimization with 375 total citations. Next follows Crone and Finlay (2012)'s study on credit scoring with 275 citations and Momtaz (2021)'s article on tokens and corresponding moral dangers with 274 citations. A review of average citation per year depicts a different ranking. Specifically, the most influential article is Giudici and Raffinetti (2023) (CPY = 80), focusing on AI in finance, followed by Momtaz (2021) (CPY = 68.5) and Ban et al. (2018) (CPY = 53.6). These three articles indicate recent yet highly impactful research, especially in the area of AI applications in finance and behavioral analytics.

Moreover, the most influential articles are primarily published in Management Science, International Journal of Forecasting, Journal of Business Venturing and Finance Research Letters, underscoring high academic quality and visibility across numerous fields. Specifically, management, finance, banking and econometric appear as the most representative research fields based on ABDC classification.

Broadly speaking, the most impactful articles relate to topics such as AI and ML (e.g. Ban et al., 2018), behavioral finance and biases (e.g. Athota et al., 2023), and credit scoring and discrimination (e.g. Crone and Finlay, 2012). The most recent shift to assessing AI and biases reflects an increased importance in these topics and a trend that most likely will continue in future research publications.

Around 200 authors have contributed to the domain of AI and bias in banking and financial services. The majority of publications (89.2%) are co-authored and only 7 publications are authored by a single scholar (e.g. Nguyen, 2024). The articles feature an average of 3.1 authors per publication. Over time, the average number of authors remained relatively steady around 3.0–3.5 authors with a small decrease in 2021 (mean = 2.1). The most common collaboration patterns are teams of four (32.3%), followed by teams of two (23.1%) and teams of three (21.5%). To examine how team size influences impact, total citations were analyzed based on the number of authors per publication. Results show that articles with two authors are cited the most with an average of 73.7 citations, closely followed by teams of three (mean = 73.0) and single-authored papers (mean = 72.3). Interestingly, larger groups of four or five/six authors seem to be cited less frequently with corresponding averages of 28.2 and 16.6. Overall, articles with two or three authors reflect the greatest citation impact.

Next, we review authorship trends based on location, i.e. countries. Specifically, countries were counted based on the author and institution level instead of the article level. As such, one article can reflect multiple countries. Moreover, only countries with at least two publications were included in Table 4 to summarize the most productive countries. The top three countries based on publication output are the United Kingdom (UK) (20 publications), the United States (USA) (18 publications) and China (16 publications), producing over 50% of the total literature. Considering total citations, the USA surpasses the UK with 1,795 citations compared to 1,106 citations, followed by China (378 citations) and India (337 citations). Therefore, India appears to have produced fewer yet more influential articles. Interestingly, the average citation counts indicated that Ireland produces the most impactful research with 130.5 average citations, followed by the USA (99.7 average citations) and Italy (89 average citations). Overall, while the UK dominates in terms of quantity, Ireland produces the most impactful research with Europeans, Asian and North American countries dominating global research productivity. High average citations in a few countries, such as India and Singapore, indicate impactful research despite fewer publications and a more diverse scholarly presence in this specific research domain.

A total of 233 author keywords provide insights into the content and topics explored within AI and bias in banking and financial services research. Overall, the most frequently occurring keywords are ML (23 times), AI (19) times and credit scoring (11). Moreover, discrimination, risk and fairness are also popular keywords with 8 occurrences each. While keyword occurrence offers initial insights into the popular content, the keyword co-occurrence analysis identifies the interconnections among terms and the subsequent foundation of the literature (Donthu et al., 2021; Krey et al., 2022).

Figure 4 depicts the top 50 author keywords co-occurrence network. Four distinct clusters emerge: 1. red cluster “ethical AI and bias in financial decision making” (19 keywords), 2. green cluster “ML and bias in financial systems” (17 keywords), 3. blue cluster “financial risk in behavioral finance” (7 keywords) and 4. yellow cluster “trustworthiness in corporate governance and audits” (7 keywords). Overall, the first two clusters and the last two clusters are relatively balanced regarding cluster size.

The red cluster is the largest and most diverse, dominated by AI and foundational challenges. Some of the core concepts include deep learning, decision-making and financial planners as discussed by Hasan et al. (2022) and Athota et al. (2023). Beyond these topics, ethics, bias, fairness and explainable AI also define this specific cluster and indicate a growing interest in responsible AI adoption in financial decision-making. For example, Tigges et al. (2024) emphasize that while AI and ML can enhance credit accessibility, they also raise important concerns about bias and discrimination in financial systems. Additionally, newer topics such as blockchain, cryptocurrency and bitcoin complete the red cluster. Carbó-Valverde et al. (2025) find that financial literacy bias, particularly overconfidence, significantly influences cryptocurrency ownership, underscoring how cognitive biases affect financial behavior in digital assets markets.

The green cluster includes the most frequently occurring keyword, namely ML. Based on the size of the nodes, discrimination and algorithm bias in ML and credit scoring are especially of interest. Moreover, this cluster also highlights potential issues in the banking and financial industry, such as sampling bias, stereotypes and gender discrimination (e.g. Brock and De Haas, 2023). Newer financial services, including crowdfunding (Yoo et al., 2022) and fintech (Huang, 2026), are also nested within this cluster. Specifically, Yoo et al. (2022) suggest that AI or ML can be used to detect facial expressions of project initiators in crowdfunding. While these expressions can (positively) influence funding decisions, they can also introduce bias in banking or financial institutions. Brock and De Haas (2023) also identify that restricting information available to loan officers does not disproportionately impact female loan applications, indicating that AI or ML could help mitigate gender bias in lending decisions.

While the blue and yellow clusters are comparable in size, both clusters encompass less than half of the keywords than the first two clusters. The blue cluster is characterized by risk and associated terms such as behavioral finance (e.g. Hasan et al., 2022) and portfolio optimization (e.g. Ban et al., 2018). More recent terms such as robo-advisor also appear in this cluster. Tubadji et al. (2021) suggest that the consumers' willingness to adopt robo-advisory AI in banking is impacted by cultural aspects, with lower social capital and fear of being cheated being important factors that reduce consumers' preference to adopt AI in banking services. Furthermore, Hasan et al. (2022) posit that AI and ML can help financial planners overcome cognitive biases, such as confirmation and hindsight biases, improving the overall decision-making process. The yellow cluster reflects balanced nodes with regard to size, indicative of the absence of a main keyword. The two most frequently occurring keywords are big data and facial expression with cognitive bias and trustworthiness following close behind. Hsieh et al. (2020) discuss how auditors tend to charge higher audit fees to companies led by CFOs perceived as untrustworthy based on facial features, indicating how ML-based facial trustworthiness measures can reflect auditors' cognitive biases. Furthermore, Athota et al. (2023) argue that while AI and ML can help reduce cognitive biases in financial planning, uncertainty about whether financial planners will fully rely on AI-based decision-making remains.

Next, we explore citable topics based on average citation counts for the top 50 author keywords. In Figure 5, warmer colors (e.g. yellow) represent higher average citation counts compared to cooler colors (e.g. blue), which represent lower average citation counts. The most-cited keywords include blockchain, cryptocurrency, bitcoin, behavioral finance and diversification. Moreover, risk, robo-advisor and crowdfunding also appear highly cited. In contrast, mortgage, reject interference, loan and race are among the least cited keywords. Interestingly, ML, fintech and algorithm bias are more cited than general terms like AI and bias. Consequently, authors who prioritize quantitative impact should consider content and keywords related to highly cited terms such as cryptocurrency, robo-advisor and diversification.

As a final keyword assessment, we evaluate the thematic evolution of the top 50 keywords as depicted in Figure 6. The temporal map features recent time periods in warmer colors and older time periods in cooler colors. Foundational topics in AI and bias in banking and financial services consist of risk, index tracking, behavioral finance and reinforcement training. Over time, the focus shifts toward credit scoring, sampling bias and eventually discrimination, ML and big data. More recent research explores areas of fairness, behavioral bias, bias mitigation and AI. Finally, topics such as sustainability, trust, ethics and social justice have gained the attention of researchers in the last few years.

For the second phase of our research, we implement a systematic literature review to offer a deeper understanding of AI and bias within the banking and finance sector domain. Specifically, this assessment examines theories, common methodologies, frequently applied variables, types of bias and research designs.

Studies examining AI and bias within banking and financial services research rely on a diverse set of theoretical frameworks and paradigms (see Table 5). About 64.6% of these works are grounded in at least one theoretical premise. Although some papers do not explicitly mention an underlying theory, around 32.3% apply multiple theoretical perspectives to frame their investigations.

The most frequently employed theoretical framework is the concept of cognitive bias (6.9%), introduced by Tversky and Kahneman (1974). This framework explains how people relying on mental shortcuts, or heuristics, can make systematic errors. Within the current research stream, this framework examines both human and technological biases, such as using AI to detect bias in financial decision-making (e.g. Król and Król, 2021). Another widely adopted theoretical framework is prospect theory (5.6%) by Kahneman and Taversky (1979). This theory suggests that individuals assess results based on perceived gains and losses rather than absolute results. This theory shows how loss aversion and risk perceptions influence biased decisions and how AI models can mitigate such biases to improve financial judgement (e.g. Hasan et al., 2022).

The evaluation of variables is based on 48 quantitative publications. Figure 7 summarizes antecedents, outcomes, mediators, moderators and control variables across the assessed research domain. During the systematic literature review, a clear pattern emerged distinguishing between internal and external variables. Specifically, internal variables refer to factors intrinsic to the decision-making entity (e.g. individual attributes, psychological traits and biases), whereas external variables capture outcomes and conditions in the broader financial and market environment. Since several papers use secondary data to report variables grouped together (e.g. “loan and borrower characteristics”) without specifying each element, exact counts for individual variables (e.g. age) could not always be determined. For this reason, main constructs are grouped under the broader clusters, such as personal attributes.

As shown in Figure 7, antecedents cover a wide range of internal and external variables. The most common internal category is personal attributes (35.4%), including age, gender, race and income level, among others. Other notable internal antecedents include creditworthiness indicators, such as debt-paying ability and growth ability (4.17%). Among external antecedents, culture (6.3%) and finance market sentiment (6.3%) are also prominent, with financial market sentiment specifically representing a market-level variable. Together, these variables address how biases influence loan applications and affect individuals across demographic groups. Bias (8.3%) is also frequently included in studies, as it spans both internal (e.g. financial literacy bias) and external (e.g. aggregation bias) dimensions.

Regarding outcome variables, the most frequently analyzed internal outcomes are investment behavior (10.4%), followed by technology adoption behavior (8.3%) and fairness perception (4.2%). Among external outcomes, loan performance (54.2%) is the most prominent, with default status being the most commonly studied indicator within this category. Moreover, lending decision outcomes (14.6%), market performance (14.6%) and model performance (14.6%) also appear as commonly studied outcome variables.

Among the moderating variables, the most frequently examined are personal attributes (12.5%) and bias (10.4%), which are both internal in nature. For mediating variables, trust (6.3%) and bias (4.2%) are the most commonly utilized in this research area. Interestingly, personal attributes appear across multiple roles, serving as an antecedent, moderating or a control variable depending on the study's design, highlighting their central significance in the AI bias literature.

Based on Figure 7, four propositions are advanced across four main areas: internal antecedent factors, external antecedent factors, mediating mechanisms and moderating mechanisms. Collectively, these propositions offer a structured foundation for the conceptual framework and can be utilized in future research to examine individual relationships within this specific research domain.

Grounded in behavioral finance, theories on credit risk and technology acceptance literature (e.g. Bin-Nashwan et al., 2025), the conceptual model posits that internal user attributes and external factors determine behavioral and structural outcomes. Proposition 1 argues that internal and external antecedents directly shape internal user outcomes (e.g. fairness perceptions, technology adoption and repayment behaviors) by forming cognitive and contextual perspectives for individual decision-making. Additionally, Proposition 2 asserts that these cross-level inputs, specifically external antecedents like market sentiment paired with internal personal attributes, directly dictate external structural outcomes such as credit risk, lending decisions and algorithmic model performance. Together, these propositions establish that financial technology ecosystems are fundamentally embedded within a matrix of micro-level psychology and macro-environmental realities.

To further explore relationships and limits of these pathways, the framework integrates important mediating and moderating dynamics. Drawing on trust-commitment and socio-technical systems theories, Proposition 3 positions internal mediators (e.g. system trust, reliability) and internal biases (e.g. perceived facial expression bias) as vital forces that translate baseline antecedents into downstream behavioral and financial outcomes, acknowledging that raw inputs require cognitive or affective evaluation before manifesting as performance. Finally, utilizing contingency theory (Donaldson, 2001), Proposition 4 establishes that these relationships are conditionally bounded rather than static. It introduces internal moderators (e.g. AI perceived usefulness, ease of use) and external moderators (e.g. decision-making context, system characteristics) as variables that amplify or attenuate the antecedent-outcome linkages, offering a highly nuanced, ecologically valid model of technology-mediated financial behavior.

Lastly, the conceptual framework introduces controls researchers need to incorporate to accurately examine relationships. Empirically testing the propositions needs to account for internal controls such as facial attractiveness and user demographics. Additionally, to accurately isolate main, mediating or moderating effects, external controls such as specific AI model type or ML model need to be accounted for.

Table 6 summarizes commonly applied methodological approaches across the examined publications. Overall, 57 articles (87.7%) employ empirical analysis, 7 articles (10.8%) are conceptual and 1 study is a review (1.5%). Among the empirical work, most adopt a quantitative approach, with only 15.8% using qualitative methods. Around 37.5% of studies rely on a single analytical technique, while the rest apply multiple analytical approaches, indicating a trend of employing a multi-method research approach. Regarding qualitative tools, interviews (5.1%) are the most common, followed by ethnographic analysis (1.7%), secondary data analysis (1.7%) and hypothetical situation analysis (0.8%).

The most commonly employed statistical method is regression analysis (38 studies), including a range of techniques such as binary logistic, OLS, linear mixed-effects, linear probability, logistic, univariate quantile and bivariate probit. Following this, ML and AI-based predictive models are widely used (36 studies), including methods such as artificial neural network, support vector machine, causal forest, k-prototype clustering, ML dimensionality reduction, reject inference and object selection algorithm. Additionally, structural equation modeling is another notable approach, applied in seven studies (5.9%).

To better understand how biases manifest in AI and ML research in banking and financial services, we categorize biases into three overarching categories: algorithm-to-user, data-to-algorithm and user-to-data (Mehrabi et al., 2021). Table 7 provides an overview. While Mehrabi et al. (2021) identify a broader range of bias types, we only present those that were explicitly observed in the reviewed articles. Within the algorithm-to-user category, four sub-biases were identified, namely algorithmic bias (n = 11), evaluation bias (n = 8), user interaction bias (n = 3) and emergent bias (n = 3). These biases occur when algorithmic outputs influence or interact with user perceptions, decisions or evaluations processes. For instance, Schultz and Fabozzi (2022) argue that AI and ML in banking and financial institutions can perpetuate bias, particularly in lending and credit decisions, representing algorithmic bias.

The data-to-algorithm category captures biases embedded in how data are collected, represented, or processed by AI systems. It includes sampling bias (n = 6), representation bias (n = 6), measurement bias (n = 8), linking bias (n = 1) and omitted variable bias (n = 1). These biases primarily arise when flawed, incomplete, or unrepresentative data distort algorithmic learning or outcomes. For instance, Shi et al. (2023) suggest that multi-step convex optimization approach based on MSW-LASSO can reduce bias in parameter estimation for sparse index tracking in high-dimensional settings, which is relevant to financial institutions and AI/ML applications, representing measurement bias.

Finally, the user-to-data category encompasses biases introduced through human behavior, culture or cognition that shape data generation and interpretation. This category includes aggregation bias (n = 1), social bias (n = 3), behavioral bias (n = 14), content-production bias (n = 1), historical bias (n = 1) and population bias (n = 1). Such biases reflect how human and societal factors influence data inputs and ultimately affect AI-driven financial decision-making. For instance, Athota et al. (2023) suggest that AI and ML can help reduce cognitive biases in financial planning, but there is uncertainty about whether financial planners will fully rely on AI-based decision-making, representing behavioral bias.

To further the study design, we examined important elements such as procedure and sample size across the evolutionary phases of AI (Huang and Rust, 2021; Bornet et al., 2025): mechanical (i.e. performing physical tasks), analytical AI (analyzing data), generative AI (creating new content) and agentic AI (acting autonomous) (Table 8). Our coding focused on the specific AI type (i.e. mechanical, analytical, generative and agentic) that was substantively examined and meaningfully explored in each study (e.g. discussed in the results section). As such, incidental mentions of AI types in introductions or discussions, among others, were not classified in our coding. Of the 65 articles, none of the studies explicitly examined mechanical AI as the primary focus of analysis; 59 studies employed analytic AI, 16 used generative AI and 5 used agentic AI. Studies employing analytic AI had the largest sample size (mean = 286,796), followed by generative (mean = 3,2081.1) and agentic (mean = 2,248.8). The large analytic AI samples reflect the predominance of secondary data (31 studies), which typically involve extensive datasets. Female participants averaged 50.3% in agentic AI studies, 41.3% in generative AI and 32.2% in analytic AI.

Regarding procedures, analytic AI relied mainly on secondary or remote data collection (41 studies) versus in-person methods (8 studies). For agentic and generative AI, the difference between remote and in-person collection was smaller. Among studies reporting sample details, most analytic AI research drew data from loan applicants (22%), secondary firm-related samples (18.6%), employees (11.9%) and students (5.1%).

This article identifies the characteristics, trends and future directions of the literature on AI and bias in banking and financial services. Findings suggest that while AI can be used to both detect and mitigate biases, consistent with the finding of Annu and Tripathi (2025), it can also be associated with biases that arise through continuous interactions between data, algorithms and users.

The bibliometric results show that publication output has increased steadily since 2017, with a peak in citations in 2021. Top publication outlets, such as the Journal of Behavioral and Experimental Finance and Management Science, highlight the inherently multidisciplinary nature of this research stream. The thematic analysis of author keywords identifies four main clusters: ethical AI and bias in financial decision-making; ML and bias in financial systems; financial risk in behavioral finance and trustworthiness in corporate governance and audits. Emerging topics such as blockchain and cryptocurrency reflect a shift toward the use of AI in ethically sensitive financial applications.

The systematic literature review reveals that most articles are theory-driven, with cognitive bias and prospect theory frequently applied to explain how heuristics and loss aversion influence financial decisions and how AI may moderate these effects. Studies largely focus on personal attributes (e.g. age, gender and race) as the major antecedents and loan-related outcomes (e.g. default and performance) as the major outcome variables. Taken together, these findings underscore the growing importance of AI and fairness in shaping financial decision-making and access to financial opportunities. The review also addresses three major bias categories (i.e. algorithm-to-user, data-to-algorithm and user-to-data) (Mehrabi et al., 2021). It suggests that, among the four types of AI, existing research has predominantly examined analytic AI, while mechanical AI remains comparatively underexplored.

To further examine the context of AI applications in banking and financial services and to develop a future research agenda, all 65 articles were reviewed to determine how AI was used. This step provides a more context-sensitive understanding of how AI is actually implemented across different financial settings. Specifically, two authors independently coded the studies and identified four themes: (1) AI as a strategic infrastructure for digital transformation in banking and financial services, (2) human–algorithm interaction and the duality of algorithmic opacity – toward bias mitigation and behavioral correction, (3) fairness frameworks and responsible AI governance – trust, ethics and systemic stability and (4) behavioral finance in the age of AI and ML. Moreover, articles were also coded to identify understudied methods and variables. Specifically, we conducted a cross-theme comparison to identify gaps in methods and variables that were not examined in all four themes. Table 9 provides an overview of the four themes, corresponding context examples, future research agendas, along with proposed methods and variables.

AI has progressively evolved into a core strategic infrastructure that underpins the digital transformation of banking and financial services (Ho and Chow, 2024; Mostafa, 2025). As institutions transition toward data-driven ecosystems, it is possible to observe an increase in the deployment of ML algorithms for credit scoring, fraud detection, portfolio optimization, liquidity management and systemic-risk monitoring (Ban et al., 2018; Fu et al., 2021; Hurlin et al., 2026; Repetto, 2025). Within this transformation, AI no longer represents a collection of isolated applications but a multi-layered infrastructure that aligns technology, data and governance to reshape the financial enterprise (Giovine et al., 2024).

This multi-layered structure is evident in the organizational moves made by leading banks, which have begun to establish executive-level AI mandates and center on excellence, thus institutionalizing capabilities and standards across their businesses (Giovine et al., 2024). Collectively, these studies show that AI is increasingly embedded across core financial functions, indicating a shift from isolated applications to system-level integration. Combining this systemic perspective with the studies under this theme, an integrated framework composed of four interdependent layers can be conceptualized, representing the New Strategic Infrastructure (see Table 10).

The “engagement layer” facilitates interactive exchanges and personalized service recovery (Bin-Nashwan et al., 2025; Ho and Chow, 2024; Mostafa, 2025) through conversational AI (text/voice/visual), robo-advisors, digital twins to simulate customer and employee behavior, and biometric authentication systems that enable continuous, adaptive and emotionally attuned interactions (Bin-Nashwan et al., 2025; Ho and Chow, 2024). The “decision-making layer” functions as the cognitive AI core, integrating big data analytics, ML, NLP and RPA into orchestrated systems that automate complex judgments (Bornet et al., 2025).

Next, the “core technology and data layer” provides the infrastructure necessary for scalable and secure data flows (Cao and Zhang, 2025; Vlačić et al., 2021) and enables real-time sensing, learning and decision execution, allowing banks and financial institutions to respond dynamically to market volatility, cybersecurity threats and evolving consumer expectations (Amaral et al., 2023; Guenther et al., 2023). The “operating model layer” embeds governance, transparency and accountability across the system, while also coordinating the functioning of the other layers to ensure transparency, fairness and accountability (Giudici and Raffinetti, 2023; Hurlin et al., 2026). From a managerial perspective, this multi-layered configuration implies that successful AI adoption depends on alignment across technological capabilities, data governance and organizational structures, rather than fragmented or function-specific implementation efforts.

Despite these advances, the literature remains limited in explaining how such infrastructures are implemented across different organizational contexts and how institutional constraints shape their effectiveness. In particular, there is still little understanding of how these systems operate in practice across diverse banking environments. Therefore, future research on Theme 1 should deepen the understanding of AI as a multi-layered infrastructure across different geographies and organizational types. For example, local and regional banks often display risk aversion due to high implementation costs, limited executive sponsorship and skepticism toward opaque algorithmic decisions (Diniz et al., 2024).

Further, because organizational inertia and uncertainty surrounding accountability, data governance and customer acceptance constrain implementation (Tigges et al., 2024), scholars could explore how different stages of development across data, technology and governance shape firm performance. In addition, longitudinal and multi-level designs could examine how AI infrastructure interacts with institutional strategies, internal policies and regulatory environments, influencing strategic agility, risk management and innovation diffusion across banking and financial services ecosystems. Moreover, based on the SLR, existing articles within this theme predominantly rely on quantitative methods (e.g. regression, SEM and ML models), which leaves important process-level insights underexplored. This highlights the need for more qualitative research, which remains understudied. Therefore, future research could benefit by focusing more on conducting qualitative studies (such as case studies) of banks implementing AI systems.

Human–algorithm interaction can be viewed as a new frontier in understanding how cognition shifts, and it is shared between human and machine agents within algorithmic ecosystems. Studies under this theme reveal that cognitive biases are not eliminated by automation. Rather, AI systems often replicate and amplify human biases and heuristics through biased data, heuristic model design and human oversight practices (Hsieh et al., 2020; Yoo et al., 2022; Athota et al., 2023). This process produces what can be described as an algorithmic bias inheritance, where ML models internalize the implicit preferences, value hierarchies and cognitive shortcuts of their human creators (Hurlin et al., 2026; Mehrabi et al., 2021). Research suggests that this process operates as a hybrid cognitive bias system, where human and algorithmic errors reinforce one another across decision cycles (Gigerenzer, 2024). Taken together, these findings suggest that AI does not eliminate bias but instead persists and evolves through continuous interactions between human cognition and algorithmic systems.

Furthermore, scholars, including Grishina et al. (2017), describe algorithmic bias as a form of digital governance. In this sense, algorithmic systems do more than classify or predict outcomes; they also shape how individuals interpret themselves and their choices. Behavioral manifestations include overreliance on algorithmic authority, self-disciplining around credit ratings and reduced trust when decisions appear opaque (Gurdgiev and O'Loughlin, 2020; Matthews et al., 2024). This perspective highlights that AI systems are not merely passive tools, but active participants in shaping user behavior and self-perception, creating feedback loops between human judgment and algorithmic outputs.

This co-dependence exposes institutions in banking and financial services to a duality of algorithmic innovation. While AI augments human cognition through scalability and predictive precision, it simultaneously introduces algorithmic opacity, constraining interpretability and accountability (Giudici and Raffinetti, 2023). For instance, ML credit models broaden access for nontraditional borrowers but embed proxy variables that reproduce discriminatory histories (Fu et al., 2021; Yang et al., 2025). These contradictions underscore how AI's dependence on historical data and probabilistic inference introduces algorithmic opacity and the risk of model drift (Giudici and Raffinetti, 2023; Hurlin et al., 2026). Conceptually, this duality reflects a structural tension between predictive efficiency and interpretability, suggesting that algorithmic opacity is not simply a temporary limitation but rather an inherent feature of advanced AI systems.

In the context of financial services, this duality has critical implications for trust calibration, as users must balance reliance on algorithmic recommendations with uncertainty about their procedural fairness. As Vlačić et al. (2021) and Cao and Zhang (2025) emphasize, this asymmetry challenges the premise of responsible AI, requiring institutions to treat explainability not as a technical add-on but as an organizational principle embedded in governance and auditing frameworks. From a managerial perspective, this implies that institutions must actively manage trust calibration, ensuring that users neither over-rely on nor underutilize algorithmic recommendations in decision-making processes.

Addressing this duality requires an informed approach to mitigating bias and correcting cognitive errors. Traditional bias control methods, such as data balancing, fairness constraints or adversarial debiasing, focus primarily on the algorithmic layer. However, behavioral correction extends beyond model design to include human augmentation (e.g. decision aids, feedback loops and interpretive scaffolds) depending on the task complexity (Grewal et al., 2025). In practice, these interventions help users better align their mental models with algorithmic reasoning, thus mitigating overreliance and bias (Amaral et al., 2023). Furthermore, adaptive systems leveraging meta-learning can detect when user behavior diverges from normative benchmarks, prompting corrective explanations or adaptive nudges (Lu and Calabrese, 2023). These mechanisms create a more dynamic learning environment in which both humans and algorithms evolve toward greater cognitive symmetry and fairness. However, the literature remains fragmented in explaining how technical bias mitigation strategies and behavioral interventions interact, and under which conditions they effectively reduce bias without introducing new distortions.

Therefore, future research for Theme 2 should consider the cognitive, emotional and behavioral dynamics of human–algorithm collaboration. Exploring how bias moves between human and algorithmic agents opens opportunities to study how trust, overreliance and interpretive reasoning evolve in adaptive AI environments. Scholars should also examine how behavioral correction mechanisms enhance human–algorithm interactions and systemic stability. Integrating behavioral economics with ML research could reveal how biases evolve across time and how organizational learning processes prevent their amplification. Alternatively, research should implement algorithmic fidelity (Argyle et al., 2023) by conditioning LM models that simulate the sociocultural characteristics of a specific demographic group to better capture different patterns of attitudes and ideas present across many groups. Furthermore, new experimental methods (e.g. neurophysiological measures of attention and confidence) can enhance the understanding of how users calibrate trust under varying levels of algorithmic opacity and explainability. Finally, as prior research largely relies on quantitative methods such as regression, t-tests and experiments, future studies could adopt understudied methods such as survey-based SEM to examine perceived fairness and the adoption of AI tools in the banking sector or to test how transparency affects trust in AI-based banking.

The third theme centers on the conceptual and regulatory architecture of fairness in AI. There is growing literature on fairness-design approaches that embed ethical constraints into the model development phase (Giudici and Raffinetti, 2023). New quantitative methods, such as cohort-based Shapley values (Lu and Calabrese, 2023), probabilistic percentile-equivalence models (Varley and Belle, 2021) and interpretable deep-learning architectures (Repetto, 2025), reflect the increasing maturity of the field, where fairness is treated as intrinsic to optimization rather than an external audit requirement.

Taken together, these studies show that fairness is increasingly operationalized through formalized and quantifiable approaches embedded within algorithmic systems. However, as Momtaz (2021) and Millo et al. (2024) argue, technical fairness does not automatically translate into ethical fairness. In practice, overly rigid constraints can destabilize systemic-risk models or generate new forms of bias. Fairness must therefore be calibrated to the socio-economic context, and a universal standard is both unrealistic and undesirable.

Conceptually, this reveals a fundamental gap between computational fairness and socio-ethical fairness, pointing to the idea that fairness cannot be fully addressed through technical optimization alone. Legal and regulatory analyses underscore the inadequacy of current regulatory frameworks. Traditional anti-discrimination and consumer-protection laws are ill-equipped to address adaptive algorithms, as they lack mechanisms to handle model opacity or dynamic learning (Gillis and Spiess, 2019; Hurlin et al., 2026). This gap highlights a structural misalignment between rapidly evolving AI systems and slower-moving regulatory frameworks, creating uncertainty in the governance of algorithmic decision-making.

Moreover, studies under Theme 3 connect fairness to organizational responsibility, consumer trust, innovation competency and systemic stability. Responsible AI can therefore be understood not only as a governance model, but also as a strategic resource, integrating behavioral and institutional perspectives (Guenther et al., 2023; Lu and Calabrese, 2023; Repetto, 2025). Importantly, although customers may question and reject technically biased systems, the trust of a bank or financial institution can mediate their perception of fairness (Rubin et al., 2025). From a managerial perspective, this suggests that fairness is not only a compliance requirement but also a strategic capability that shapes organizational legitimacy, customer trust and competitive differentiation.

Further, firms can reframe fairness as an innovation competency rather than a compliance burden through organizational learning and increased bias-awareness capabilities, aligning ethical design with strategic differentiation (Lauretta and Santamaria, 2025). In this context, fairness management thus becomes integral to corporate culture, fostering transparency, accountability and long-term brand trust (Ho and Chow, 2024). This implies that fairness must be embedded within organizational processes and culture, rather than treated as an external regulatory constraint. At the systemic level, fairness and stability are interlinked. Biased algorithms can distort credit markets, misallocate capital and erode institutional confidence (Carbó-Valverde et al., 2025). Consequently, emerging frameworks emphasize macro-level governance mechanisms, combining interpretability, continuous auditing and multi-stakeholder accountability (Yardi, 2024; Packin and Lev-Aretz, 2024).

Despite these advances, the literature remains limited in explaining how fairness frameworks are operationalized across different institutional contexts and how organizations balance trade-offs between fairness, performance and regulatory compliance. In particular, there is still limited empirical insight into how these trade-offs are managed in practice. Therefore, future research on Theme 3 should extend the conceptual and empirical study of fairness from technical to socio-organizational domains. Scholars might investigate how fairness-by-design principles interact with firm culture, regulatory mandates and consumer perceptions.

Comparative analyses across regulatory regimes (e.g. EU, USA and Asia–Pacific) can illuminate how different institutional architectures foster or constrain the adoption of responsible AI. For instance, mixed-method approaches could bridge quantitative fairness metrics with qualitative evaluations of legitimacy and trust. Additionally, as shown in Table 9, no studies in this theme adopt qualitative analysis. This highlights a critical gap, suggesting that future research should prioritize qualitative analyses to better understand how governance policies shape algorithmic fairness across banks. Future research could also conduct event studies to examine market reactions to AI governance regulations.

Theme 4 focuses on behavioral finance, which suggests that financial decision-makers are prone to cognitive and emotional biases that shape their judgments and choices (Thakker et al., 2023). The literature on this theme consistently shows that AI is positioned as both a corrective mechanism that reshapes how these biases manifest in financial decision-making contexts.

The adoption of AI-based tools in forecasting, portfolio optimization and credit evaluation reflects a shared objective of enhancing decision accuracy by addressing bounded rationality (Tubadji et al., 2021), information asymmetries (Momtaz, 2021) and systematic bias (Yang et al., 2025) inherent in both human and algorithmic judgements. Conceptually, this suggests that AI does not eliminate bounded rationality but instead reconfigures it within a human–algorithm decision-making process. Accordingly, studies under this theme conceptualize behavioral finance not merely as the psychology of investors, but also as a domain in which AI can replicate, correct or even amplify human biases in financial contexts (Greenwald et al., 2024; Gurdgiev and O'Loughlin, 2020; Hurlin et al., 2026). This reframing positions AI systems as active participants in behavioral processes, rather than neutral tools, thereby extending traditional behavioral finance theory.

Recent research applies AI to forecasting, portfolio optimization and asset pricing as a means of reducing human subjectivity and cognitive distortions in financial decision-making. ML algorithms have been utilized to refine portfolio selection (Zhang et al., 2022), online portfolio strategies (Zhang et al., 2024) and high-dimensional index tracking (Shi et al., 2023), offering data-driven optimization techniques that outperform traditional heuristics. Eilers et al. (2014) and Bockel-Rickermann et al. (2025) further suggest that reinforcement learning and causal ML approaches simulate adaptive investor behavior and market dynamics, which produce more robust and objective trading outcomes. While these approaches improve predictive performance, they also reveal a tension between performance optimization and bias mitigation, particularly when models rely on historically biased data.

Research also shows that machine-learning models in credit scoring are more accurate overall and perform comparably to, and in some cases better than, traditional models on relevant fairness criteria, minimizing the role of subjective human evaluations in lending decisions (Bono et al., 2021). Empirical studies in microfinance (Chai et al., 2025), peer-to-peer lending (Fu et al., 2021) and SME financing (Lu and Calabrese, 2023) further show that AI enhances the fairness and precision of credit scoring, particularly in emerging markets characterized by limited formal credit histories and high informational asymmetry. These findings show that AI can expand access and improve efficiency, but its outcomes remain contingent on data quality and underlying model assumptions.

Even though AI and semi-supervised ML have been shown to reduce sample bias and improve predictive accuracy in credit scoring (Wu et al., 2025), biases in training and limited representativeness of prior studies still challenge their real-world applicability (Crone and Finlay, 2012). This indicates that improvements in model performance do not necessarily translate into equitable or unbiased outcomes across different populations. A further gap lies in understanding how human and algorithmic learning interact over time, creating feedback loops that may either reinforce or mitigate bias in financial decision-making.

The discussion on Theme 4 identifies several promising avenues for future research. Future comparative studies could assess whether certain ML architectures are inherently more bias-resistant than others. Future research may also investigate how reliance on AI-based forecasting changes the mental workload and decision strategies of financial decision-makers. Furthermore, it would be valuable to see how human behaviors and biases influence AI model training and how these models reshape future human judgments, creating a loop between human and algorithmic learning.

According to the SLR, methods of analysis within this theme primarily focus on regression techniques, t-tests SEM and machine-learning–based predictive models. In contrast, experimental designs and mediation/moderation analyses appear to be largely absent. Future research should incorporate these approaches to examine the causal inference, underlying mechanisms and boundary conditions among the related constructs. Such approaches would be particularly useful for understanding investor reliance on AI-generated recommendations and how AI advice influences risk perception.

This study contributes to the literature on AI and bias in banking and financial services in several ways. First, it develops a more integrated view of AI bias, conceptualizing it as a multi-level and dynamic phenomenon that emerges from the interaction between technological systems, human cognition and institutional contexts. While prior studies tend to examine bias from either a technical or a behavioral lens, this work brings these perspectives together, showing that bias is not confined to a single source but instead unfolds through ongoing interactions between data, algorithms and users.

Second, the study contributes to theory by introducing a human–algorithm co-evolution perspective, highlighting how biases are reproduced, amplified or mitigated through recursive feedback loops between human decision-makers and AI systems. Rather than treating AI as a passive recipient of human bias, this perspective emphasizes that AI systems also reshape those biases over time, creating a more dynamic and reciprocal process than previously assumed.

Third, our article revisits the notion of fairness by distinguishing between computational fairness and socio-ethical fairness and by highlighting the limitations of relying solely on technical solutions for bias mitigation. This distinction points to fairness as a socio-technical construct that depends on the alignment among algorithmic design, organizational practices and regulatory frameworks.

Fourth, the study extends behavioral finance by positioning AI as an active participant in decision-making processes, rather than a neutral analytical tool. This shift helps explain how AI reshapes bounded rationality, influences risk perception and alters decision strategies, offering a more nuanced understanding of financial behavior in algorithmic environments.

Finally, the study brings together a fragmented body of research into four interrelated themes, providing a structured lens through which AI and bias in financial services can be understood. By connecting previously separate streams across technology, behavior and governance, this framework offers a clearer foundation for future theory development.

These findings have significant implications for various stakeholders, including banks, financial institutions, managers, policymakers and academic researchers. Banks and financial institutions need to think of AI not as a set of isolated tools, but as an integrated infrastructure, aligning data systems, decision models and governance processes across the organization. To operationalize this, managers should conduct internal audits of their AI capabilities across these layers and identify misalignments between data systems, decision models and governance practices.

The capacity to leverage AI effectively depends on how well these technological, data and governance components are aligned in practice. Managers should move beyond episodic digital initiatives toward enterprise-wide AI architectures that enable long-term learning, transparency and cross-functional coordination. This may involve establishing dedicated AI governance teams, investing in data integration platforms and developing cross-functional workflows that connect analytics, risk management and customer-facing units. More broadly, this requires aligning AI deployment with strategic objectives, orchestrating resources and developing human capital to support business agility and foster systemic resilience.

The duality of algorithmic innovation requires carefully balancing efficiency with explainability. Because biases can persist and evolve within AI-supported decision-making processes, managers should combine technical bias-mitigation tools with behavioral interventions (Argyle et al., 2023). For example, managers can implement explainability tools (e.g. model interpretability dashboards) and complement them with employee training programs that enhance algorithmic literacy and awareness of cognitive biases. Such initiatives help ensure that employees not only use AI systems but also understand their limitations. Training programs in algorithmic literacy and cognitive bias awareness, along with interpretable decision-support tools, can foster more balanced human–algorithm interactions and strengthen accountability in AI-assisted decision-making. Managers should treat responsible AI governance as a strategic priority by embedding fairness monitoring, transparency practices and accountability mechanisms into decision systems. Banks and financial institutions that embed fairness frameworks and ethical standards within their architecture (e.g. integrating socio-technical audits and continuous monitoring of decision outcomes) can strengthen consumer trust, organizational legitimacy and regulatory alignment. In practice, this requires implementing continuous auditing systems, publishing transparency reports and establishing interdisciplinary committees that oversee AI fairness and accountability.

Ultimately, effective AI transformation depends on the alignment of infrastructure, cognition and ethics. Managers who institutionalize iterative learning between humans and algorithms are better positioned to use bias management not only as a risk-control mechanism, but also as a source of innovation and trust, positioning their institutions to compete responsibly in an AI-driven financial ecosystem. This can be supported by embedding feedback loops into decision systems, regularly updating models based on user behavior and monitoring how algorithmic decisions affect different customer segments over time.

While this research offers some insightful contributions, it still has some limitations. While a comprehensive search protocol was applied to identify all relevant articles within the scope of the current study, utilizing two prominent databases (i.e. Scopus and WoS), the nature of this research imposes a risk of missing relevant articles (Pasquino and Lucarelli, 2025). Second, the focus on A and A* Journals based on the ABDC list published in English limits the inclusion of potentially insightful articles from non-English or lower-tiered journals. Third, the use of author-reported data introduces potential inconsistencies or missing information. Despite these limitations, the current research offers new insights into AI and biases in banking and financial services.

The current study advances the literature on AI and bias in banking and financial services by synthesizing results from a bibliometric analysis and a systematic literature review. Findings from the bibliometric analysis identified four clusters: (1) ethical AI and bias in financial decision-making, (2) ML and bias in financial systems, (3) financial risk in behavioral finance and 4) trustworthiness in corporate governance and audits. Building on this, a deeper evaluation of the contextual use of AI across the 65 reviewed articles revealed four additional emerging themes: (1) AI as a strategic infrastructure for digital transformation in banking and financial services, (2) human–algorithm interaction and the duality of algorithmic opacity toward bias mitigation and behavioral correction, (3) fairness frameworks and responsible AI governance, trust, ethics and systemic stability and (4) behavioral finance in the age of AI and LM. Future research should adopt interdisciplinary and comparative methodologies to address identified gaps and generate new knowledge on this important topic. Overall, responsible AI integration remains important and is shaping up to be a core organizational competency for long-term success in companies.