Human–AI teams for decision-making in banking and finance: A systematic scoping review

Abstract

Purpose

This systematic scoping review synthesizes empirical evidence on human–AI team decision-making in banking and financial services, a domain characterized by rapid AI adoption but limited documentation of how that adoption shapes collaborative decision processes.

Methods

Following the Arksey and O’Malley framework and PRISMA-ScR guidelines, three databases (Scopus, Web of Science, PsycINFO) were searched, along with snowball sampling. Thirteen empirical studies met eligibility criteria.

Results

Three collaborative architectures were identified: human-as-final-decision-maker, human-as-supervisor, and iterative co-construction with AI consistently operating as the initial analytic agent. Trust dynamics ranged from calibrated reliance to algorithm aversion and commitment bias. Explainable AI supported sensemaking, but only when it was designed interactively. Organizational incentive alignment and AI literacy emerged as decisive moderators.

Conclusions

In the included studies, performance gains from human–AI teaming in finance were evident but depended on collaboration architecture, trust calibration, and sociotechnical design. Human-centered outcomes and task performance were partially independent dimensions of effectiveness. Empirical documentation of AI deployment in banking remains thin and geographically concentrated, warranting longitudinal field research and empirically grounded governance standards.

Keywords

human–AI teaming banking financial decision-making trust calibration explainable AI scoping review

Introduction

Artificial intelligence is increasingly embedded across banking and financial services, supporting functions such as credit risk assessment, fraud detection, portfolio optimization, and customer advisory.^1,2 In some institutions, AI has moved beyond peripheral applications toward integration in core decision workflows, though the extent and form of this integration vary considerably across contexts.^3,4 The organizational and human conditions required for effective deployment, however, remain poorly understood relative to the pace of technical adoption,^5,6 and empirical documentation of how human–AI teaming unfolds in financial decision workflows remains scarce, methodologically fragmented, and geographically concentrated.⁷

Financial decisions carry legal accountability, and their consequences extend across clients, institutions, and markets. Governance frameworks increasingly require human oversight of AI systems in high-stakes domains,⁸ yet evidence suggests that the formal presence of a human decision-maker does not automatically produce meaningful engagement, trust miscalibration, algorithm aversion, and commitment bias can each undermine collaborative quality in ways that existing frameworks may not adequately address.^9,10 These considerations suggest that understanding what effective human–AI teaming requires in financial contexts is both a scientific and a governance priority, though the evidence base needed to inform that understanding remains limited.

Existing reviews have addressed human–AI teaming and trust across domains,^11–13 with evidence from healthcare suggesting that collaborative gains depend on task structure and implementation context.¹⁴ This review focuses specifically on how these dimensions interact within banking and financial services, a domain whose accountability demands, regulatory constraints, and professional norms warrant dedicated examination. We therefore prioritize theoretically grounded mechanisms likely to shape human–AI teaming across technological generations, rather than cataloguing the effects of any system.¹⁵ Guided by the Arksey and O'Malley¹⁶ approach, we pursue four objectives: (1) map the financial decision contexts and AI system characteristics present in empirical human–AI teaming research; (2) analyze task allocation and decision authority structures; (3) synthesize task-level performance and human-centered outcomes; and (4) identify recurrent implementation barriers and governance gaps warranting future research and regulatory attention.

Literature review and theoretical background

AI deployment in banking and financial services

Credit scoring and loan evaluation are the most studied domains, with machine learning models using borrower characteristics, transaction histories, and alternative data to achieve performance comparable to traditional scorecards.^3,4 Such models have also identified risk patterns associated with the 2008 crisis, illustrating the potential of algorithmic credit assessment.¹⁷ Fraud detection systems use anomaly detection to identify suspicious transactions at scales beyond manual review,⁴ while portfolio optimization systems apply reinforcement learning and multi-objective algorithms to generate allocation recommendations under varying constraints.¹⁸ Customer service applications range from rule-based chatbots to large language model-based advisory systems.¹

Human involvement remains integral due to regulatory requirements and risks such as bias and fairness in AI-supported decisions.¹⁹ Employee skills, data reliability, and management alignment have been identified as decisive organizational conditions for effective human–AI collaboration in financial services firms.⁶ Generative systems add oversight challenges because narrative outputs require scrutiny of reasoning and provenance,^1,20 raising governance questions that remain underdeveloped in current frameworks.

Evidence on human–AI teaming and research gap

Empirical research on human–AI teaming in banking has developed across three streams: machine learning studies focused on model performance, technology acceptance research examining adoption, and human–computer interaction studies addressing interface design.^21–25 Cross-domain research suggests that collaborative gains depend on task structure, expertise alignment, and interaction design,^12,14 with conceptual frameworks for human–AI teaming emerging across service domains including healthcare¹⁴ and tourism.²⁶ In banking specifically, longitudinal evidence suggests that value creation through human–AI teaming depends on interaction quality and dynamic organizational capabilities rather than algorithmic performance alone.⁵

Trust in AI is shaped not only by accuracy but also by transparency and perceived fairness.¹³ In financial contexts, trust dynamics are sensitive to AI performance disruptions and error exposure, with recovery depending on explanation quality and interaction history.^9,10 Recent work further identifies teamwork frameworks, engagement, transparency, process control, and the substitution or supplementation of knowledge components as relevant to understanding employee-AI co-production and domain expertise in financial service contexts.^27–29

Theoretical foundations of human–AI teaming

We draw on three interconnected theoretical traditions: human factors and automation theory, judgment and decision-making research, and sociotechnical systems theory, each addressing a distinct dimension of human–AI teaming.

Human factors and automation theory helps explain task structure and authority allocation. The levels-of-automation taxonomy^30,31 distinguishes functional stages of decision-making and provides a basis for identifying where miscalibrated reliance may emerge. In financial AI deployments, where systems typically support early-stage analysis while humans retain decision authority, this configuration may create risks related to automation bias, over-reliance, and vigilance.³¹ Automation bias varies with task complexity, time pressure, and prior experience.³⁰ Effective human–AI teaming might require alignment of AI purpose, value commitments, and system properties, increasingly reflected in regulatory expectations for high-risk AI systems.^8,32,33 Early studies³⁴ identified coordination and communication demands in AI-agent integration, suggesting that human–AI teaming challenges are fundamentally relational and organizational rather than purely technical.

Judgment and decision-making research helps explain the cognitive processes through which humans engage with AI outputs. Under time pressure or cognitive load, individuals may accept AI outputs with insufficient scrutiny.³¹ Algorithm aversion, the tendency to discount AI advice after seeing it make errors, is a separate but related problem that often appears alongside other trust failures.³⁵ These patterns can coexist within the same populations,^36,37 indicating that miscalibrated reliance is not unidirectional. Algorithm aversion may happen even when objective performance favors the algorithm,³⁵ and domain expertise can moderate these effects, although findings remain context dependent.³⁸

The literature on trust provides a conceptual bridge between task structure and organizational context. Across disciplines, AI reliance depends on actual performance, perceived transparency, contextual risk, social norms, and the broader organizational environment.^39,40 This carries governance implications: effective human oversight cannot be reduced to the formal presence of a human decision-maker, as evidence suggests it depends on the cognitive and organizational conditions that enable calibrated trust.³⁸ Evidence across human factors, HCI, and interpretability research^41–43 suggests that explanation quality shapes how users integrate AI outputs with domain knowledge, though its relationship with decision quality is not consistently positive. Surface-level transparency can increase confidence without improving accuracy,⁴² and the effectiveness of explanation design depends on format, user expertise, and task structure.^23,25,44 Rigorous evaluation of interpretability therefore requires task-specific measures.⁴² Recent evidence from financial contexts suggests that end-users and technical stakeholders diverge substantially in their explanation format preferences 45.

Sociotechnical systems theory addresses the organizational and institutional conditions shaping collaborative outcomes, a dimension particularly relevant to banking contexts.⁴⁶ It posits that technology adoption outcomes are co-determined by technical system properties and surrounding social and organizational structures.^12,46 Organizational incentives and HR practices influence⁴⁷ whether AI capabilities translate into performance gains.^2,48,49 Evidence further suggests that AI-induced technostress may reduce employee engagement and well-being,⁵⁰ while AI literacy can mitigate these effects, although evidence in financial professional populations remains limited.⁵¹ At the organizational level, investment in digital skills is associated with greater returns from AI adoption.^52,53 At the same time, shifts in task ownership may contribute to professional identity erosion, and algorithmic features such as opacity and immediacy may reduce perceived autonomy and fairness, potentially undermining conditions for calibrated reliance.⁵⁴ This perspective informs the review’s empirical synthesis by framing human–AI teaming as contingent on organizational as well as technical factors. Table 1 summarizes the literature review and theoretical background.

Table 1.

Summary of literature: AI deployment, human–AI teaming evidence, and theoretical foundations.

Dimension	Focus	Key themes	Core insights	Key references
AI Deployment in Banking	What AI systems do in financial decision workflows	Credit scoring, fraud detection, portfolio optimization, customer service, generative AI	(1) AI deployment in banking spans credit scoring, fraud detection, portfolio optimization, customer service, and increasingly generative advisory applications, though the extent and form of integration varies considerably across institutional contexts. (2) Explainability and oversight requirements differ by application type—predictive systems raise different governance challenges than generative systems whose narrative outputs require scrutiny of both reasoning and provenance. (3) The organizational and human conditions required for effective AI deployment remain underexamined relative to the technical capabilities being adopted.	^{1,3–6,17–20}
Human–AI Teaming Evidence	How human–AI collaboration has been studied and what gaps remain	ML performance, technology acceptance, HCI design, emerging organizational HAT research	(1) Finance-specific empirical evidence on human–AI teaming remains scarce, methodologically fragmented, and geographically concentrated, limiting transferability across regulatory and institutional contexts. (2) Cross-domain research suggests collaborative performance gains are not guaranteed and depend on task structure, expertise alignment, and interaction design, conditions that may not follow from technical system properties alone. (3) Trust miscalibration appears bidirectional, both algorithm aversion and over-reliance are documented within financial contexts, suggesting that the challenge is calibration rather than simply increasing trust.	^{9,10,12–14,21–26,28,29}
Theoretical Foundations	What frameworks explain HAT mechanisms in financial contexts	Automation bias, algorithm aversion, trust calibration, explainability effects, sociotechnical conditions	(1) Miscalibrated reliance appears to be shaped simultaneously by cognitive factors (automation bias, algorithm aversion), design factors (explainability format and interactivity), and organizational factors (incentive alignment, AI literacy), no single factor appears sufficient based on available evidence. (2) Trustworthy systems do not automatically produce calibrated reliance; the gap between system reliability and user trust depends on cognitive and organizational conditions that formal human oversight may not substitute for. (3) Individual-level interventions such as explanation design may be insufficient without supporting organizational and governance conditions, pointing to a sociotechnical rather than purely technical understanding of effective human–AI teaming.	^{8,30–33,35–44,50–54}

Method

A systematic scoping review was conducted following the five-stage framework proposed by Arksey and O'Malley,¹⁶ appropriate for heterogeneous and emerging bodies of literature where studies vary widely in design, context, and reported outcomes. To enhance transparency and reproducibility, the review was reported in accordance with PRISMA-ScR guidelines,⁵⁵ which informed the reporting of study identification, screening, eligibility, and data charting. The review protocol is documented in an Open Science Framework (OSF) registration. The registration is currently embargoed for peer review and will be made publicly accessible upon publication. Database searches were conducted in December 2025; background references were updated during revision and did not affect corpus composition.

Step 1: Identification of the research question

The study focused on situations in which humans and AI systems jointly contribute to decision-making, rather than fully automated systems. Extracted dimensions included AI system type and purpose, task allocation, and outcomes covering trust, reliance, sensemaking, perceived control, and autonomy. By systematically mapping these dimensions, we discuss empirical patterns and gaps in the literature, regarding decision contexts, methodological limitations, and emerging AI technologies.

Step 2: Identification of relevant studies

Searches were conducted in December 2025 across three databases (Scopus, Web of Science, and PsycInfo). Snowballing techniques complemented database searches following established guidelines,⁵⁶ through backward citation searching and forward citation tracking applied to the full set of included studies as the seed corpus following completion of database screening. One round of snowballing was conducted, with candidate references screened against the same eligibility criteria applied to database records. The low yield of snowball sampling (n = 2) may reflect the comprehensiveness of the primary database search, although it also indicates that snowballing contributed only minimally to the final corpus. Full search strings for all databases, database-specific adaptations, and procedural documentation are reported in the OSF-registered protocol, which is currently embargoed for peer review and will be made publicly accessible upon publication. Boolean operators and wildcard symbols were applied to refine retrieval. The search syntax for Scopus was as follows: (“AI” OR “Artificial intelligence” OR “decision support system*” OR “DSS*” OR “human-AI team*” OR “human AI team*” OR “human-agent team*” OR “hybrid team*” OR “AI team*” OR “AI collab*” OR “human-AI agent collab*”) AND (“bank*” OR “banking” OR “finance*” OR “financial”) AND (“decision*” OR “decision-making” OR “decision support”).

Equivalent search strategies were adapted for the other databases, considering their specific indexing conventions and search fields, while maintaining the same conceptual scope. The initial search yielded 1016 records (Scopus: n = 372; Web of Science: n = 476; PsycInfo: n = 166; Snowball: n = 2). After removing duplicates (n = 428), 588 articles remained for screening.

Step 3: Study selection

In the first stage, 332 records were excluded for absence of a banking or financial domain, human–AI teaming, or decision-making task, leaving 256 articles. In the second stage, 208 records were excluded for lacking AI as a decision-support system or for being non-empirical, yielding 48 full-text articles assessed for eligibility. Eligible studies were empirical, published in peer-reviewed journals or conference proceedings in English since 2020, and situated within a banking or financial decision-making context involving explicit human–AI teaming. Conceptual papers, literature reviews, industry reports, and studies of fully automated systems were excluded. Following full-text review, 35 articles were excluded, leaving 13 studies in the final synthesis. One study meeting all other criteria was published in late 2019 and was retained given its direct relevance to the review’s decision-making focus. Disagreements were resolved through discussion. The study selection process is summarized in the PRISMA flow diagram (Figure 1).

Figure 1.

PRISMA flowchart summarizing the study selection process.

Step 4: Quality assessment of studies

Eligible studies were appraised using the Joanna Briggs Institute (JBI) checklist for quasi-experimental studies as a supplementary step to contextualize findings, without excluding studies on quality grounds. Results are summarized in Table 2.

Table 2.

Quality assessment and risk of bias through the JBI checklist.

Citation	Variables causality	Participant similarity	Comparable intervention	Control group presence	Pre-post measurement	Follow-up/differences analysis	Measurement consistency	Outcome measurement reliability	Statistics appropriateness
Dikmen and Burns (2022)	✓	?	✓	✓	✖	?	✓	✓	✓
Ding et al. (2023)	✓	?	✓	?	✓	✓	✓	✓	✓
How et al. (2020)	✓	✓	✓	?	?	✓	✓	✓	✓
Judek (2024)	✓	✓	✓	✓	✖	✓	✓	✓	✓
Khalil et al. (2025)	✓	?	✓	✓	✓	?	✓	✓	✓
Liu et al. (2024)	✓	✓	✓	✓	✖	?	✓	✓	✓
Luong et al. (2019)	✓	?	✓	✓	✓	✓	✓	✓	✓
Mandava et al. (2025)	✓	?	?	✓	✓	?	✓	✓	✓
Maier et al. (2022)	✓	?	✓	✖	✓	?	✓	✓	✓
Oliveira Filho et al. (2023)	✓	✓	✓	✖	?	✓	✓	✓	✓
Sachan et al. (2024)	✓	?	✓	✖	✓	✓	✓	✓	✓
Strunk et al. (2025)	✓	✓	✓	?	?	?	✓	✓	✓
Yang et al. (2025)	✓	✓	?	✓	✓	?	✓	✓	✓

Note. ✓ = low risk; ? = unclear; ✖ = high risk.

Step 5: Charting the data and reporting the results

Data were extracted using a predefined, iteratively refined data charting framework, designed to create consistency across studies while accommodating methodological diversity to reveal patterns among organizations or contexts. Extracted information included author and year, country, AI system model and purpose, AI interface, human–AI team composition, working task, AI tasks, human tasks, performance outcomes, and human-centered outcomes.

Results

Risk of bias and quality of studies

Of the 13 studies, 7 (53.8%) received no high-risk ratings across evaluated domains,^57–63 while 6 (46.2%) presented at least one high-risk rating on a specific criterion. No study was classified as high risk across all criteria. All studies presented a sufficiently clear causal or comparative structure for the purposes of the JBI appraisal and received low-risk ratings for measurement consistency, outcome reliability, and statistical appropriateness. Three studies did not employ pre-post measurements,^64–66 and three lacked a defined control group.^67–69 Follow-up and difference analysis received unclear ratings in seven studies, reflecting a common design feature of this literature rather than a systematic quality concern. Where inferential statistics were reported, effect sizes and p-values are included in the synthesis.

Decision contexts and AI system characteristics

The 13 studies covered diverse financial decision-making contexts. Credit, lending, and trade finance decisions were the most frequent domain, examined in five studies.^{59,60,63,64,68} These studies addressed loan application screening, peer-to-peer lending investment selection, trade finance document verification, and specialist loan underwriting. AI systems typically generated predictive risk assessments, discrepancy flags, or structured compliance outputs to support human evaluation and final decisions.

A second cluster focused on investment-related decision-making, including stock price forecasting, portfolio allocation optimization, and investment suitability assessment.^65,69 Here, AI systems provided probabilistic predictions, portfolio recommendations, or optimization outputs for interpretation by investors or financial advisors. A third group addressed fraud detection and broader risk assessment, where AI systems performed anomaly detection, generated fraud risk indicators, or produced explainable risk classifications to support analysts working under uncertainty.^57,61

The remaining studies examined customer-facing financial services, including AI-supported retail banking interactions⁶⁶ and predictive targeting for financial inclusion marketing.⁵⁸ One study examined financial decision-making under risk more generally, without anchoring the task in a specific operational banking process.⁶²

Across contexts, AI systems functioned as decision-support tools, with humans retaining evaluative, supervisory, or final decision roles. Their main functions included predicting outcome probabilities, generating recommendations, and detecting anomalies or compliance discrepancies in financial data.

A key source of variation concerned explainability. Most studies used transparency or interpretability mechanisms, including feature-based and knowledge-based explanations,⁶⁴ SHAP, LIME, and counterfactual explanations,⁶¹ local anomaly explanations,⁵⁷ Bayesian network visualizations,⁵⁸ and evidential reasoning frameworks.⁶⁸ In contrast, some studies implemented systems with limited or no explicit transparency regarding internal decision logic.^62,65,67 These systems typically presented numerical forecasts, risk indicators, or recommendation signals without detailed explanation of feature contributions or reasoning pathways. This distinction appeared, in several studies, to be associated with patterns of trust, reliance, and advice-taking.

Interfaces included visual analytics dashboards,⁵⁷ Bayesian visualization tools,⁵⁸ prototype explainable AI systems,^61,68 conversational chatbots,⁶⁶ portfolio optimization interfaces,⁶⁹ and numerical forecasting displays.⁶⁵ One study did not report interface details.⁵⁹ This diversity suggests that standardized interaction patterns have not yet emerged in financial decision-making. Figure 2 summarizes these distributions.

Figure 2.

Decision-making domains (a). AI system’s primary purpose (b). Decision authority configuration (c). AI system explainability (d).

Task allocation and decision authority distribution

Across the studies, AI systems consistently acted as the initial analytic component in the decision flow. They aggregated data, detected patterns, modeled predictions, identified anomalies, and generated recommendations or risk scores. Human actors then interpreted these outputs, evaluated contextual information, and made or supervised final decisions. This sequential configuration, AI generation followed by human adjudication, was the dominant model.

Decision authority varied across three configurations. The most common was the human-as-final-decision-maker model, observed in seven studies.^59–65 In this arrangement, AI systems provided predictions, recommendations, or risk scores, while humans retained authority to accept, modify, or reject the output. AI shaped the informational context but did not override human judgment. Human tasks involved evaluating AI-generated information against contextual knowledge, professional standards, or strategic objectives.

A second configuration, human-as-supervisor, appeared in two studies.^66,67 Here, AI systems played a more operational role, autonomously generating customer responses or managing investment portfolio segments. Human actors monitored performance, determined delegation levels, and intervened selectively. Authority was partially shared, with human influence exercised through oversight and escalation rather than evaluation of each discrete output.

A third configuration, iterative co-construction, characterized studies in which human and AI contributions were integrated through repeated interaction rather than a single handoff.^57,64,68 Analysts explored AI-generated outputs, examined explanations, adjusted parameters or preference weights, and refined decision criteria through interaction with the system. Authority was distributed across multiple stages, with both human and AI contributions shaping the final decision.

Across configurations, human actors appeared to engage in sensemaking and evaluative reasoning rather than merely confirming AI outputs, although this inference requires validation in broader deployment settings. Table 3 summarizes the cognitive dynamics, trust patterns, and quantitative findings associated with each teaming architecture. Given heterogeneity in decision tasks, participant populations, outcome measures, and reported statistics, these findings are compared descriptively by collaboration architecture rather than pooled statistically.

Table 3.

Collaboration architecture, cognitive dynamics, and outcomes in human–AI financial decision-making.

	Human-as-final-decision-maker	Human-as-supervisor	Iterative co-construction
Studies	^{58,59,61–65}	^66,67	^57,68,69
Decision contexts	Loan screening, trade finance, peer-to-peer lending, investment advisory, risk scenarios	Retail banking customer service, robo-advisor portfolio management	Fraud detection, loan underwriting, private banking portfolio optimization
Human task	Evaluate AI output; apply domain knowledge; accept or override recommendation	Monitor performance; determine delegation degree; intervene selectively	Explore outputs iteratively; adjust parameters; refine criteria across cycles
AI task	Generate predictions, risk scores, or recommendations as informational input	Autonomously execute operational tasks	Generate initial outputs; respond to human adjustments; support progressive refinement
Trust pattern	Calibrated when explainability is interactive and domain expertise sufficient^61,64; aversion under high risk⁶² or low-adoption social signals⁶⁵; over-reliance when AI advice is human-endorsed⁶³	Commitment bias independent of performance⁶⁷; fewer trust recalibrations with increasing interaction turns⁶⁷; perceived service fairness without AI scrutiny⁶⁶	Predominantly calibrated; progressive trust development through iterative engagement; discrepancies between AI outputs and expert judgment resolved across cycles^68,69
Sensemaking quality	Variable and expertise-dependent sensemaking; risk-driven disengagement⁶²; aversion overriding accuracy signals⁶⁵; counterfactual exploration via visualization⁵⁸	Episodic sensemaking; engagement limited to monitoring and escalation points; scrutiny progressively displaced by inertia across interaction turns⁶⁷	More sustained sensemaking; iterative interaction enabling exploration and progressive reconciliation; improvement across cycles^68,69
Task performance	Improved decision accuracy observed in controlled settings (83% vs 65% black-box⁶¹); reduced Type I and II errors when AI accuracy and incentive alignment coincided⁶⁰; 44.92% higher client payoff under Human–AI vs AI-only (β = .371, p < .01) in the studied context⁶³; 68.3% operational risk reduction in trade finance examination⁵⁹; advice-taking significantly lower under high financial risk (MD = −40.71, p < .001⁶²); algorithm outperformed human forecasts (d = 1.80) but 41.34% of investors disregarded superior AI output⁶⁵	No consistent task performance advantage documented across either study; trust changes decreased with increasing interaction turns (R² adj = .401, p < .001⁶⁷), consistent with progressive trust inertia; AI-human service evaluation parity achieved under high justice conditions (p = .263⁶⁶)	AUC improved from 0.86 to 0.96 across four iterative human–AI refinement cycles⁶⁸; F1 improved from 0.80 to 0.90⁶⁸; time-to-rule reduced from approximately 8 to 6 min in fraud detection context⁵⁷; portfolio alignment enhanced though no inferential statistics reported⁶⁹
Human-centered outcomes	Empowerment and perceived control⁶¹; perceived control over trade-off decisions⁵⁹; heightened risk sensitivity in financial decision contexts⁶²	Perceived fairness in AI-supported service outcomes⁶⁶; progressive trust inertia constraining meaningful engagement with performance feedback⁶⁷	Increased confidence in decision rationale⁶⁸; enhanced client involvement and comprehension⁶⁹; increased empowerment through iterative engagement⁶⁸

Cognitive and relational dynamics in human–AI joint decision-making

Three interrelated mechanisms shaped human interaction with AI systems across the reviewed studies: trust and reliance calibration, sensemaking, and the moderating role of risk perception and organizational context.

Trust calibration and reliance

Trust in AI was the most frequently examined human-centered construct, with studies documenting calibrated, under-reliant, and over-reliant patterns.

Several studies reported calibrated trust dynamics, where users differentiated between credible AI outputs and situations requiring skepticism. Dikmen and Burns⁶⁴ found that explainable AI enabled retail investors to rely selectively on AI predictions when explanations aligned with domain cues, while maintaining skepticism when discrepancies arose. Participants in the explainability condition made fewer risky investments (M = 4.1 vs 5.7, Mann-Whitney U = 271.5, p = .05) and fewer investments when the AI was incorrect (M = 2.9 vs 3.8, U = 283.5, p = .02). They also reported lower trust scores (M = 4.53 vs 5.11, t (38) = 2.1, p = .04), consistent with calibrated, skepticism-informed reliance rather than blind acceptance. Similarly, Mandava et al.⁶¹ found that SHAP, LIME, and counterfactual explanations increased perceived transparency and confidence among participants with finance backgrounds, suggesting a possible pathway toward comprehension-based trust.

Algorithm aversion and under-reliance appeared in other contexts. Judek⁶⁵ found that 41.34% of investors disregarded accurate AI forecasts despite superior algorithmic performance. AI forecasts showed lower error than human forecasts (8.56% vs 18.51%, t (252) = 14.19, p < .001, d = 1.80), and AI use yielded approximately 63% higher financial payoff (t (252) = 17.47, p < .001). Adoption was influenced by social adoption cues (χ² (n = 254) = 4.69, p = .030) and technology affinity, although participants with lower ATI scores were more likely to use the algorithm (χ² (n = 127) = 8.52, p = .004). Strunk et al.⁶² further showed that trust shaped advice-taking under varying financial risk conditions. These findings suggest that trust may depend not only on system performance but also on perceived risk and individual differences.

A third pattern involved over-reliance or inertia. Maier et al.⁶⁷ found commitment bias among robo-advisor users: participants were less likely than random to decrease trust after having increased it and often maintained their allocation regardless of previous performance. Turn number predicted fewer trust changes over time (adjusted R² = .401, p < .001), indicating increasing reliance inertia. Yang et al.⁶³ found that embedding AI-generated advice within a human advisory context increased client reliance, suggesting that social framing may influence adoption independently of technical accuracy in advisory settings.

Sensemaking and comprehension

Sensemaking appeared across several studies as a potential mediator between AI system design and decision quality, although mechanisms and effect sizes varied. In fraud detection, Ding et al.⁵⁷ reported that interactive fraud rule authoring reduced analyst time-to-rule to approximately 6 min, compared with around 8 min using ad-hoc tools on the Czech dataset and 10–30 min for domain rules on the Card dataset. Rule quality was broadly comparable across conditions, although coverage advantages varied by dataset. Sachan et al.⁶⁸ found that an evidential reasoning framework helped underwriters identify and resolve inconsistencies between data-driven outputs and expert judgments, improving underwriting consistency across four iterations (AUC: 0.86 to 0.96; F1: 0.80 to 0.90).

How et al.⁵⁸ showed that Bayesian network visualizations enabled analysts to explore counterfactual scenarios and variables influencing customer purchase intentions, with good model performance (ROC = 90.41%, R² = .33). De Oliveira Filho et al.⁶⁹ reported that interactive portfolio optimization tools enhanced clients’ understanding of risk-return trade-offs. Collectively, these findings suggest that explainability alone may be insufficient; effective sensemaking appeared to depend on interaction designs enabling exploration, comparison, and iterative refinement of AI outputs.

The moderating role of risk perception and organizational context

Contextual factors, especially perceived risk and organizational alignment, moderated trust and reliance. Strunk et al.⁶² found that advice-taking was significantly lower in high-severity financial risk scenarios than in physical risk contexts (MD = −40.71, p < .001), with trust mediating the effect of risk severity on advice-taking (indirect effect = 17.19, p < .001). Luong et al.⁶⁰ showed that reliance on AI recommendations depended on alignment with organizational incentive structures; when AI outputs were congruent with institutional objectives, decision errors decreased.

Yang et al.⁶³ found that human delivery of AI-generated investment advice increased client trust and material welfare outcomes in the studied context, with 44.9% higher payoff under Human–AI than AI-only advice (β = .371, p < .01). This points to the relevance of sociotechnical integration in shaping human–AI teaming effectiveness.

Decision outcomes and implementation challenges

Most studies reported positive effects of human–AI teaming on task performance, although mainly in controlled or single-context deployments. Improvements included reduced decision errors in loan evaluation,⁶⁰ improved underwriting consistency,⁶⁸ reduced operational risk in trade finance document examination,⁵⁹ improved mean portfolio returns relative to analyst-selected portfolios despite higher volatility and a DSS Sharpe ratio of 0.13,⁶⁹ increased forecast accuracy and payoff when AI outputs were adopted,⁶⁵ enhanced efficiency in fraud rule design,⁵⁷ and improved decision accuracy with explainable outputs compared with black-box systems (83% vs 65%).⁶¹

However, performance gains were not uniform. Maier et al.⁶⁷ found that neither AI nor human performance alone strongly predicted trust change (adjusted R² = .054 and adjusted R² = .023, respectively), whereas the difference between human and AI performance was a moderate predictor (adjusted R² = .136, p < .001). Judek⁶⁵ also observed reduced performance among participants who disregarded AI outputs.

Human-centered outcomes included increased perceived transparency and empowerment,⁶¹ enhanced client involvement and comprehension,⁶⁹ greater confidence in decision rationale and improved alignment between expert judgment and data-driven evidence,⁶⁸ comparable perceived fairness between AI and human agents under high justice conditions (p = .263),⁶⁶ and flexibility for managers to adjust risk-utilization trade-offs in task allocation.⁶⁰

Importantly, task performance and human-centered outcomes did not always co-occur. In some contexts, reliance persisted without objective performance gains; in others, reduced trust limited advice adoption despite potential accuracy benefits. This pattern suggests that task performance and human-centered outcomes are partially independent dimensions of human–AI teaming effectiveness.

Implementation challenges included tailoring explanations to user expertise,^57,61,64 aligning AI outputs with organizational workflows,^60,68 managing user expectations,^63,66 and mitigating cognitive biases affecting reliance.^63,65,67 These challenges suggest that effective human–AI teaming depends on the interaction between system performance, user cognition, workflow alignment, and organizational context. The extracted data from each study is reported in Table 4.

Table 4.

Extracted data from the included studies.

Author(s)	Country	AI system purpose	AI system interface	Team composition	Working Task(s)	AI tasks	Human tasks	Task performance outcome(s)	Human-centered outcome(s)
Dikmen and Burns (2022)	CAN	Decision support for peer-to-peer lending investment via credit risk assessment	Web-based platform with explainable ML outputs	Retail investors + AI system	Loan evaluation and investment selection in peer-to-peer lending	• Prediction of loan default probability • Generation of feature-based and knowledge-based explanations	• Assessment of loan information • Interpretation of AI explanations • Investment selection and allocation • Final investment decisions	• ↓ Risky investments (Exp: M = 4.1 vs. Baseline: M = 5.7; p = .05)	• Calibrated trust (Exp: M = 4.53 vs. Baseline: M = 5.11; p = .04. • Trust correlated with risky investments when AI erred (r = .53, p = .02) • No difference between conditions in explanation quality (p > .05); overall understanding rating M = 6.0/7
Ding et al. (2023)	USA	Decision support for fraud detection through anomaly analysis and rule design	Interactive visual analytics GUI (ALARM toolkit)	Fraud analysts + AI system	Fraud anomaly verification and rule-based fraud detection design	• Detection of anomalous transactions • Generation of local anomaly explanations • Suggestion of candidate detection rules	• Assessment of fraud cases • Interpretation of AI outputs • Verification of true vs. false positives • Design and refinement of fraud detection rules	• ↑ Rule design efficiency and ↓ time-to-rule vs. ad-hoc (banking: ∼6 vs. ∼8 min; card fraud: ∼6 vs. ∼10–30 min	• ↑ Sensemaking support • High perceived usability • Mixed trust in AI-generated rules
How et al. (2020)	SGP	Decision support for financial inclusion marketing via prediction of customer purchase intention	Interactive Bayesian Network visualization for probabilistic scenario simulation	Financial service provider analysts + AI system	Customer targeting and outreach strategy design	• Prediction of purchase intentions • Generation of counterfactual scenario simulations • Identification of influential variables through sensitivity analysis	• Assessment of customer profiles and marketing conditions • Interpretation of AI outputs • Exploration of counterfactual scenarios • Formulation of targeting strategies	• ↑ Customer segment identification (N = 45,211; baseline purchase intention = 11.61% “yes”; ROC = 90.41%, Gini = 71.37%)	• ↑ Understanding of causal relationships (R² = .33; RMSE = .26; strongest predictors: duration, previous outcome, month) • High accessibility for non-technical users (no-coding BN approach)
Judek (2024)	DEU	Decision support for financial forecasting through algorithmic stock price prediction	Numerical forecasting interface displaying stock price predictions	Investors + AI system	Stock price forecasting to support investment decision-making	• Prediction of future stock prices based on historical data	• Assessment of historical stock price information • Interpretation of AI outputs • Final investment decisions	• ↑ Forecast accuracy (8.56% vs. 18.51, p < .001) • ↑ Financial payoff when AI used (≈63% higher, p < .001) • ↓ Performance when AI disregarded (up to 88% higher forecast error; 41.34% rejected AI despite superior accuracy)	• ↑ Actual use of AI influenced by social adoption cues (χ² (1, n = 254) = 4.69, p = .030) • AI adoption correlated with technology affinity/ATI (interaction χ² (1, n = 127) = 8.52, p = .004)
Khalil et al. (2025)	QAT	Decision support for trade finance document examination	Not described	Trade finance officers + AI system	Document examination and compliance checking in letters of credit processing	• Extraction of document information • Detection of discrepancies against trade finance rules • Risk scoring • Optimization-based assignment of documents to human checkers	• Verification of trade finance documents • Interpretation of AI-generated discrepancy and risk assessments • Final decision on compliance and exception handling	• ↓ Operational risk (68.3% reduction under optimized allocation) • ↑ Efficiency (checker utilization: 73% efficiency-focused vs. 34% risk-focused)	• High perceived control over task allocation (managers adjust λ to prioritize risk vs. utilization) • High transparency of workload and risk distribution
Liu et al. (2025)	CHN	Decision support for banking customer service through service agents	Conversational chatbot interface (simulated)	Bank customer service employees + AI system	Handling customer inquiries and complaints in retail banking	• Generation of customer responses • Generation of informational explanations related to service outcomes	• Oversight of customer service interactions • Resolution of complex or sensitive issues	• ↑ Customer satisfaction with service outcomes (interpersonal justice: g = 1.16, p < .001) • No difference between AI and human agents when appropriate explanations provided (p = .263)	• High perceived fairness of AI-supported service; no difference in acceptance between AI and human agents (informational justice: η²p = .253, p < .001) • ↓ Negative reactions to unfavorable outcomes (human penalized more under low justice: d = 0.48, p = .011)
Luong et al. (2019)	USA	Decision support for organizational loan application evaluation through ML-based risk prediction	Interface displaying AI predicted loan risk recommendations	Loan officers + AI system	Loan application evaluation and approval decisions	• Prediction of loan default risk based on historical loan data • Generation of approve/reject recommendations	• Assessment of loan application • Interpretation of AI outputs • Final decision on loan approval	• ↑ Decision quality when AI accuracy and incentive structures aligned (all four hypotheses supported; specific p-values not reported.	• ↑ Intention to use AI (qualitative result) • ↑ Reliance on AI when accuracy and incentives aligned (interaction effect supported; p-values not reported)
Mandava et al. (2025)	USA	Decision support to improve quality and accuracy in financial decision-making	Prototype financial DSS with embedded XAI explanations (SHAP, LIME, counterfactuals)	Finance professionals (investment advisors, loan officers, financial analysts) + AI system	Evaluation of loan approval, investment and fraud detection scenarios	• Generation of decision recommendations • Prediction of risk and outcome likelihoods • Generation of explainable outputs	• Evaluation of AI outputs • Interpretation of AI outputs • Final decision on task outcomes	• ↑ Decision accuracy with XAI vs. black-box (83% vs. 65%; +18 pp) • ↑ Performance across tasks: Loan approval (85% vs. 68%), investment (82% vs. 66%), fraud detection (88% vs. 70%).	• ↑ Perceived transparency (trust score: +32% with XAI vs. +8% black-box) • ↑ Perceived empowerment and confidence (empowerment index: 78% XAI vs. 50% black-box) • ↑ User satisfaction (87% XAI vs. 60% black-box)
Maier et al. (2022)	USA	Decision support for investment decision-making through an AI-based robo-advisor	Interactive simulation interface presenting AI-supported investment performance	Investors + AI system	Investment allocation decisions in simulated financial markets	• Generation of investment allocations • Re-investment of user-delegated funds • Presentation of performance feedback	• Allocation of funds between self-managed stocks and AI-managed portfolio • Monitoring and evaluation of AI performance • Changes in allocation over time	• No consistent ↑ in objective investment performance (AI performance alone weak predictor of trust change: R² adj = .054) • Human–AI performance difference predicts reliance changes (R² adj = .136, p < .001)	• ↑ Reliance on AI when AI outperforms human (β = 1.15 for performance difference) • Emergence of commitment bias and inertia over time (turn number predicts fewer trust changes: R² adj = .401)
Oliveira Filho et al. (2023)	BRA	Decision support for private banking portfolio selection, integrating suitability assessment and capital allocation	Interactive interface (OptPrivate) with questionnaire-based suitability assessment and portfolio optimization outputs	Private banking clients + financial analysts + AI system	Portfolio selection and investment allocation in private banking, considering investor profile, regulatory requirements, risk, and return	• Assessment of investment suitability based on client profile and regulatory constraints • Classification of investment alternatives • Optimization of portfolio allocation • Generation of recommended portfolios	• Definition of investment objectives, preferences, constraints • Completion of suitability questionnaire • Adjustment of preference weights • Review and validation of AI-supported recommendations	• ↑ Mean return in 60% of cases vs. real portfolios (mean difference: +0.28%; p = .13); 100% regulatory suitability maintained • ↑ Portfolio profitability in majority of cases; ↑ volatility (DSS portfolios showed higher standard deviation in 73% of cases; no significant differences reported). • ↓ Unsuitable investment allocation (DSS maintained regulatory suitability in 100% of cases)	• ↑ Client involvement in decision-making • ↑ Transparency and comprehension of trade-offs (interactive weight adjustment and real-time sensitivity analysis) • ↓ Influence of conflicts of interest (DSS portfolios independent of analyst discretion)
Sachan et al. (2024)	GBR	Decision support to improve consistency in loan underwriting	Interactive decision-support interface using evidential reasoning with explainable outputs	Loan underwriters + domain experts + AI system	Loan underwriting and credit decision-making in specialist lending	• Aggregation of data sources • Transformation of raw data into evidence • Aggregation of evidence using evidential reasoning framework • Generation of confidence indicators • Identification of inconsistencies between data-driven outputs and expert judgments	• Application of expert judgment and domain knowledge • Definition and validation of underwriting rules • Evaluation of AI recommendations • Resolution of ambiguous or borderline cases • Final decision on loan approval	• ↑ Consistency of underwriting decisions (AUC: 0.86 → 0.96; F1: 0.80 → 0.90 after 4 iterations) • ↑ Accuracy vs. baselines (ER-X F1 = .901 vs. DNN = .886, Decision Tree = .864, Logistic Regression = .843)	• ↑ Sensemaking under uncertainty (AUC: 0.86 → 0.96 through auditing ambiguous cases) • ↑ Confidence in decision rationale • ↑ Alignment between expert judgment and data (performance gains after 4 feedback iterations)
Strunk et al. (2025)	DEU	Decision support through AI-generated advice under varying risk contexts	Advisory interface displaying AI recommendations with risk indicators	Individual decision-makers + AI system	Financial decision-making based on AI advice under risk	• Generation of context-specific decision advice • Presentation of risk facet (physical, financial, psychosocial) and risk severity (low vs. high) • Prediction of potential consequences associated with advice adoption	• Evaluation of AI-generated advice • Evaluation of risk severity • Final Decision to adopt or reject AI advice	• No consistent ↑ in decision quality (trust predicts advice taking: OR = 4.39; classification accuracy = 79.0%) • ↓ Advice taking in high-risk financial scenarios (financial vs. physical: MD = −40.71, p < .001)	• Trust calibration based on perceived risk severity (F (1,830) = 127.16, p < .001; indirect β = 17.19, p < .001) • ↓ Reliance on AI in high-risk conditions (interaction effect: F (2,830) = 4.19, p = .015)
Yang et al. (2025)	DEU	Decision support for financial investment advice through AI-based risk assessment	Advisory interface providing AI-generated risk assessments and investment recommendations (field deployment)	Bank advisors + customers + AI system	Investment decision-making in personal lending under risk	• Prediction of loan default risk • Generation of investment recommendations • Categorization of risk levels	• Evaluation and adjustment of AI-supported advice • Final decision on investment recommendations • Communication of advice to customers	• No degradation of advice quality with human involvement • ↑ Alignment between advice and customer decisions (Human–AI: +15.5 pp vs. AI-only; β = .155, p < .01) • ↑ Material welfare for customers (payoff +44.9% overall, β = .371, p < .01; for risky investments: β = .791, p < .01)	• ↑ Reliance on AI advice (Human–AI: 79.2% vs. AI-only: 68.3%, p < .05) • ↑ Emotional trust mediates alignment under decision uncertainty (indirect effect = .040, 95% CI [.005, .076], p < .05; not significant under certainty) • ↑ Advice alignment for risky investments under Human–AI vs. AI-only (β = .213, p < .001), consistent with reduced algorithm aversion under uncertainty

Note. Outcome reporting reflects the level of detail provided in the original articles; quantitative effect sizes and statistical parameters are reported where available, and qualitative descriptions are presented when numerical results were not reported.

Discussion

Key findings

Across the four objectives, the findings suggest that collaborative gains in human–AI financial decision-making depend not only on algorithmic performance, but also on the organizational and cognitive conditions surrounding its use. Evaluating financial AI through predictive metrics alone risks overlooking dimensions of collaborative quality that Schmutz et al.¹² and Chiriatti et al.⁷⁰ identify as central to effective human–AI teaming. This concern is supported by the partial independence of task performance and human-centered outcomes observed across several studies. Khalil et al.,⁵⁹ for example, showed that operational risk reduction and managerial flexibility over risk-utilization trade-offs could co-occur, a pairing that human-centered frameworks position as a design target rather than an incidental outcome.¹²

The reviewed studies also show that human professionals function as active cognitive agents rather than passive recipients of AI outputs.⁴³ They engaged in sensemaking, contextual reasoning, and evaluative judgment grounded in domain knowledge, accountability, and organizational norms. However, this engagement was neither uniform nor consistently effective. When collaboration architecture, explainability design, and organizational incentives were aligned, studies reported gains in task performance or human-centered outcomes, although most evidence came from controlled or single-context deployments. These included reduced risky investments through calibrated reliance on explainable AI,⁶⁴ reduced analyst time-to-rule through interactive fraud rule authoring,⁵⁷ improved decision quality through transparency design,⁶¹ improved underwriting consistency through iterative refinement,⁶⁸ and increased client welfare and trust when AI advice was routed through a human intermediary.⁶³

In contrast, where these enabling conditions were absent, collaboration produced no consistent advantage and sometimes introduced systematic biases. In experimental and simulated settings with non-expert participants, algorithm aversion,⁶⁵ trust inertia,⁶⁷ and risk-driven advice rejection⁶² suggest that miscalibrated collaboration may produce outcomes no better than unaided judgment. Whether these patterns persist under real accountability structures remains an open empirical question.

Theoretical contributions

The findings offer preliminary grounding for the frameworks outlined in the literature review section by showing how automation bias, reliance, and collaborative outcomes vary across decision architectures. The three collaboration architectures identified in this review map onto distinct vulnerability profiles. Human-as-final-decision-maker configurations concentrate automation bias risk at the evaluation stage,³¹ consistent with the commitment bias⁶⁷ and risk-driven trust erosion⁶² observed in the reviewed studies. Human-as-supervisor configurations shift this risk toward delegation calibration.¹² Iterative co-construction distributes cognitive labor across repeated interaction cycles, allowing expert judgment and data-driven evidence to be progressively reconciled.⁶⁸ This configuration may create conditions under which miscalibration is corrected rather than compounded.³¹ Domain expertise also appeared important for calibrated reliance, consistent with Hickman and Langer’s²⁹ proposition that AI may augment performance through the substitution or supplementation of declarative and procedural knowledge.

Trust miscalibration was bidirectional, context-sensitive, and socially mediated. This aligns with HMI scholarship suggesting that trust in AI draws partly on social-cognitive schemas from interpersonal interaction while operating under distinct accountability and epistemic conditions.⁷¹ The coexistence of algorithm aversion^62,65 and commitment bias⁶⁷ supports the view that the target state is calibration rather than trust per se. Multilevel trust frameworks^71,72 are consistent with these patterns, as Yang et al.⁶³ showed that relational and emotional trust mediated AI adoption through a human intermediary. The distinction between trustworthiness and trust is also visible: Judek⁶⁵ found distrust of high-performing algorithms, while Maier et al.’s⁶⁷ robo-advisor generated commitment bias without objective warrant.¹³

At the organizational level, incentive alignment, knowledge adaptability, and AI literacy appeared as relevant moderators.⁷³ Luong et al.’s⁶⁰ finding that decision errors decreased only when high AI accuracy coincided with organizational incentive alignment is consistent with the argument that AI effectiveness is co-determined by organizational design.^12,74 AI-induced technostress may reduce engagement and well-being in banking populations,⁷⁵ although no reviewed study directly measured it. Implementation challenges such as misaligned incentives, unmanaged expectations, and cognitive bias effects are consistent with organizational conditions associated with perceived injustice, reduced autonomy, and privacy concerns,⁵⁴ and are unlikely to be addressed through individual-level interventions alone.

Practical implications

For financial institutions, three priorities emerge. First, deployment should begin with organizational readiness rather than tool selection alone. Performance and empowerment gains documented across studies^61,63,68 depended on institutional conditions such as alignment between human and AI performance goals and ongoing monitoring of reliance calibration.⁶⁰

Second, evidence from contrasting trust patterns across explainable and opaque systems suggests that architecturally embedded, interactive explainability may support calibrated reliance more effectively than static post hoc explanations.⁴¹ Interactive tools that allow users to explore outputs, compare alternatives, and interrogate reasoning pathways appear more likely to support meaningful oversight.^57,61 Clarity over whether systems are designed to “assist” or “augment” human decision-makers also remains a design-level decision with implications for accountability.⁷⁶

Third, individual differences in technology affinity and domain expertise were among the most consistent moderators of reliance calibration across the reviewed studies.^64,65 Institutions may therefore consider these characteristics in role assignment, team composition, and onboarding, including training on how AI outputs should be interpreted, challenged, and integrated with domain judgment.

For regulators, three areas for attention emerge. First, existing frameworks that mandate human involvement without specifying meaningful engagement may leave important failure mechanisms unaddressed, including commitment bias, social framing effects,⁶³ and risk-driven trust erosion.⁶² Minimum standards for explanation quality therefore represent a tractable regulatory intervention, provided they support sensemaking rather than merely satisfy transparency requirements. Such standards might require interactive explanation mechanisms, counterfactual exploration, and evaluation against decision quality outcomes rather than user satisfaction alone.^57,61,64

Second, the three collaboration architectures identified in this review, human-as-final-decision-maker, human-as-supervisor, and iterative co-construction, may offer regulators a practical taxonomy for specifying oversight requirements proportionate to delegated decision authority. This could help address accountability ambiguities that current frameworks leave unresolved. Third, guidance on AI literacy as an operational condition of deployment would complement technical standards with human capital requirements. The organizational failure conditions identified by Zhou et al.,⁵⁴ including perceived injustice and reduced autonomy, suggest that explanation standards may support not only transparency, but also calibrated reliance and employee well-being.

Limitations and future research

The small evidence base warrants interpretive caution. Only 13 studies met the inclusion criteria, reflecting the early stage of empirical documentation relative to the scale of AI deployment in banking. Most were conducted in controlled or simulated settings, limiting ecological validity and leaving open whether findings transfer to professional populations operating under real accountability structures. Geographic concentration in North America, Europe, and East Asia further limits transferability across regulatory, institutional, and cultural contexts.

The predominance of cross-sectional designs also leaves unresolved how trust calibration and collaboration quality evolve during sustained deployment, a critical gap given evidence of both temporal inertia⁶⁷ and iterative improvement.⁶⁸ Several financial decision domains, including trading, anti-money-laundering, insurance underwriting, and regulatory compliance, remain empirically underexamined from a teaming perspective.^2,24 Healthcare research has begun to formalize governance frameworks for continuously learning human-in-the-loop systems, highlighting the absence of comparable regulatory clarity in financial decision environments.⁷⁷

Future research should prioritize longitudinal field studies of trust calibration and reliance in banking environments, complemented by organizational ethnographies of incentive structures, knowledge dynamics, and workflow adaptations. Cross-country comparative studies would clarify how regulatory frameworks and institutional cultures shape collaboration architectures. Further work should assess both objective performance and human-centered outcomes using validated measures of empowerment, fairness, and well-being. AI literacy also warrants investigation, including whether structured training improves reliance calibration and reduces automation bias among financial professionals.⁵¹ Finally, validated teamwork frameworks have not yet been applied to human–AI teams in financial contexts,²⁷ and future studies should examine how generative AI capabilities, including autonomous reporting and contextual reasoning,⁷⁸ alter authority configurations.

Conclusions

The current evidence suggests a coherent but not definitive account of human–AI teaming in financial decision-making. Across the reviewed studies, AI systems typically operated as the initial analytic component in the decision flow, while humans interpreted, evaluated, and authorized outputs. The human role appears more cognitively demanding and evaluative than “human-in-the-loop” rhetoric often implies, involving sensemaking, calibrated trust, and contextual judgment. However, whether these patterns hold under live conditions with professional populations, across regulatory and cultural contexts beyond the geographically concentrated current evidence base, and over sustained longitudinal deployment remains to be established.

Explainability was more consistently associated with calibrated reliance when architecturally embedded and interactively designed than when delivered as static post hoc output,⁴¹ although the evidence remains too limited for generalization. Organizational conditions, including incentive alignment, governance clarity, AI literacy, and adaptable knowledge structures, appeared at least as consequential as technical AI properties. Effective banking decision-making may therefore depend on preserving and augmenting distinctively human capabilities, including judgment, relational intelligence, and professional accountability, through AI systems designed to support rather than erode them.^12,54

For institutions, interactive explanation design and AI literacy may support more calibrated human–AI teaming,^{51,57,61,64,65,68} although live-deployment evidence remains limited. For regulators, specifying conditions for meaningful engagement, including explanation quality and collaborative process design, may offer stronger governance than mandating human involvement alone.^62,63,67 Longitudinal field studies with financial professionals across diverse regulatory and institutional contexts remain a methodological priority.

Footnotes

ORCID iDs

Fatma Aksu

Sofia Morandini

Luca Pietrantoni

Ethical considerations

This study is a scoping review of published literature and did not involve primary data collection. Ethical approval was therefore not required.

Consent to participate

Informed consent was therefore not required.

Author contributions

Fatma Aksu: Conceptualization, methodology, literature screening, data extraction, analysis, writing, original draft preparation.

Sofia Morandini: Conceptualization, methodology refinement, supervision, critical revision of the manuscript.

Luca Pietrantoni: Conceptualization, supervision, critical revision of the manuscript.

All authors reviewed and approved the final manuscript.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The data supporting the findings of this study are derived from publicly available published articles. The data extraction table is available.*

References

Sai

Arunakar

Chamola

, et al. Generative AI for finance: applications, case studies and challenges. Expert Syst 2025; 42: e70018. https://doi.org/10.1111/exsy.70018

Qudah

Almustafa

Almnadheh

. Exploring the relationship between Fintech solutions and HRM efficiency in Jordanian banking. Hum Syst Manag 2024; 44: 408–423. https://doi.org/10.1177/01672533241295608

Fares

Butt

Lee

SHM

. Utilization of artificial intelligence in the banking sector: a systematic literature review. J Financ Serv Market 2022; 27: 241–259. https://doi.org/10.1057/s41264-022-00176-7

Singh

Mishra

Kumar

, et al. Applications of artificial intelligence for optimizing banking and financial operations: a systematic literature review and future research agenda. J Econ Surv 2025; 39: 2284–2302. https://doi.org/10.1111/joes.12693

Simón

Revilla

Saenz

. Integrating AI in organizations for value creation through human-AI teaming: a dynamic capabilities approach. J Bus Res 2024; 182: 114783. https://doi.org/10.1016/j.jbusres.2024.114783

Cho

. Factors affecting human–AI collaboration performances in financial sector: sustainable service development perspective. Sustainability 2025; 17(10): 4335. https://doi.org/10.3390/su17104335

Vuković

Dekpo-Adza

Matović

. AI integration in financial services: a scientometric review of trends, challenges, and governance. Hum Soc Sci Commun 2025; 12: 562. https://doi.org/10.1057/s41599-025-04850-8

European Parliament and Council of the European Union . Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Off J Eur Union 2024; L 2024/1689, Available at: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=OJ:L_202401689

Kim

Park

. Trust dynamics in financial decision making: behavioral responses to AI and human expert advice following structural breaks. Behav Sci 2024; 14(10): 964. https://doi.org/10.3390/bs14100964

10.

Han

. Trust formation, error impact, and repair in human–AI financial advisory: a dynamic behavioral analysis. Behav Sci 2025; 15(10): 1370. https://doi.org/10.3390/bs15101370

11.

Berretta

Tausch

Ontrup

, et al. Defining human-AI teaming the human-centered way: a scoping review and network analysis. Front Artif Intell 2023; 6: 1250725. https://doi.org/10.3389/frai.2023.1250725

12.

Schmutz

Outland

Kerstan

, et al. AI-teaming: redefining collaboration in the digital era. Curr Opin Psychol 2024; 58: 101837. https://doi.org/10.1016/j.copsyc.2024.101837

13.

Afroogh

Akbari

Malone

, et al. Trust in AI: progress, challenges, and future directions. Hum Soc Sci Commun 2024; 11: 1568. https://doi.org/10.1057/s41599-024-04044-8

14.

Liu

Zhang

Chen

, et al.

Human-AI teaming in healthcare: 1+1>2?

npj Artif Intell 2025; 1: 47. https://doi.org/10.1038/s44387-025-00052-4

15.

Bigman

Briker

Langer

. Human-AI interaction research needs to be embedded in psychological theory. Nat Rev Psychol 2026; 5: 233–234. https://doi.org/10.1038/s44159-026-00551-4

16.

Arksey

O’Malley

. Scoping studies: towards a methodological framework. Int J Soc Res Methodol 2005; 8: 19–32. https://doi.org/10.1080/1364557032000119616

17.

Khandani

Kim

. Consumer credit-risk models via machine-learning algorithms. J Bank Finance 2010; 34: 2767–2787. https://doi.org/10.1016/j.jbankfin.2010.06.001

18.

Mer

Virdi

. The influence of artificial intelligence on personal and corporate finance: a comprehensive literature review with prospects for future research. In: Kour

Taneja

Özen

, et al. (eds). Financial Landscape Transformation: Technological Disruptions. Emerald Publishing, 2025. https://doi.org/10.1108/978-1-83753-750-120251014

19.

Jagtiani

Lemieux

. The roles of alternative data and machine learning in fintech lending: evidence from the LendingClub consumer platform. Financ Manag 2019; 48: 1009–1029. https://doi.org/10.1111/fima.12295

20.

Cao

. AI in finance: challenges, techniques, and opportunities. ACM Comput Surv 2022; 55: 1–38. https://doi.org/10.1145/3502289

21.

Davis

. Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS Q 1989; 13: 319–340. https://doi.org/10.2307/249008

22.

Venkatesh

Morris

Davis

, et al. User acceptance of information technology: toward a unified view. MIS Q 2003; 27: 425–478. https://doi.org/10.2307/30036540

23.

Cai

Winter

Steiner

, et al. “Hello AI”: uncovering the onboarding needs of medical practitioners for human-AI collaborative decision-making. Proc ACM Hum-Comput Interact 2019; 3(CSCW): 1–24. https://doi.org/10.1145/3359206

24.

Alhyasat

Khattab

. Technology trends in strategic management in the AI era: a systematic literature review. Hum Syst Manag 2025; 44: 764–781. https://doi.org/10.1177/01672533251322396

25.

Buccinca

Malaya

Gajos

. To trust or to think. Proc ACM Hum-Comput Interact 2021; 5(CSCW1): 1–21. https://doi.org/10.1145/3449287

26.

Stylos

Okumus

Onder

. Beauty or the Borg: agentic artificial intelligence organizational socialization in synergistic hybrid transformative dynamic flows. Tour Manag 2025; 111: 105205. https://doi.org/10.1016/j.tourman.2025.105205

27.

Salmon

King

Hall

, et al. The big five model of teamwork and human autonomy teams: a scoping review. Appl Ergon 2026. https://doi.org/10.1016/j.apergo.2026.104761, Epub ahead of print 26.

28.

Blaurock

Büttgen

Schepers

JJL

. Designing collaborative intelligence systems for employee-AI service co-production. J Serv Res 2025; 28(4): 544–562. https://doi.org/10.1177/10946705241238751

29.

Hickman

Langer

. Integrating generative AI into theories of job performance. Res Hum Resour Manag. In press. Preprint: https://doi.org/10.13140/RG.2.2.18113.01128

30.

Parasuraman

Sheridan

Wickens

. A model for types and levels of human interaction with automation. IEEE Trans Syst Man Cybern Syst Hum 2000; 30: 286–297. https://doi.org/10.1109/3468.844354

31.

Cummings

. Automation bias in intelligent time critical decision support systems. In: AIAA 1st intelligent systems technical conference, Chicago, IL, 20–22 September 2004, paper no. AIAA-2004-6313. https://doi.org/10.2514/6.2004-6313

32.

Schmager

Pappas

Vassilakopoulou

. Understanding human-centred AI: a review of its defining elements and a research agenda. Behav Inf Technol 2025; 44: 3771–3810. https://doi.org/10.1080/0144929X.2024.2448719

33.

Gao

. Applying HCAI in developing effective human-AI teaming: a perspective from human-AI joint cognitive systems. Interactions 2024; 31: 32–37. https://doi.org/10.48550/arXiv.2307.03913

34.

Sycara

Lewis

. Integrating agents into human teams. In: Salas

Fiore

(eds). Team Cognition. APA, 2004, pp. 203–231.

35.

Dietvorst

Simmons

Massey

. Algorithm aversion: people erroneously avoid algorithms after seeing them err. J Exp Psychol Gen 2015; 144: 114–126. https://doi.org/10.1037/xge0000033

36.

Roser

Meinikheim

Muzalyova

, et al. Artificial intelligence-assisted endoscopy and examiner confidence: a study on human-artificial intelligence interaction in Barrett’s esophagus. DEN Open 2025; 6: e70150. https://doi.org/10.1002/deo2.70150

37.

Logg

Minson

Moore

. Algorithm appreciation: people prefer algorithmic to human judgment. Organ Behav Hum Decis Process 2019; 151: 90–103. https://doi.org/10.1016/j.obhdp.2018.12.005

38.

Lee

See

. Trust in automation: designing for appropriate reliance. Hum Factors 2004; 46: 50–80. https://doi.org/10.1518/hfes.46.1.50.30392

39.

Valori

Kraus

Fairhurst

. Interdisciplinary perspectives and current findings on the role of trust as a psychological mediator in human interaction with artificial intelligence: editorial overview. Comput Hum Behav 2026; 164: 108957. https://doi.org/10.1016/j.chb.2026.108957

40.

Ulfert

Georganta

Centeio

, et al. Shaping a multidisciplinary understanding of team trust in human-AI teams. Eur J Work Organ Psychol 2024; 33: 158–171. https://doi.org/10.1080/1359432X.2023.2200172

41.

Doshi-Velez

Kim

. Towards a rigorous science of interpretable machine learning. ArXiv Preprint arXiv:1702.08608 2017. https://doi.org/10.48550/arXiv.1702.08608, Epub ahead of print 28.

42.

Wang

Yin

. Are explanations helpful? A comparative study of the effects of explanations in AI-assisted decision-making. In: Proceedings of the 26th International Conference on Intelligent User Interfaces, College Station, TX, 13–17 April 2021, pp. 318–328. https://doi.org/10.1145/3397481.3450650

43.

Weick

. Sensemaking in organizations. Sage, 1995.

44.

Bauer

von Zahn

Hinz

. Expl(AI)ned: the impact of explainable artificial intelligence on users’ information processing. Inf Syst Res 2023; 34(4): 1582–1602. https://doi.org/10.1287/isre.2023.1199

45.

Maathuis

Stalenhoef

Van Otterloo

, et al. Designing effective explainable AI: a human-centered evaluation of explanation formats in financial decision-making. Front Artif Intell 2026; 9: 1668029. https://doi.org/10.3389/frai.2026.1668029

46.

Trist

Bamforth

. Some social and psychological consequences of the longwall method of coal-getting. Hum Relat 1951; 4: 3–38. https://doi.org/10.1177/001872675100400101

47.

Kreuzer

Papapetrou

Zdravkovic

. AI explainability methods in digital twins: a model and a use case. In: Enterprise design, operations, and computing. Springer, 2025. https://doi.org/10.1007/978-3-031-78338-8_1

48.

Purohit

Banerjee

. Artificial intelligence-based organizational decision-making in recruitment practice. Hum Syst Manag 2025; 44: 173–186. https://doi.org/10.3233/HSM-240044

49.

Mendy

Jain

Thomas

. Artificial intelligence in the workplace – challenges, opportunities and HRM framework: a critical review and research agenda for change. J Manag Psychol 2025; 40: 517–538. https://doi.org/10.1108/JMP-05-2024-0388

50.

Hang

Hussain

Amin

, et al. The moderating effects of technostress inhibitors on techno-stressors and employee’s well-being. Front Psychol 2022; 12: 821446. https://doi.org/10.3389/fpsyg.2021.821446

51.

Teh

Lim

, et al. The resourceful academic: AI literacy’s influence on service-oriented organizational citizenship behavior via work engagement and stress in academia. Hum Syst Manag 2026; 01672533261430175. https://doi.org/10.1177/01672533261430175

52.

Benitez

Ruiz

Castillo

, et al. How corporate digital transformation affects the firm’s value through IT capitalization: the moderating role of IT strategic leadership. Inf Manag 2020; 57: 103387. https://doi.org/10.1016/j.im.2020.103387

53.

Zhu

. The impact of corporate digital responsibility in the digital technology environment. Hum Syst Manag 2025; 44: 1017–1032. https://doi.org/10.1177/01672533251342993

54.

Zhou

Wang

Chen

. The dark side of AI-enabled HRM on employees based on AI algorithmic features. J Organ Change Manag 2023; 36: 1222–1241. https://doi.org/10.1108/JOCM-10-2022-0308

55.

Tricco

Lillie

Zarin

, et al. PRISMA extension for scoping reviews (PRISMA-ScR): checklist and explanation. Ann Intern Med 2018; 169: 467–473. https://doi.org/10.7326/M18-0850

56.

Wohlin

. Guidelines for snowballing in systematic literature studies and a replication in software engineering. In: Proceedings of the 18th international conference on Evaluation and Assessment in Software Engineering (EASE’ 14), London, 13–14 May 2014, pp. 1–10, article no. 38. https://doi.org/10.1145/2601248.2601268

57.

Ding

Seleznev

Kumar

, et al. From detection to action: a human-in-the-loop toolkit for anomaly reasoning and management. In: Fourth ACM International Conference on AI in Finance (ICAIF), New York, November 2023, pp. 279–287. https://doi.org/10.1145/3604237.3626872

58.

How

Cheah

Khor

, et al. Artificial intelligence-enhanced predictive insights for advancing financial inclusion. Big Data Cogn Comput 2020; 4: 8. https://doi.org/10.3390/bdcc4020008

59.

Khalil

Hadid

Padmanabhan

, et al. An integrated AI and optimization model for operational efficiency in letter of credit examination. Decis Anal J 2025; 14: 100552. https://doi.org/10.1016/j.dajour.2025.100552

60.

Luong

Kumar

Lang

. Man-machine collaboration in organizational decision-making: an experimental study using loan application evaluations. In: Proceedings of the 2019 Pre-ICIS SIGDSA Symposium, 2019. paper no. 17. https://doi.org/10.2139/ssrn.5028371

61.

Mandava

Vellela

Gorintla

, et al. Evaluating the impact of explainable AI on user trust in financial decision-support systems. In: International Conference on Computational Robotics, Testing and Engineering Evaluation (ICCRTEE), Virudhunagar, India, 2025, pp.1–6. https://doi.org/10.1109/ICCRTEE64519.2025.11052919

62.

Strunk

Nissen

Smolnik

. All risks ain’t the same – a risk facets perspective on AI-based decision support systems. Decis Support Syst 2025; 199: 114557. https://doi.org/10.1016/j.dss.2025.114557

63.

Yang

Bauer

, et al. My advisor, her AI and me: evidence from a field experiment on human-AI collaboration and investment decisions. Manag Sci 2025; 72: 242–264. https://doi.org/10.1287/mnsc.2022.03918

64.

Dikmen

Burns

. The effects of domain knowledge on trust in explainable AI and task performance: a case of peer-to-peer lending. Int J Hum Comput Stud 2022; 162: 102792. https://doi.org/10.1016/j.ijhcs.2022.102792

65.

Judek

. Willingness to use algorithms varies with social information on weak vs. strong adoption. FinTech 2024; 3: 55–65. https://doi.org/10.3390/fintech3010004

66.

Liu

Yang

Lyu

, et al. The impact of interactional justice on service evaluations in banking. Acta Psychol 2025; 261: 105805. https://doi.org/10.1016/j.actpsy.2025.105805

67.

Maier

Menold

McComb

. The relationship between performance and trust in AI in e-finance. Front Artif Intell 2022; 5: 891529. https://doi.org/10.3389/frai.2022.891529

68.

Sachan

Almaghrabi

Yang

, et al. Human-AI collaboration to mitigate decision noise in financial underwriting. Int Rev Financ Anal 2024; 93: 103149. https://doi.org/10.1016/j.irfa.2024.103149

69.

De Oliveira Filho

De Almeida Filho

Ferreira

, et al. Fuzzy-based decision support system for private banking. Pesqui Oper 2023; 43: e264857. https://doi.org/10.1590/0101-7438.2023.043.00264857

70.

Chiriatti

Bergamaschi Ganapini

Panai

, et al. System 0: transforming artificial intelligence into a cognitive extension. Cyberpsychol, Behav Soc Netw 2025; 28: 534–542. https://doi.org/10.1089/cyber.2025.0201

71.

Ulfert-Blank

Georganta

Tielman

, et al. Piecing together the puzzle: understanding trust in human-AI teams. In: Murukannaiah

Hirzle

(eds). Proceedings of the Workshops at the Second International Conference on Hybrid Human-Artificial Intelligence (HHAI 2023). CEUR Workshop Proceedings. CEUR-WS.org, 2023, vol 3456, pp. 169–174.

72.

Georganta

Ulfert

. My colleague is an AI! trust differences between AI and human teammates. Team Perform Manag 2024; 30: 23–37. https://doi.org/10.1108/TPM-07-2023-0053

73.

Khushk

Zhiying

, et al. Navigating human-AI dynamics: implications for organizational performance (SLR). Int J Organ Anal 2025; 33: 1293–1311. https://doi.org/10.1108/IJOA-04-2024-4456

74.

Amershi

Weld

Vorvoreanu

, et al. Guidelines for human-AI interaction. In: Proceedings of the 2019 CHI conference on human factors in computing systems, Glasgow, Scotland, 4–9 May 2019, paper no. 3. https://doi.org/10.1145/3290605.3300233

75.

Pothuganti

Ramachandra

Rama Devi

. The impact of technostress on employee engagement and well-being. Hum Syst Manag 2025; 44: 819–835. https://doi.org/10.1177/01672533251331503

76.

Raisch

Krakowski

. Artificial intelligence and management: the automation–augmentation paradox. Acad Manag Rev 2021; 46(1): 192–210. https://doi.org/10.5465/amr.2018.0072

77.

Olawade

Plabon

Ojo

, et al. Human-in-the-loop artificial intelligence in healthcare. Int J Med Inf 2026; Article 106362. https://doi.org/10.1016/j.ijmedinf.2026.106362

78.

Barile

Secundo

Bussoli

. Exploring artificial intelligence robo-advisor in banking industry. Manag Decis 2025; 63: 3323–3349. https://doi.org/10.1108/MD-08-2023-1324