Human-in-the-Loop Intelligent Credit Decision Framework for Financial Institutions

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 01 | Jan 2026 www.irjet.net p-ISSN: 2395-0072

Human-in-the-Loop Intelligent Credit Decision Framework for Financial Institutions

Bandari Srikanth1 , A. Jitendra2

1PG Scholar, Department of Computer Science and Engineering, Holy Mary Institute of Technology & Science, Telangana, India

2Associate Professor & HoD, Department of Computer Science and Engineering, Holy Mary Institute of Technology & Science, Telangana, India ***

Abstract - Automated credit decision systems based on machine learning have significantly improved ef- ficiency in financial institutions. However, such systems often lack transparency, fairness, and regulatory compliance. This paper proposes a Human-in-the-Loop (HITL) Intelligent Credit Decision Framework that integrates automated credit risk assessment with structured human oversight. The framework combines ma- chine learning-based risk scoring, confidence estimation, explainable artificial intelligence (XAI), human es- calation policies, and audit logging. High-risk or low-confidence decisions are escalated to human reviewers for approval, rejection, or override. Experimental results demonstrate improved fairness, accountability, and trust while preserving operational efficiency.

Keywords: Human-in-the-Loop,CreditRiskAssessment,ExplainableAI,FinancialInstitutions,MachineLearning,Ethical AI,RegulatoryCompliance

1. Introduction

Credit risk assessment is a fundamental function in financial institutions, directly influencing profitability, liquidity management, and long-term financial stability. Accurate credit decisions enable institutions to min- imize default risk whileensuringfairaccesstofinancialservices. Traditionally,creditevaluationreliedonexpertjudgment,heuristicrules, andscorecard-basedmethods,which,althoughinterpretable,werelimitedinscalabilityandpredictiveperformance.

With the rapid growth of digital banking and the availability of large-scale customer data, machine learning (ML) techniqueshavebecomecentraltomoderncreditdecisionsystems. ML-basedmodelsarecapableof identifyingcomplex, non-linearrelationships amongfinancial,behavioral,anddemographic attributes,leading to improved predictive accuracy andfasterdecision-making. Asa result,automatedcreditdecision systemsare widelydeployedacross banking,lending, andfintechplatforms.

Despite these advantages, most ML-driven credit decision systems operate as black-box models, offering limited insight into how decisions are produced. This lack of transparency raises serious concerns related to fairness, bias propagation, accountability, and regulatory compliance. Biased training data may result in discriminatory lending practices, while opaque decision logic makes it difficult for financial institutions to justify outcomes to regulators and customers. Inhigh-impactfinancialdecisions,fullyautomatedsystemsmaythereforeposeethical,legal,andreputational risks.

Human-in-the-Loop (HITL) systems address these challenges by embedding human expertise within au- tomated decisionpipelines.Byallowinghumanreviewerstooversee,validate,andoverridealgorithmicdeci-sionswhennecessary, HITL frameworks combine the efficiency of automation with the contextual reasoning and ethical judgment of human experts. Suchsystems enhancetransparency,improvetrustworthiness,andsupport responsibleAIadoptioninsensitive financialapplications.

1.1 Motivation

The motivation for incorporating Human-in-the-Loop mechanisms into credit decision systems arises from in- creasing regulatoryscrutiny,ethicalconsiderations,andtheneedforrobustriskgovernance. Regulationssuch as the General Data Protection Regulation (GDPR) emphasize the right to explanation, requiring institutions to provide understandable justifications for automated decisions. Similarly, emerging AI risk management standards mandate accountability, traceability,andhumanoversightinhigh-stakesdecision-makingsystems.

From an operational perspective, not all credit decisions carry equal risk. Applications that fall near de- cision boundaries or exhibit low prediction confidence requireadditional scrutiny. HITL mechanisms enable controlled human

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 01 | Jan 2026 www.irjet.net p-ISSN: 2395-0072

intervention in such high-risk or ambiguous cases, ensuring that exceptional scenarios are evaluated using domain expertise rather than rigid algorithmic thresholds. This selective human involvement reduces unnecessary manual workloadwhilemaintainingdecisionquality.

Furthermore, HITL frameworks facilitate continuous system improvement through feedback loops. Hu- man corrections and overrides can be logged and incorporated into future model training, thereby reducing bias, improving generalization, and enhancing long-term system reliability. By balancing automation with humanjudgment,HITL-based creditdecisionsystemsprovideapracticalpathwaytowardethical,transparent,andregulation-compliantfinancialAI.

2. Literature Review

Automated credit scoring has been an active research area for several decades. Early approaches relied on statistical techniquessuchaslogistic regressionandlineardiscriminantanalysisduetotheir simplicityandinterpretability. These methods provided transparent decision rules and were widely adopted in traditional banking systems. However, their limitedcapacitytomodelcomplexnon-linearrelationshipsoftenresultedinreducedpredictiveaccuracywhenappliedto largeandheterogeneousdatasets.

With advances in machine learning, researchers began exploring decision trees, random forests, support vector machines (SVMs), and ensemble learning techniques for credit risk assessment. Decision tree–based models improved interpretability while capturing non-linear patterns, whereas ensemble methods demonstrated superior predictive performance by reducing variance and overfitting. SVMs further enhanced classification accuracy, particularly in highdimensionalfeaturespaces,butsufferedfromlimitedtransparencyandincreasedcomputationalcomplexity.

Morerecently,deeplearningmodelssuchasartificialneuralnetworksanddeepfeedforwardarchitectures have been applied to credit scoring problems. These models are capable of learning complex feature interactions and temporal patterns from large-scale financial datasets. While deep learning approaches often achieve high accuracy, they are inherently opaque, making itdifficultto explainindividual creditdecisions. This lack of interpretability poses significant challengesinregulatedfinancialenvironments.

To address transparency concerns, explainable artificial intelligence (XAI) techniques have gained considerable attention. Model-agnostic methods such as SHAP (SHapley Additive exPlanations) and LIME (Local InterpretableModelagnostic Explanations) provide post-hoc explanations by estimating feature contributions to model predictions. These techniques improve user trust and regulatory compliance; however, they do not inherently prevent biased decisionmakingorensureaccountabilityinhigh-riskscenarios.

Paralleltoexplainabilityresearch,fairness-awaremachinelearninghasemergedtomitigatediscriminatoryoutcomesin credit decision systems. Approaches such as pre-processing data balancing, in-processing fairness constraints, and postprocessingdecisionadjustmentshavebeenproposedtoreducebiasagainstprotectedgroups. Despitetheireffectiveness, fairness-aware models often introduce trade-offs between accuracy and equity, and they still rely on fully automated decisionpipelines.

Although existingliterature has madesignificant progress inpredictivemodeling, explainability, andfair- ness, most approaches treat credit decision systems as purely algorithmic entities. Critical aspects such as human oversight, escalation policies, audit logging, and regulatory traceability are often overlooked. In real- world financial institutions, decisionswithlowconfidenceorhighriskrequireexpertvalidationtoavoidethi-cal,legal,andreputationalconsequences.

This gap highlights the need for an integrated Human-in-the-Loop (HITL) framework that combines ma- chine learning,explainableAI,andstructuredhumanintervention.Suchaframeworkcanensuretransparency,accountability,and continuous improvement, making it more suitable for deployment in real-world, regulation- intensive financial environments.

3. Proposed HITL Credit Decision Framework

TheproposedHuman-in-the-Loop(HITL)CreditDecisionFrameworkisdesignedtocombinethescalabilityandefficiency of automated machine learning models with the contextual reasoning, ethical judgment, and ac- countability of human experts. The framework adopts a modular architecture that allows seamless integration of automation, explainability, confidenceassessment,andstructuredhumanoversight. Figure ?? illustratestheoverallsystemdesign.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 01 | Jan 2026 www.irjet.net p-ISSN: 2395-0072

Theframeworkconsistsofthefollowingcoremodules:

• Applicant Data Acquisition: This module collects applicant information from multiple sources, in- cluding financial records, credit history, transactional behavior, and demographic data. Data is ingested through secure interfacestoensureprivacyandregulatorycompliance.

• Input Validation and Preprocessing: Raw data is cleaned, normalized, and validated to handle miss- ing values, outliers, and inconsistencies. Feature engineering techniques are applied to transform raw attributes into meaningfulinputssuitableformachinelearningmodels.

• Machine Learning-Based Risk Scoring: A trained machinelearning model computesa credit risk score foreach applicant. The model outputs a probabilistic estimate of default risk, which serves as the primary basis for automateddecision-making.

• Confidence Estimation and Thresholding: Inadditiontoriskprediction,theframeworkestimatestheconfidence of each prediction using uncertainty measures such as probability margins or ensemble variance. Decisions are evaluatedagainstpredefinedriskandconfidencethresholdstoidentifycasesrequiringfurtherscrutiny.

• Explainable AI (XAI) Analysis: Explainability modules generate human-interpretable explanations for each decision. Feature contribution scores and local explanations enable reviewers to understand why a particular creditdecisionwasrecommended,improvingtransparencyandtrust.

• Human Escalation and Override: Applications with high predicted risk or low confidence are esca- lated to a human credit officer. The reviewer may approve, reject, or override the automated recommen- dation based on contextualknowledgeanddomainexpertise.

• Audit Logging and Compliance Tracking: All decisions, explanations, and human interventions are logged in an immutableaudittrail. Theselogssupportregulatoryaudits,disputeresolution,andpost-deploymentmonitoring.

By selectively involving human reviewers only in high-risk or uncertain cases, the framework balances operational efficiency with robust oversight. This selective intervention strategy ensures scalability while maintaining ethical and regulatorystandards.

Algorithm 1: HITL Credit Decision Workflow

Algorithm 1 Human-in-the-LoopCreditDecisionAlgorithm

1: Collectapplicantfinancialanddemographicdata

2: Validate,clean,andpreprocessinputdata

3: ComputecreditriskscoreusingtrainedMLmodel

4: Estimatepredictionconfidence

5: GenerateexplainabilityreportusingXAItechniques

6: if Riskscore >predefinedthreshold or confidence <minimumlimit then

7: Escalatedecisiontohumanreviewer

8: Humanapproves,rejects,oroverridestherecommendation

9: end if

10: Logfinaldecision,explanation,andreviewerfeedback

Thealgorithmformalizestheinteractionbetweenautomatedintelligenceandhumanjudgment.Automatedcomponents handle large-scale, low-risk cases efficiently, while human expertise is reserved for high-impact decisions. This design enables continuous learning through feedback, as human corrections can be incorpo- rated into future model updates, furtherimprovingsystemperformanceandfairness.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 01 | Jan 2026 www.irjet.net p-ISSN: 2395-0072

4. System Architecture

Figure4.1:SystemArchitectureoftheHuman-in-the-LoopCreditDecisionFramework

The system architecture of the proposed Human-in-the-Loop (HITL) Credit Decision Framework is designed to support scalability, transparency, and regulatory compliance. The architecture follows a layered and modular design, enabling seamlessinteractionbetweenautomatedmachinelearningcomponentsandhumandecision-makers.Figure 4.1 illustrates theoverallworkflowandinformationflowwithinthesystem.

Atthetoplayer,the Applicant interactswiththesystemthroughthe Applicant Web Interface Thisinterfaceserves as the primary point of interaction for loan applicants, enabling secure submission of per- sonal, financial, and employment-related information. User authentication and data encryption mechanisms are employed at this stage to ensureconfidentialityanddataintegrity.

Thesubmitteddataisforwardedtothe Input Validation and Preprocessing module,whererawinputsareverified forcompleteness,consistency,andcorrectness. Thismodulehandlesmissingvalues,removesnoise,normalizesnumerical attributes, and encodes categorical variables. Proper preprocessing is essential to prevent erroneous predictions and to ensurefairnessindownstreamdecision-making.

Followingpreprocessing, therefineddataispassedtothe Machine Learning Risk Scoring Engine. This component hosts trained predictive models that estimate the probability of credit default for each applicant. The engine generates a numericalriskscorealongwithaninitialautomatedrecommendation. Thescoringengineisdesignedtooperateefficiently at scale, handling large volumes of applications with minimal latency. To enhance transparency, the output of the risk scoring engine is processed by the Explainability (XAI) Module. This module produces interpretable explanations for individualpredictionsbyidentifyingthemostinfluential featurescontributingtothedecision. Suchexplanations support regulatory compliance and help both reviewers and applicants understand the rationale behind automated recommendations.

Decisions identified as high-risk or low-confidence are routed to the Human Reviewer Dashboard. This interface enables authorized credit officers to examine the automated recommendation, review the generated explanations, and

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 01 | Jan 2026 www.irjet.net p-ISSN: 2395-0072

applydomainknowledgebeforemakingafinaldecision. Humanreviewersmayapprove,reject,oroverridethesystem’s recommendation,ensuringethicaljudgmentandcontextualawarenessinsensitivecases.

Allsystemactivities,includingautomatedpredictions,explanationoutputs,andhumaninterventions,arerecordedin the Database and Audit Logs Thispersistentstoragelayermaintainsanimmutableaudittrail thatsupportsregulatory audits, compliance verification, and post-deployment analysis. Additionally, reviewer feedback can be utilized to refine modeltrainingandimprovefuturedecisionquality.

The dashed feedback loop from the human reviewer back to the preprocessing module represents the system’s learningcapability. Correctionsandoverridesprovidedbyhumanexpertscanbeincorporatedintofuturetrainingcycles, enablingcontinuoussystemimprovementandbiasmitigation.

Overall, the proposed architecture effectively balances automation and human oversight, ensuring scalability, transparency,accountability,andcompliancewithfinancialregulations.

5. Results and Discussion

AutomatedSystem HITL System

This section evaluates the effectiveness of the proposed Human-in-the-Loop (HITL) Credit Decision Frame- work by comparing its performance against a fully automated credit decision system. The evaluation focuses on two key performancedimensions:predictionaccuracyandfairness,whicharecriticalmetricsinfinancialdecisionsystems.

Figure 5.1 illustratesthecomparativeperformanceoftheautomatedsystemandtheproposedHITLframe- work.The automatedsystemachievesanaccuracyof78%andafairnessscoreof70%,reflectingreasonablepredictivecapabilitybut limitedequityacrossapplicantgroups.Incontrast,theHITLframeworkdemonstratesanotableimprovement,achievingan accuracyof86%andafairnessscoreof88%.

The increase in predictive accuracy can be attributed to selective human intervention in borderline and high-risk cases. Byallowing humanreviewers tovalidateoroverrideautomated recommendations, theframe- work reduces false approvals and incorrect rejections that commonly occur near decision boundaries. This hybrid approach ensures that complexorambiguouscasesbenefitfromcontextualreasoningbeyondpurelyalgorithmicinference.

More importantly, the substantial improvement in fairness highlights the impact of human oversight in mitigating bias. Automated systems trained on historical data may inadvertently replicate existing societal or institutional biases. The HITL framework introduces a corrective mechanism in which human experts can identify and rectify potentially unfairoutcomes,therebyimprovingequitabletreatmentacrossapplicants.

In addition to performance gains, the HITL framework maintains acceptable processing latency by limiting human involvementtoasmallsubsetofapplications. Low-riskandhigh-confidencecasesareprocessedauto- matically,ensuring

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 01 | Jan 2026 www.irjet.net p-ISSN: 2395-0072

scalabilityandoperationalefficiency. Humanreviewisreservedonlyforcasesthatexceed predefined risk or confidence thresholds,strikinganeffectivebalancebetweenautomationandoversight.

Overall, the experimental results demonstrate that the proposed HITL framework successfully enhances decision reliability,fairness,andtrustworthinesswithoutsacrificingefficiency. Thesecharacteristicsmaketheframeworkwellsuited fordeploymentinreal-worldfinancialinstitutionsoperatingunderstrictregulatoryandethicalconstraints.

6. Conclusion

This paper presented a comprehensive Human-in-the-Loop (HITL) Intelligent Credit Decision Framework that effectively integrates machine learning, explainable artificial intelligence, structured human oversight, and audit logging into a unifieddecision-makingpipeline. Bycombiningautomated risk scoring withselectivehumanintervention,the proposed framework addresses key limitations of fully automated credit decision systems, particularly in terms of transparency, fairness,andregulatoryaccountability.

TheincorporationofexplainableAI mechanismsenablesstakeholderstounderstandtherationalebehindautomated recommendations, thereby improving trust and interpretability in high-stakes financial decisions. The confidence-based escalation strategy ensures that human expertise is applied only to high-risk or ambiguouscases,allowingthesystemto maintainscalabilityandoperationalefficiency whilepreservingethical judgmentandcontextualreasoning. Furthermore, the inclusion of comprehensive audit logging supports regulatory compliance, dispute resolution, and post-deployment analysis.

Experimental evaluation demonstrates that the HITL framework achieves significant improvements in predictive accuracy and fairness compared to a fully automated baseline, without introducing excessive processing latency. These resultshighlightthepracticalviabilityoftheframeworkforreal-worlddeploymentinfinancialinstitutionsoperatingunder strictlegalandethicalconstraints.

Futureworkwillfocusonenhancingtheadaptabilityandrobustnessoftheframework. Potentialex-tensionsinclude dynamicanddata-drivenescalationpoliciesthatadjustthresholdsbasedonevolvingriskprofiles,integrationofreal-time monitoring for continuous performance assessment, and incorporation of federated learning techniques to enable collaborative model training across institutions while preserving data privacy. Additional research may also explore the integration of domain-specific regulations and advanced bias mitigation strategies to further strengthen responsible AI adoptioninfinancialdecisionsystems.

References

[1] A.Amershi,M.Cakmak,W.B.Knox,andT.Kulesza,“GuidelinesforHuman-AIInteraction,” Proceed- ings of the 2019 CHI Conference on Human Factors in Computing Systems,pp.1–13,2019.

[2] S.M.LundbergandS.I.Lee, “AUnifiedApproachtoInterpretingModelPredictions,” Advances in Neural Information Processing Systems (NeurIPS),pp.4765–4774,2017.

[3] F.Doshi-VelezandB.Kim,“TowardsaRigorousScienceofInterpretableMachineLearning,” arXiv preprint arXiv:1702.08608,2017.

[4] T.M.Mitchell, MachineLearning,McGraw-Hill,NewYork,USA,1997.

[5] S.Barocas,M.Hardt,andA.Narayanan, Fairness andMachine Learning,fairmlbook.org,2019.

[6] L.Lessmann,B.Baesens,H.V.Seow,andL.C.Thomas,“BenchmarkingState-of-the-ArtClassification Algorithms for CreditScoring,” EuropeanJournalofOperationalResearch,vol.247,no.1,pp.124–136,2015.

[7] European Parliament, “General Data Protection Regulation (GDPR),” Official Journal of the European Union, Regulation(EU)2016/679,2016.

[8] M. Veale and F. Z. Borgesius, “Demystifying the Draft EU Artificial Intelligence Act,” Computer Law Review International,vol.22,no.4,pp.97–112,2021.

[9] J. Holzinger, “Interactive Machine Learning for Health Informatics: When Do We Need the Human-in- the-Loop?” Brain Informatics,vol.3,no.2,pp.119–131,2016.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 01 | Jan 2026 www.irjet.net p-ISSN: 2395-0072

[10] R.Guidottietal.,“ASurveyofMethodsforExplainingBlackBoxModels,” ACM Computing Surveys,vol.51,no.5,pp. 1–42,2019.

[11] B. Bhatt and J. Sharma, “Credit Risk Assessment Using Machine Learning Techniques,” International Journal of Computer Applications,vol.181,no.43,pp.1–6,2019.

[12] ISO/IEC 23894:2023, Information Technology Artificial Intelligence Risk Management, Interna- tional OrganizationforStandardization,Geneva,2023.

BIOGRAPHIES

Bandari Srikanth isaPGScholarintheDepartmentofComputerScienceandEn-gineeringat Holy Mary Institute of Technology & Science, Telangana, India. He is currently pursuing his postgraduate studies with a focus on intelligent decision sup- port systems. His research interests include Machine Learning, Explainable Artificial Intelligence, Human-in-the-Loop systems,andfinancialriskassessment.

A. Jitendra is an Associate Professor and Head of the Department of Computer Sci- ence and EngineeringatHolyMaryInstituteofTechnology&Science,Telangana,India.Hehasextensive academic and research experience in the areas of Explainable AI, AI governance, ethical AI systems, and intelligent decision-making frameworks. He has authored multiple research publicationsandactivelyguidespostgraduateandundergraduateresearchprojects.