Artificial Intelligence (AI) powered credit scoring and loan underwriting System Proposal for Savings and Credit Cooporative Organizations (SACCO) in Uganda

 Corresponding author: Jimmy Kinyonyi Bagonza Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution License 4.0. Artificial Intelligence (AI) powered credit scoring and loan underwriting System Proposal for Savings and Credit Cooporative Organizations (SACCO) in Uganda Jimmy Kinyonyi Bagonza * School of Computer and Information Sciences, University of the Cumberlands, Kentucky USA. World Journal of Advanced Research and Reviews, 2025, 27(01), 2074-2088 Publication history: Received on 09 June 2025; revised on 15 July 2025; accepted on 17 July 2025 Article DOI: https://doi.org/10.30574/wjarr.2025.27.1.2689 Abstract This proposal outlines the comprehensive implementation of an AI-powered Credit scoring and loan underwriting system for Savings and Credit Cooperative organizations (SACCO) in Uganda. The system leverages alternative data sources like mobile money transactions, utility payments and behavioral indicators to build accurate and inclusive borrower risk profiles. By integrating advanced machine learning (ML) and generative pre-trained transformer (GPT) models, this system aims to overcome key challenges in traditional credit assessments like limited credit histories, collateral requirements, and subjective decisions which have led to high default rates and financial exclusion The proposed AI model enhances operational efficiency by delivering real time risk evolutions, improving loan turnaround times and supporting proactive risk management. Human oversight is maintained through explainable AI outputs and override mechanisms, aligning with Uganda’s data protection laws and SACCO values. A dedicated AI Steering Committee will oversee ethical governance while comprehensive training and stakeholder engagement will ensure smooth adoption. Custom built for SACCOs, the system is designed for full integration with existing platforms such as the Mobis management information systems, mobile money platforms, credit bureaus and includes agentic monitoring for fraud and early warnings. Ultimately the solution strengthens financial inclusion, improves portfolio quality and empowers SACCOs to deliver faster fairer and more inclusive lending services across Uganda. Keywords: AI Powered Credit Scoring; Loan Underwriting; Savings and Credit Cooperative Organization (SACCO); Financial Inclusion; Machine Learning; AI Governance and Oversight 1. Introduction SACCOs in Ugandan operate as community-based member owned financial cooperatives offering savings and loan services to the underserved and rural population 1. Despite their role in financial inclusion initiatives, SACCOs face challenges in credit underwriting ranging from limited access to formal credit histories, insufficient collateral and reliance on subjective decision-making processes 1. Traditional loan assessments base on character reference and sparse documentation that make it difficult to effectively evaluate the risk of members with irregular income such as farmers or traders leading to inefficiencies and high loan defaults. These challenges in credit assessment have dare consequences like SACCOs often suffer with high numbers of nonperforming loans (NPL) forcing them to restrict lending, eroding trust among SACCO members and limiting growth1. Furter more existing underwriting procedures tend to favor well established collateral rich members inadvertently excluding many potentially creditworthy individuals3. With only about 13.4% of Uganda’s population accessing credit World Journal of Advanced Research and Reviews, 2025, 27(01), 2074-2088 2075 from formal institutions, the gap left by these traditional methods presents a significant missed opportunity for SACCOs to expand financial services equitably3. Introducing an AI-powered credit scoring model offers a transformative solution leveraging alternative data sources such as mobile money transaction and utility payment behavior to build a more accurate and inclusive risk profiles 4,5. This model enables SACCOs to evaluate applicants holistically including those with no formal credit history increasing approval rates without compromising on risk 6. The proposed AI model will streamline the loan process, reduce human error and accelerate decision thereby helping SACCOs scale efficiently while enhancing member trust and satisfaction7. 2. Artificial intelligence governance To ensure responsible and strategic oversight of the proposed AI model, SACCOs will establish a cross functional AI Steering Committee (AI Stereo) that will comprise of key leadership roles including8. • Board representative • IT Manager • Head of Credit • Head of Compliance and Risk • Member representative The core responsibilities for the AI Stereo, include but not limited to. • Aligning AI use with the SACCOs cooperative values • Revising the credit policies • Setting usage protocols • Monitoring compliance The Stereo will meet frequently preferably biweekly during the development phase and monthly post launch phase to guide the project, address emerging issues and align with regulatory requirements. 2.1. Artificial Intelligence Policy Framework SACCOs will adopt a structured AI policy framework that is grounded in fairness, transparency, human oversight, privacy and continuous improvement4. Measures will be taken to • Prevent bias in AI outputs by excluding sensitive attributes and conducting fairness audits. • The system will offer clear explainable reasons for each AI recommendation supporting transparency and member trust. • Human in the loop approach will be embedded tasking loan officers or a committee to review for significant decisions to maintain accountability and allow human judgment where needed. • In accordance with the Uganda data protection and privacy act 2019, ethical data use is prioritized with informed consent required from members and strict protocols for data minimization and encryption. • The committee will Continuously monitor the systems performance using parameters like default rate, fairness, approval accuracy, conduct quarterly reviews and organize annual AI audit. • These measures will ensure the model evolves responsibly, remans compliant and aligns with SACCOs mission of member trust and financial empowerment. 3. Proposed AI Solution Overview 3.1. Member and User Experience The proposed AI model is designed to enhance the existing loan process without removing the human interaction members value9. Members can apply for loans through familia channels either in person at the SACCO office or digitally via app or USSD code10. Once the application is submitted, the model instantly evaluates the members data and generates real time credit risk assessment significantly speeding the decision-making process5 7. In office applications loan officers receive AI generated scores and recommendations within minutes allowing them to discuss loan terms with the member almost immediately. For mobile app or USSD applicants the system can prequalify World Journal of Advanced Research and Reviews, 2025, 27(01), 2074-2088 2076 members or fully approve low risk loans automatically with friendly clear message. Even with automated decisions, the SACCO will maintain human contact points to handle questions and appeals ensuring transparency and trust11. For loan officers, the model serves as a decision support tool embedded in their loan management system. It provides a credit score, confidence level, recommendation action, and a natural language explanation of key risk factors. Loan Officers retain the authority over final decisions and can override AI outputs when needed with reasons logged for continuous learning. This setup ensures a faster, more data informed yet still personalized lending experience. 3.2. Artificial Intelligence Architecture Model and Data Pipeline At the heart of SACCOs AI solution is a specialized Generative Pre-trained Transformer (GPT) model designed to enhance credit underwriting through data driven insights12. 3.2.1. Data Ingestion Layer The architecture begins with a robust Data Ingestion Layer pulling in diverse sources including SACCO core banking data, mobile money transactions, utility bill payments, credit bureau reports and SACCO specific behavioral indicators. This multisource approach ensures a comprehensive view of each members financial reliability and is managed through secure APIs or batch integrations with real time or near real time updates 10, 13. 3.2.2. AI Credit scoring Engine The core AI credit scoring engine blends traditional Machine Learning (ML) lie gradient boosting models with GPT capabilities. While the ML model identifies quantitative risk patterns and generates raw credit scores, the GPT component interprets this structured data to produce narrative risk assessments and loan recommendations (OpenAI, 2023). This hybrid approach allows for both predictive accuracy and human readable explanations. The GPT is fine-tuned on past SACCO loan decisions enabling it to mirror real life lending outcomes and improve reasoning2,12. 3.2.3. Output And Integration Layer This layer ensures that AI generated insights are seamlessly embedded within existing SACCO systems. Loan officers interact with the AI model through their Management Information System (MIS) dashboard viewing scores, recommendations and explanations directly within the application interface. These outputs are stored for audit and analytics and can trigger further processes such document preparation or credit bureau reporting. Security is reinforced at every point of integration ensuring data privacy and regulatory compliance. This architecture empowers faster, fairer and more informed lending decisions while continuously learning from outcomes to improve future predictions Figure 1 AI Credit scoring and loan underwriting data flow diagram World Journal of Advanced Research and Reviews, 2025, 27(01), 2074-2088 2077 4. Justification for the AI Generated Credit Risk scoring over Manual Artificial Intelligence especially GPT based models surpass manual credit scoring through several advantages. 4.1. Holistic Data Analysis processing hundreds of traditional and alternative data points rapidly to uncover trends that humans may miss. This allows for richer credit profiles using sources like mobile money activity and utility payments which are often overlooked in manual assessments. By drawing on diverse inputs, the model creates a more comprehensive and accurate view of the member risk 3. 4.2. Consistency and objectivity By eliminating human biases and decision variability between officers. It applies standardized criteria uniformly increasing fairness and compliance with lending regulations 14. Additionally, AI models have Superior Predictive Power identifying complex nonlinear patterns in borrower behavior that traditional models or intuition may not detect. This reduces both false approvals and unjustified rejections thereby improving portfolio quality and reducing defaults 7 4.3. Speed and efficiency delivering real time assessments and reducing loan processing times from days to minutes. GPT models also support adaptive learning that improvs performance as new outcomes are incorporated over time. Finally, GPTs natural language generation enables personalized educational feedback to members strengthening engagement and helping members improve their credit worthiness. Overall, the model transforms underwriting into faster fairer and smarter process aligned with the SACCO mission of inclusive and efficient service delivery 12. 5. Build Vs Buy Considerations SACCOS must weigh whether to build a custom AI credit scoring model solution or buy an off the shelf product. A custom-built model offers critical advantages in aligning with the SACCO unique member base, data environment, and values. Unlike generic models, custom development allows SACCOs to incorporate localized insights such as member participation, rural behavior patterns and loyalty metrics while tightly integrating alternative data sources like mobile money and utility payments. this tailored approach enables more accurate and culturally relevant credit assessments that reflect SACCO specific risk appetite 11. From a governance and compliance perspective, building in house offers better control over data privacy and model transparency. SACCOs can ensure that the model adheres to ethical standards, explainability requirements and Uganda’s data protection laws by retaining data locally and designing the system for interpretability. In contrast commercial vendors may provide opaque “black box” models that are harder to audit or adapt and may expose sensitive member data to third party risks 14. While off the shelf solutions offer speed and ease of deployment they often come with recurring costs and limited flexibility. External scores may supplement the process like credit bureau checks but won’t harness the full power of SACCO specific data. A hybrid strategy of buying tools like cloud infrastructure or GPT APIs while building custom integration and logic may provide a balanced approach. Given strategic importance of credit scoring and the opportunity to leverage local alternative data the preferred route is custom development possibly using open-source AI frameworks and partnering with local developers to ensure long term adaptability and member centric outcomes 6. 6. Integration With SACCO Management Information System (MIS) and External Platforms A successful credit scoring model must seamlessly integrate with existing systems to effectively ingest data and deliver intended outcomes. World Journal of Advanced Research and Reviews, 2025, 27(01), 2074-2088 2078 Core SACCO MIS. The model will plug into the SACCOs core banking management system Mobis. The model will connect via APIs or direct database access while ensuring security. This integration allows the model to fetch member profile data like savings balances, loan records etc. automatically whenever a credit scoring is needed 6. It also allows writing back the model results including scores, recommendations, reason codes into the MIS as part of the member profile data. The user interface can be extended to show the model outputs. 6.1. Mobile Banking and payments Many SACCOs now enable transactions through mobile money integrations10. They use MTN mobile money and Airtel Money to allow members deposit or receive loans. The model will integrate with these mobile money platforms as data sources through an aggregator service or direct APIs provided by Telecoms. Integration may also occur through the SACCO mobile App or USSD code for loan applications where the front end calls for the model scoring API to get instant decision. Thus consequently, the model becomes part of the mobile banking workflow. 6.2. Credit Bureaus Integration with Metropol CRB and Gnu Grid CRB is planned in a way that when a scoring a loan, the system can automatically query the bureaus database using the members National Identification Number (NIN) as the key to retrieve any existing credit history score or report (Gnu Grid CRB, 2022). This requires a memorandum of understanding with the bureau to use their API or web service. The model will incorporate the report from the bureau into assessment. Conversely, the SACCO can report back to the bureau thereby contributing to the members external credit history. This two-way integration ensures that the model is not working in isolation from the broader financial ecosystem. 6.3. Other Data Providers If available, integration with utilities like electricity company Uganda Electricity distribution Company Ltd (UEDCL) database or National Water and Sewerage Cooperation (NWSC) for water provided bill payment verified data. Realistically integration with utility companies may not exist via APIs in Uganda yet but the SACCO can urge members to upload proof of payment for recently cleared utility bills and then use optical character recognition (OCR) to read the member payment history and feed that into the model 6. The model can also integrate with the National ID verification systems to prevent fraud ensuring the person is who they say they are and KYC databases as part of the loan processing life cycle. To ensure all these systems talk to each other, the SACCO will involve a middleware or integration platform that handles data formatting, encryption and errors gracefully. Figure 2 Data sources and Integration World Journal of Advanced Research and Reviews, 2025, 27(01), 2074-2088 2079 7. Proactive risk management AI agent Beyond just scoring loans at application time, the model will have agentic capabilities of acting like an agent to continuously monitoring and taking limited initiatives in risk management. 7.1. Portfolio Monitoring and Early Warning alerts The model will run periodic scans of the SACCOs loan portfolio. By analyzing ongoing data like current loan repayment trends, any new transactions of borrowers, changes in savings balances, it can detect anomalies or signs of risk in real time. For example, if borrowers’ mobile money inflows drop significantly for two consecutive weeks or if a usually active saver suddenly stops saving, the model can flag this as potential early risk indicator5. The model would proactively generate an alert message to the loan officer or the credit risk manager e.g. “alert: Member RM005 income appears to have dropped by 40% this month, consider reaching out to assess any issues”. Similarly, if a member starts to miss small payments even before formal defaults the model will highlight them for follow up. By catching problems early, the SACCO can intervene through loan restructure or providing financial counselling to prevent defaults. This is a shift from reactive to proactive credit management. 7.2. Automated Recommendations and Actions In some cases, the model agent might be allowed to take limited actions on its own under predefined rules. For instance, the SACCO might authorize the model to auto approve very low risk small loans say below Ugx 500,000 for top rated members without waiting for a human to deliver instant credit for small needs or the model could automatically decline applications that fall below certain thresholds with a message the member should for example build more savings first to reduce workload. Any such autonomous actions will be carefully governed and fall back to human review should the member appeal or if certain uncertainties are present. Another agentic feature could be dynamic credit line adjustment, for example if a member risk profile improves over time, the model might suggest increasing their eligible credit limit and notify them or the loan officer of this opportunity11. 7.3. Fraud Detection The model can also act as a guard through using pattern recognition it can watch for fraudulent behavior e.g. if multiple loan applications come in with the same national ID or phone number or if a members data pattern matches known fraud cases perhaps the account suddenly shows large deposits before loan applications could be borrowed just to inflate statements 15. On detecting possible fraud, the model can freeze the process and alert staff for investigation. The agentic monitoring continuously runs in the background12. Figure 3 Proactive risk management AI agent World Journal of Advanced Research and Reviews, 2025, 27(01), 2074-2088 2080 8. Human In the Loop Controls and Feedback Mechanisms (HITL) The SACCO will implement a robust human in the loop protocols to ensure model credit decisions are guided by human judgment and accountability1,14. In the initial phases, all significant model generated recommendations will be reviewed by loan officers or the credit committee. If the staff overrides the model decision, they must provide a justification thereby feeding valuable insights into the system refinement. Guidelines will govern when overrides are appropriate such as cases of limited data or exceptional circumstances. To improve the model, overtime the system will log every decision and its outcome forming a feedback loop for retraining. Analysts will periodically review mismatches between model predictions and actual results. For example, defaults or accurate overrides to identify gaps, retrain the model and enhance performance. Explainability tools like SHAP or LIME will be included to help officers understand which data features influenced each AI recommendation building trust and enabling informed overrides. Members will also have the right to appeal decisions. if denied, they can provide additional documents or context to support their case prompting reevaluation by staff. This ensures the model supports and not replacing the SACCOs commitment to fairness, transparency, and member centered service. As the model matures and proves reliable, some routine decisions may be automated, but humans will always retain the final say in critical credit determinations. Figure 4 Human in loop controls 9. Implementation plan 9.1. Project phases and timelines Implementing the Credit Scoring model system will be approached in structured phases. A proposed timeline with an assumption of kick starting the project in the second quarter (Q2) of the financial year is shown below. World Journal of Advanced Research and Reviews, 2025, 27(01), 2074-2088 2081 Figure 5 Phased implementation timeline Figure 6 Detailed implementation framework Overall, the above timeline estimates the project to be complete between 9 – 10 months from project initiation to full deployment. This can however be adjusted based on resource availability and having built buffers for iteration since getting the model right may take multiple tries. Key milestones like prototype ready, Pilot success criteria met will be used to decide transition from one phase to the next or reviewing timelines. 9.2. Team Structure and Roles Implementing and AI solution requires a multidisciplinary team whose role is so critical to the overall project success and their responsibilities will be tracked with clear deliverables under the Project Managers coordination with periodic reviews by the AI Steering committee. i.e. World Journal of Advanced Research and Reviews, 2025, 27(01), 2074-2088 2082 9.2.1. Executive leadership most likely the SACCO General Manager or Deputy who is charged with championing the project, secure funds, aligns with strategic goals and interfaces with the Board and AI Steering Committee. 9.2.2. Project Coordination taken on by the Project Manager to coordinate progress, manage timelines, and acts as a bridge between the technical team and the business units. he also manages vendor relations and requirements translation 9.3. Technical Team that includes • Machine Learning Engineers to build credit risk models, finetune GPT and evaluate model performance and Data Engineers to set up ETL pipelines, data cleaning and anonymization • Software Developers to build MIS user interface and backend logic and handle system integrations and present model results • Cloud Engineers to ensure secure environment for the model processing and API calls through to management of servers/ cloud infrastructure, security, backup and scaling. 9.4. Business and Domain experts like Credit officers to provide training data labelling, interpret model outputs, validate business relevance and contribute to user adoption. While compliance officers will ensure data privacy, legal compliance, ethical AI usage and align the solution with the SACCO bylaws and regulatory standards. 9.5. External support teams to offer technical guidance like the Makerere University AI Lab, Fintech companies like Ensibuko to offer support with system integration and development partners like the world bank to offer financial support. 10. Data sources and Quality Assurance To ensure a robust and reliable AI powered credit scoring system, the project emphasizes rigorous data quality and preparation. Historical loan data from SACCO, forms the core training dataset capturing details like loam amounts, terms and conditions, repayment outcomes and member profiles. This data will undergo cleaning processes such as handling missing values, correcting errors, and ensuring consistent labeling of outcomes e.g. defining “Default”. Personally Identifiable Information will be anonymized to maintain privacy and focus the model on behavioral patterns rather than individuals. Additional data sources include member savings behavior and mobile money transaction logs. These will be transformed into useful features like average balances, transaction frequency and carefully crossmatch with loan records to maintain consistency. Outliers and missing or incomplete mobile money data will be addressed by validating with members and standardizing inputs. the project also plans to use alternative data like utility bill payment and external credit data which will be verified through random sampling and reconciliation with member reported information to ensure integrity3. Ongoing quality assurance spans from pre-deployment model validation using cross validation and performance metrics like AUC and precision / recall to fairness testing across demographic segments2. Post deployment, the pilot phase will continue monitoring AI recommendations against real outcomes to catch discrepancies or data integration issues. With strong data governance, encrypted storage, secure transfers and compliance oversight, the project ensures that data quality is not a one-time task but a continues process forming the foundation of a trustworthy AI solution7, 13.