Swapnil Awasthi, a data scientist and former Barclays and Citibank fraud analytics leader, is helping redefine how AI fights financial crime while promoting fairness. “AI gives banks powerful tools to find early warning signs of fraud,” he says, “but without clear oversight, it can also carry forward unfair biases from past data.” His work—spanning credit risk, fraud detection, and explainable AI—illustrates how financial institutions can navigate the Banker’s Dilemma: harnessing intelligent systems to reduce fraud and expand financial inclusion without sacrificing transparency or trust.
The financial industry is navigating a high stakes balancing act, adopting artificial intelligence to combat increasingly sophisticated fraud while simultaneously working to ensure these systems promote fairness and financial inclusion. This challenge, often called the “Banker’s Dilemma,” requires a deep understanding of both technology and risk management. With over a decade of experience at institutions like Barclays, Citibank, and the AI platform DataRobot, Swapnil Awasthi has been at the forefront of this evolution.
Awasthi’s journey from a corporate developer at IBM to a data scientist specializing in credit risk and fraud analytics provides a unique perspective on AI’s dual role. His work, which involves managing portfolios with billions in exposure, highlights the critical need for frameworks that harness AI’s power responsibly. As financial services become more automated, the principles guiding the deployment of these complex models are more important than ever.
The transition to AI-driven finance began with projects that revealed both its immense potential and its inherent risks. For Awasthi, an early project with a major U.S. bank illuminated how technology could fundamentally reshape lending and risk assessment.
He notes, “AI gives banks powerful tools to find early warning signs of fraud, uncover patterns in customer behavior, and improve how they decide who qualifies for loans or other services.” This can lead to significant reductions in financial losses and help identify creditworthy individuals overlooked by traditional methods.
However, the power of these systems comes with significant responsibility. Without proper governance, the same tools can introduce new problems, from opaque decision-making to the amplification of historical biases.
“Without clear rules and oversight, the same technology can create problems,” Awasthi states. “It may make decisions that are difficult to explain to customers, carry forward unfair biases from past data, and, if used at scale without safeguards, contribute to larger risks that affect the whole financial system.”
This reality underscores the challenge of modernizing legacy insurance systems, where the goal is to unlock new capabilities without inheriting old flaws. A successful transformation can yield a significant reduction in IT costs and improve operational efficiency.
The fight against financial fraud has become a technological arms race, with institutions constantly adapting to the evolving tactics of fraudulent actors. From his front-line roles at Barclays and Citibank, Awasthi has witnessed this dynamic firsthand. The cat-and-mouse game involves deploying advanced AI to detect anomalies while fraudsters devise new ways to appear legitimate.
“As banks introduced more advanced systems to spot unusual spending patterns and connections between accounts, fraud groups adapted by using synthetic identities, setting up ‘mule’ accounts, and creating complex spending and payment flows designed to look legitimate,” he explains. This is a challenge Awasthi has analyzed in detail; his recent scholarly article on Synthetic Identity Fraud in U.S. Payments quantifies the economic costs of this threat and explores the effectiveness of graph-based detection models.
Awasthi adds, “In practice, that meant my teams paired real-time scoring and updated risk signals with rigorous checks on model performance and clear playbooks for escalating suspicious cases — the exact operational safeguards regulators now expect in credit card lending.” This adaptive strategy is crucial, as collaborative approaches like Horizontal Federated Learning are now used to detect common fraud patterns across different banks. Studies show these consortia can achieve a notable fraud detection improvement while reducing false positives.
At the heart of modern finance lies a fundamental tension between technological innovation and ethical responsibility. Building AI models sharp enough to catch sophisticated fraud without unfairly penalizing individuals based on biased historical data is the central challenge. Awasthi frames this as a core consideration that must be addressed from the very beginning of the model development process.
“The main challenge for banks, often called the ‘Banker’s Dilemma,’ is finding the right balance between using new technology to fight fraud and making sure it treats people fairly,” he says. An effective strategy involves embedding fairness into the system’s design rather than treating it as an afterthought.
“I focus on fairness from the start when creating these systems. This includes testing to ensure the system treats everyone fairly, requiring human review for important decisions, and keeping clear records of how choices are made,” Awasthi explains.
This approach aligns with research into the inherent trade-off between fairness and accuracy. Techniques like the Quasi-Pareto Improvement approach are being developed to enhance model fairness for underrepresented subgroups without degrading overall performance.
Beyond fraud detection, AI holds significant promise for advancing financial inclusion by looking beyond traditional credit metrics. Many individuals remain outside the formal financial system due to a lack of conventional credit history, even if they have a consistent record of paying rent or utility bills. Awasthi emphasizes that AI can help bridge this gap by responsibly incorporating alternative data into credit risk assessments.
“Across the country, millions of people are still left out of the financial system because they lack traditional credit histories,” he notes. By analyzing data such as on-time rental payments or steady cash flows, lenders can identify creditworthy applicants who were previously invisible.
“This shows that smarter use of data can open doors for more families while keeping lending safe and sound.” Ensuring these models are equitable requires rigorous evaluation, moving beyond simple metrics to comprehensive benchmark suites that analyze trade-offs. The choice of fairness metric is critical, as a comparative assessment of fairness definitions has shown that some measures are more effective at detecting bias than others.
Developing quantitative risk frameworks for portfolios valued in the billions requires a multi-faceted conversation among business partners, senior management, and risk experts. The dialogue must balance regulatory compliance, customer outcomes, and business objectives. Awasthi highlights that this process is essential for ensuring that risk mitigation strategies do not inadvertently create barriers to equitable access.
“First, regulatory compliance and model oversight are essential,” Awasthi states. “We need to follow strict rules from regulators like the Federal Reserve and the OCC to validate and monitor our models and strategies.”
This is balanced against the impact on consumers and the institution’s goals. “Third, there is a trade-off between business value and public trust. Stricter controls can reduce fraud but may also limit revenue and customer access.”
Navigating this requires sophisticated data architectures that can integrate information from different sources, a challenge addressed by mainframe modernization with data streaming. For real-time analytics, processes like ELT (Extract, Load, Transform) are better suited than traditional methods, as they leverage cloud resources for high-speed processing.
As AI models grow more complex, the ability to explain their decisions to regulators, business teams, and customers becomes paramount for building and maintaining trust. Awasthi champions a three-part approach to explainability, ensuring that transparency is tailored to the needs of each audience. This structured approach is crucial for demonstrating fairness and compliance in automated systems.
For regulators, the focus is on deep technical documentation. “I provide detailed technical explanations like audit trails, model summaries, and reports showing how different factors affect decisions, and validation documents that meet regulatory standards,” he says.
For internal business teams, operational dashboards provide insights into model performance, while customers receive clear, simple reasons for decisions affecting them. “By building explainability into every step, I help ensure AI decisions are fair, reliable, and meet national priorities for responsible AI use,” Awasthi concludes. His research on “Adverse Action-Ready Explainable AI” further explores this, focusing on how to provide a predictive lift in underwriting while maintaining the regulatory specificity required for consumer protection.
This is supported by Explainable AI (XAI) tools like SHAP and LIME, which are used in modern fraud detection frameworks. An evaluation of explainability frameworks shows that while SHAP provides more stable explanations, LIME offers faster insights, highlighting a key trade-off.
Looking ahead, several key advancements in data and modeling techniques are poised to create a more resilient and inclusive financial system. Awasthi points to federated learning, graph AI, and open banking as transformative technologies that can enhance credit underwriting while protecting consumer privacy and reducing systemic risk.
“One is federated learning, a privacy-focused way for institutions to improve their models together without sharing personal data,” he explains. This collaborative approach aligns with national innovation goals. Another key area is the use of real-time alternative data.
“Finally, real-time alternative data from sources like open banking and payroll APIs, combined with strict consent and privacy safeguards, allows lenders to fairly assess borrowers with limited credit history.” While federated models have shown performance comparable to centralized ones, they require robust governance. Integrating federated learning with blockchain can enhance auditability for high-exposure applications like fraud detection.
Managing massive financial portfolios at scale offers a profound lesson in responsibility. Every data-driven decision has tangible consequences for individuals and the broader economy. Awasthi emphasizes that at this level, AI is not just a tool for optimizing profit; it is a force that shapes access to essential services and must adhere to stringent public policy and civil rights standards.
“Every data-driven decision carries real public impact,” he reflects. “At this scale, AI doesn’t just affect profit; it shapes who can access essential services and must meet strict rules around supervision, civil rights, and public policy.”
This responsibility demands robust technical controls and unwavering transparency. “Handling such large portfolios showed me that AI decisions are more than just numbers; they have human consequences.” This underscores the need for a comprehensive AI audit methodology and awareness of security threats, as outlined in surveys of federated learning attacks and defenses.
Navigating the Banker’s Dilemma in the AI era requires a sophisticated blend of technological expertise, ethical governance, and a steadfast commitment to fairness. By embedding transparency and accountability into the core of financial systems, leaders can work toward a future that is not only more secure against fraud but also more inclusive for everyone.
Swapnil Awasthi is a risk strategist with over a decade of experience in financial services and technology, specializing in AI-driven risk management, credit risk, and fraud analytics. His work focuses on building fair, transparent, and effective systems for major financial institutions. You can learn more about his work on LinkedIn or ORCID.
The expansion introduces agentic decision flows to bolster investigative capabilities for a national security agency in the EMEA region.
Zoho founder Sridhar Vembu warns that companies must prioritize data sovereignty by building internal AI models to protect sensitive intellectual property.
Cities are deploying artificial intelligence on sanitation vehicles to monitor private property, raising complex questions about the future of urban privacy and municipal oversight.
