Last year, US banks used real-time machine learning to flag over 90 percent of suspected fraud, yet almost half of chargeback disputes were still managed manually, with files moving between departments. Yet chargeback resolution has moved forward much more slowly. Banks rely on modern machine learning for fraud detection and credit risk, but many dispute systems remain outdated and manual. According to a Pega report, new AI-powered tools, such as Pega Smart Dispute, are helping banks handle transaction disputes and fraud claims more efficiently. In practice, they help teams make clearer decisions and reduce the gap between technical capability and everyday dispute work.
Chargebacks are more complicated than they seem. Each dispute requires careful review of transaction data, understanding what each party did, and applying complex legal standards and card network rules. It goes far beyond paperwork and often requires judgment calls made under strict regulatory rules. Because of this mix of rules and discretion, updating or automating chargeback systems is difficult. Any changes must reflect both changing regulations and the need for individual decisions.
As transactions grow and regulations become stricter, traditional methods are showing their limits. Manual reviews cannot keep up, and paperwork errors can lead to compliance issues. A 2026 article by Vishal Srivastava and Tanmay Sah notes that automation can also pose risks, as system failures might cause larger problems, especially if complex disputes are treated as simple sorting tasks rather than as regulated decisions. System architect Lin Yuan says this misunderstanding is a main reason the industry has not progressed, at the 2025 RegTech Innovation Summit. “They are regulated decisions. Treating them like clerical tasks is exactly why traditional systems keep failing under scale and scrutiny.” According to Chargebacks911, the company developed an AI-powered robotic emulation tool that automates chargeback remediation, prioritizes compliance, and transforms dispute handling into a decision-focused process. Software Copyright Registration: The system incorporates compliance from the outset rather than treating it as a final checkpoint.
Most automation tools rely on a single, integrated system that attempts to manage the entire dispute process within it. Most automation tools use a single system to handle the entire dispute process, from case classification to document creation and decision-making. While this can seem efficient, it often fails in regulated settings because these systems cannot explain or track how they make decisions. This lack of transparency makes it hard to meet regulatory requirements and to justify decisions to auditors. For example, research from major banks shows that manual audits of unclear systems can take up to 40% longer, since compliance officers must rebuild decision trails from scratch.
In contrast, explainable agent-based systems have reduced audit times by up to 30%, enabling institutions to complete reviews faster and focus on higher-risk areas. For compliance teams, shorter audits translate directly into less reconstruction work and clearer accountability. Rather than treating dispute handling as one continuous task, the system breaks it into a sequence of regulated decisions, each guided by its own rules. Her design uses specialized AI agents, each focused on a specific part of the process.
This breakdown mirrors how skilled human analysts work: they collect evidence, interpret rules, build arguments, and check compliance before submitting a case. By separating these steps, the system keeps the process transparent.
Each agent operates under predefined regulatory rules, so decisions are checked as they happen rather than afterward. Instead of checking for compliance only at the end, the system stops non-compliant actions at every step. This approach reduces the risk of costly penalties, such as fines for violating card network rules or other sanctions. By catching problems early, institutions avoid both direct costs, such as fines, and indirect costs, such as additional audits or reputational damage. In regulated environments, system behavior has to be shaped by rules from the outset.
This method differs from most AI systems in finance, where regulatory rules are typically added after the model is built rather than incorporated from the start. Here, multi-agent architecture means using several specialized AI agents, each handling a specific regulatory or operational task, working together to manage and resolve chargeback cases.
The platform uses four specialized AI agents that work together as a team rather than as separate automation tools.
According to Mastercard’s Chargeback Guide, the Classification Agent analyzes dispute situations, selects the most suitable reason codes, and helps determine how cases should be handled moving forward, providing the procedural basis for the chargeback process.
Next, the Evidence Agent collects and reviews supporting documents, consolidating transaction records, communication history, and other relevant data into organized bundles. It does more than gather files; it makes sure all documents meet network standards before moving forward. Using the classification results and verified evidence, and following card-network rules, the agent creates a clear explanation that meets both persuasive and procedural standards. This explanation is not just written text—it is built to comply with compliance rules and ensure validity. The Compliance Validation Agent acts as the system’s final checkpoint. Before anything is submitted, it checks every part against card network rules, including eligibility, complete documents, and timing. Only cases that meet all requirements move forward.
The key innovation is building compliance into the process itself. Instead of being just a checklist, compliance serves as an active control in the system, preventing errors and risks before they occur.
The Intelligent Chargeback Management System is a full-stack application built on a Python Flask backend, incorporating large-language-model reasoning. According to a report by Shuowei Cai, Yansong Ning, and Hao Liu, the language model plays a supporting role by helping to construct and match agents using a pool of large language models. In contrast, the structure and logic of these agents handle the main decision-making. Central to the system’s design is the principle of controllability: the architecture enforces strict governance rules, such as prohibiting self-modifying code, to ensure AI behavior remains predictable and within defined system boundaries. Explicitly defining this constraint gives regulators clearer visibility into how automated decisions remain governed and auditable.
The broader implication is straightforward: intelligence in regulated systems only works when it remains controllable.
A modern React-based frontend adds more control by keeping people in the loop. Analysts can review what agents do, adjust automated decisions if needed, and document reasons for audits. The system does not replace human judgment. Instead, it structures analysts’ work so decisions are easier to review and explain later.
Many commercial chargeback tools automate only specific components of the process, such as classification or document collection, while continuing to rely on human intervention for interpretation and final compliance verification. These discontinuities between process stages frequently result in delays and increased risk. (As noted by Hadden, Yuan’s platform addresses this gap by combining classification, evidence collection, argument building, and compliance checks into a single, seamless workflow.) This way, the system creates a complete, compliance-ready submission as a single decision, not just a set of separate tasks.
Organizations using the system have moved from manual, paperwork-heavy processes to structured decision workflows. The article “AI in Financial Services: Revolutionizing Fraud Detection and Risk Management” reports that AI has helped financial institutions reduce processing times and document errors, enabling fully traceable audits and allowing compliance teams to complete reviews much faster. Institutions report faster processing, fewer document errors, and clearer audit trails.
A major technical challenge was connecting directly with major card networks. Each network uses different data formats and submission rules, so the system had to be flexible enough to handle changing regulatory needs. (Chargeback Guide, 2024)
The solution was to use a modular adapter design. This allows network-specific rules to be updated independently, without changing the entire system, so institutions can stay compliant as rules change.
The platform lets institutions set automation policies based on their risk tolerance. Low-value disputes can be handled automatically, while high-value or cross-border cases may need human approval. (FINBOA Named Repeat Finalist for 2025 US FinTech Awards - Banking Tech of the Year, 2025) This flexibility keeps automation from becoming too rigid and ensures technology supports good decision-making rather than replacing human judgment.
Yuan points out that efficiency was not the main goal.
“The goal was never just to win more disputes,” she says. “It was to show that AI systems can generate decisions that institutions, regulators, and auditors can actually trust.”
Yuan was the only architect and author of the Intelligent Chargeback Management System, leading its development from the initial idea to the working system. She believes what sets the project apart is not just its technical skill, but a new way of thinking.
Compliance, traditionally regarded as a constraint on operations, is redefined as a central design principle that shapes the system’s functioning. The result is a system designed to be reliable in daily operations, not just efficient on paper. The system’s decisions are traceable, explainable, and auditable, thereby aligning technological capability with institutional responsibility. Building on these pillars, a new approach to measuring efficiency is emerging: trust becomes a core metric for technological ROI. For example, institutions might assign a “trust score” to automated systems, quantified by factors such as the explainability of decisions, the auditability of decision trails, and ongoing regulatory compliance. Organizations can then incorporate this score alongside traditional cost and speed metrics to provide a more comprehensive assessment of value. Treating trust as something measurable also changes how organizations discuss ROI discussions and prioritize systems that deliver reliable, transparent, and compliant outcomes over those that merely optimize for speed or cost.
Looking ahead, Yuan plans to use the multi-agent framework for other regulated tasks, like fraud appeals and regulatory reporting, expanding its impact beyond chargebacks. This aligns with the paper’s main point: building reasoning, justification, and compliance into decision-making systems can transform many financial processes. The goal is to create a reusable system design that applies these ideas in different regulatory settings, helping close the gap between new technology and good governance.
As financial technology faces more regulation, approaches like this could mark a bigger shift. Instead of trying to replace people with AI, the focus is now on designing smart systems that work within regulatory rules from the start. (Pervez et al., 2025)
Yuan says this change will shape the next wave of fintech innovation. Future systems will not only act intelligently but also give trustworthy explanations for their actions to institutions, regulators, and customers.
In an industry long focused on speed and automation, Yuan’s work points toward a different standard, one where trustworthiness matters as much as efficiency.
Hadden, R. (n.d.). Disputes and chargebacks white paper.
Journal of Financial Technology. (2023). AI in financial services: Revolutionizing fraud detection and risk management, 12(3), 45–67.
Mastercard. (2024). Chargeback guide.
Pervez, H., Gaurav, S., Heikkonen, J., & Chaudhary, J. (2025). Governance-as-a-service: A multi-agent framework for AI system compliance and policy enforcement. arXiv.
PR Newswire. (2025, October 2). FINBOA named repeat finalist for 2025 US FinTech Awards – Banking Tech of the Year.