Why dispute resolution is still broken
Card fraud losses reached $33.83 billion globally in 2023, and Deloitte projects US banking fraud losses could approach $40 billion by 2027 as AI-enabled fraud tactics grow more sophisticated. Behind every fraudulent transaction sits a dispute case - and behind every dispute case sits a process that most banks haven't fundamentally redesigned in years.
The typical lifecycle looks like this: a customer flags a charge, an operations agent logs it in one system, pulls transaction data from a second, checks fraud signals from a third, requests documentation through a fourth, and manually routes the case to a resolution team. Each step is human-coordinated, each system holds a different version of the truth, and the customer waits. Industry benchmarks put resolution times at 30 days for best-practice banks, with laggards stretching to 120 days or more on complex cases.
The cost compounds fast. Operations headcount scales with dispute volume. Exception handling consumes senior staff time. Regulatory filing creates its own administrative overhead. And every day a dispute sits open, provisional credit exposure accumulates on the bank's balance sheet.
RPA and rules-based automation have made partial inroads - routing simple cases, triggering acknowledgment emails - but they break the moment a case deviates from the expected path. A chargeback dispute involving a third-party processor, a cross-border transaction, and an unresponsive merchant doesn't fit a script. That's where most automation programs hit their ceiling.
What agentic AI does differently
RPA breaks the moment a case deviates from its script. Agentic AI doesn't need a script - it reads the case, assembles what it needs, and decides, all without waiting for a human to move it forward. This distinction matters in dispute resolution, where no two cases are identical and the information needed to reach a decision is scattered across disconnected systems.
In a McKinsey analysis of agentic AI in banking operations, multi-agent architectures coordinate specialized agents - one gathering evidence, one assessing policy compliance, one drafting regulatory filings. These architectures consistently outperform single-system automation on exception-heavy workflows. Dispute resolution is exactly the kind of exception-heavy domain where this architecture delivers.
Where RPA requires a human to handle every case that falls outside a narrow rule set, agentic AI handles the exception itself. It reasons from available evidence, applies policy constraints, and escalates only when the case genuinely requires human judgment. What 120+ bank deployments reveal about agentic AI call centers shows the same pattern: routine cases resolve autonomously, complex cases arrive at human reviewers with full context already assembled.
The end-to-end architecture: five stages, one coordinated system
A well-architected agentic dispute resolution system runs five stages without breaking the workflow into disconnected handoffs.
Intake and classification. The customer submits a dispute through any channel - mobile app, Conversational Banking interface, branch, or contact center. An intake agent classifies the dispute type (fraud, merchant error, authorization issue, duplicate charge) and assigns an urgency level. It then opens a structured case in the Orchestration Layer. The Customer State Graph in Nexus immediately provides full context - account history, previous disputes, transaction patterns, and relationship risk profile - so the agent isn't starting from scratch.
Evidence gathering. A dedicated research agent pulls transaction records from the payments system, retrieves chargeback documentation from the card network, cross-references merchant data, and checks fraud signals from the risk engine. Every data pull and every finding is recorded as part of the case evidence bundle - not siloed in the individual system that produced it. This is the coordination work that falls between systems. It is the work manual processes miss precisely because no single system owns it.
Decisioning. Once evidence is assembled, a policy agent applies the bank's decisioning rules - chargeback eligibility thresholds, provisional credit criteria, regulatory time limits, merchant liability rules. It then generates a resolution recommendation with a full justification trail. For straightforward cases, the system executes the decision autonomously. For edge cases above a defined risk threshold, the case routes to a human reviewer with all evidence pre-packaged and a suggested resolution already drafted. This is Progressive Autonomy in practice: Assistive, Delegated, or Autonomous depending on case complexity and the bank's configured risk appetite.
Regulatory filing and back-office execution. Once a decision is confirmed, a filing agent prepares and submits the required regulatory documentation - Regulation E notices, chargeback codes to card networks, internal compliance records - within the mandated timeframes. Every step generates a Decision Token from Sentinel, the Authority Layer of the AI-native Banking OS that governs every action any actor takes. No filing executes without proof of authorization, policy compliance, and the full evidence chain.
Customer notification and case closure. A communication agent drafts and sends the customer update through their preferred channel - in-app message, email, or SMS - in language the customer can read, referencing the specific dispute and outcome. Resolution confirmation updates the Customer State Graph, closing the case and feeding the outcome back into the Intelligence Layer as a training signal that sharpens future decisioning accuracy.
Compliance isn't an afterthought - it's structural
Dispute resolution operates under strict regulatory constraints. Regulation E requires provisional credit within ten business days on most electronic fund transfer disputes. Visa and Mastercard chargeback rules impose network-specific timeframes. The EU's PSD2 framework adds its own liability and notification requirements. Any agentic system that cuts corners on these mandates creates regulatory exposure faster than it saves operational cost.
The answer isn't to slow AI down - it's to build governance into the execution layer from the start. Deloitte's analysis of agentic AI risks in banking makes the point: banks should treat AI agents as active operators within their systems and design controls accordingly. They should not manage them as passive tools subject to periodic review.
On the AI-native Banking OS, Sentinel runs alongside every layer of the stack. Every agent action - document pull, evidence assessment, provisional credit issuance, regulatory filing - requires a Decision Token before it executes. That token records the policy applied, the actor identity, the model version used, and the full decision context. When a regulator asks how a specific dispute was resolved, the answer is a verifiable audit trail, not a reconstruction from memory.
Human-in-the-loop controls aren't just a governance checkbox - they're an architectural feature. The Forrester assessment of agentic AI in financial services identifies human-agent collaboration models and robust evaluation frameworks as defining characteristics of mature deployments. Banks that configure escalation thresholds - cases above a risk score, disputes above a value threshold, cases involving specific regulatory categories - ensure that autonomous execution operates within defined guardrails. Autonomy is earned, measured, and revocable.
This architecture also addresses the EU AI Act requirements that banks operating in European markets are now managing. Model authorization and explainability requirements are built into the Intelligence Layer - not added as post-deployment patches. Why 120+ bank deployments show AI ROI stalls at the architecture explains the same dynamic across AI initiatives more broadly. The banks achieving production AI outcomes are the ones that built governance into the foundation, not the ones that tried to retrofit it.
What the ROI looks like
The financial case for agentic AI in dispute resolution runs across four dimensions.
Speed and cost-to-serve. Resolution times that averaged 30-120 days on manual workflows compress to hours for straight-through cases and days for complex exceptions. Faster resolution reduces provisional credit exposure and cuts the cost of regulatory penalty risk for missed timeframes. Across Backbase deployments, agentic servicing workflows deliver 30-40% cost-to-serve reductions in operations domains. Staff who previously spent their time queuing and re-keying data across systems shift to handling genuinely complex cases that require judgment - a better use of both their capability and the bank's payroll.
Staff productivity. McKinsey estimates that between 50 and 60 percent of banking FTEs are tied to service operations. Multi-agent systems handle routine dispute cases autonomously, giving those employees back meaningful hours every day. This compounds into 3x productivity gains at the team level.
Customer retention. Customers who receive fast, transparent dispute resolution are measurably more likely to stay. Banks running AI-native end-to-end journeys across their frontline see higher satisfaction scores on operational interactions than banks where disputes disappear into a manual black box for weeks. Jouk Pleiter, Backbase's CEO, frames the customer experience potential this way:
