The story usually told is that agentic AI replaces robotic process automation (RPA), but the data doesn’t really support that. In the 2025 EY-Parthenon Generative AI in Banking survey, around 100 leading retail and commercial banks were asked about their automation estates, and only 28% of the automation use cases in development or production use generative or agentic AI at all. Traditional machine learning and RPA account for the other 72%, and they will continue to do so for a while.  

Banks that automated their reconciliation didn’t make a bad decision. The bots still run and are still saving the hours they were built to save. What changed isn’t the value of the bots. It is everything that developed around them, the exceptions they can’t handle, the human-in-the-loop processes they create, and the increasing regulatory scrutiny around the decisions they support.  

The question was never RPA or agentic AI. It is which one the process actually needs. RPA was built for predictable, rule-based work. Agentic AI is built to handle work that requires interpretation and judgment. Understanding which processes sit on which side of that line is the real exercise, and getting it wrong can put a program on the cancellation list.  

 What RPA Was Built For, and Where That Stops  

RPA does one thing extremely well. Give it an interface that doesn’t change, an input it can predict, and a rule that holds, and it will run the same sequence forever without getting tired or sloppy. EY puts RPA and traditional ML exactly where they belong, doing standardized, repetitive back-office work with low complexity and variation, such as data aggregation or transaction reconciliation.  

 The ceiling appears when any of those three conditions break. An input arrives as a scanned PDF instead of a structured file. A counterparty renames a field. A case turns up that the rule doesn’t cover. RPA has one answer to it all, which is to stop and hand the case to a person.  

That is where exceptions become queues and interface changes become maintenance tickets. Banks that scaled their bot estates hard found something uncomfortable waiting for them. The bigger the estate, the more there is to maintain, and the more of the team’s week goes into keeping it running.  

None of this is a failure of the technology. RPA is doing exactly what it was designed to do. The problem is that much of the work still sitting on a banking operations floor does not meet those three conditions.  

  The Shift Is From Executing Tasks to Owning Outcomes  

That is where the distinction between RPA and agentic AI becomes useful. Rule-shaped work stays with RPA. Judgment-shaped work is where agentic AI belongs.  

RPA is handed a procedure. An agentic system is given an objective, a set of constraints, and the freedom to decide how to reach that objective within those constraints. When it meets an input, it hasn’t seen before, it doesn’t simply stop. It interprets. When it needs information that the process never anticipated, it goes and finds it. When confidence drops, it escalates even though the case is already built, rather than dropping back to zero.   

EY frames the fit in similar terms. Agentic AI suits decision-intensive, complex work where autonomous decision-making can make the process more responsive, such as loan approvals, fraud detection, or dispute management.  

These processes contain judgment inside a workflow. That is the point where a fixed sequence of rules starts to run out of room.  

What This Looks Like in Practice  

Take a cardholder dispute. Under an RPA design, bots handle the mechanical parts, pulling the transaction record, populating the network form, and sending the acknowledgement. The work that requires interpretation still lands on an analyst.  

Under an agentic design, specialized agents own segments of the lifecycle and share state across it. An intake agent classifies the dispute and maps it to the right reason code. An evidence agent pulls together what is needed from the transaction history, merchant records, and prior case outcomes. A decisioning agent weighs recoverable value against the cost and probability of pursuing it and prioritizes accordingly. A compliance agent tests the emerging position against network rules and the clock before anything is filed.  

The difference is not that the agents automate more tasks. It is that they can carry the decision forward. Because every agent works from the same case, the context assembled at intake remains when the case reaches representment. The compliance agent doesn’t re-read the file. It tests a recommendation whose reasoning is already visible in the case.  

That is the shift from executing individual tasks to owning an outcome across the workflow.  

Why More Than 40% of These Projects Will Be Canceled  

That shift also explains why choosing agentic AI does not guarantee success. The technology must match the work, and the system has to be designed around the decisions it will make.  

Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, drawing on a January 2025 poll of more than 3,400 organizations. Gartner attributes this to escalating costs, unclear business value, and inadequate risk controls. These are failures of the overall implementation, rather than simply a question of whether the underlying model can generate an answer.  

Gartner also flags agent washing, in which existing assistants and chatbots are given new labels and rebranded as agents without meaningful autonomy. The underlying lesson is simple. Calling a process agentic does not make it capable of owning a decision. 

For banks, the bar sits higher again. EY found that regulatory compliance (71%) and data privacy (67%) were named as barriers to agentic adoption among larger institutions. Only 16% of GenAI use cases reach full deployment, and 40% of those implemented fail to meet expectations. Asked what they would change if they were starting their program over, 79% said they would prioritize governance first, ahead of every other factor on the list.  

Those numbers point back to the same requirement. An autonomous decision in banking has to be explainable to a regulator, attributable to a policy, and reversible. Those requirements have to be built into the orchestration layer, because the system needs to know when it can act, when it needs more information, and when a person needs to take over.  

How Opus Approaches This  

Opus Technologies builds agentic orchestration around the specifics of the lifecycle it operates in rather than around a general automation pattern.  

Intelligent Dispute Management is one worked example. Rather than layering more automation onto dispute handling, it coordinates specialized agents across the full lifecycle: intake and classification, evidence aggregation, decisioning and prioritization, representment and documentation, SLA and compliance monitoring, and a learning agent that improves decisioning based on realized outcomes. The orchestration maintains the context that would otherwise be lost between intake and filing, and keeps every submission aligned with network rules and timelines.   

That is where domain-native engineering becomes important. Knowing what the data actually means, where the edge cases sit, and which decisions can run unattended determines whether an agentic design can survive contact with production.  

What Changes When the Work Is Owned Rather Than Executed  

The old scorecard for an automation program counts two things. How many processes were automated, and how many hours were saved from them?   

An agentic design changes the question. What matters is how many decisions the institution can reach without reassembling context from scratch, and how many of those decisions it can defend afterward.  

That changes how financial institutions should think about automation. The goal is not to replace every bot with an agent. It is to match the technology to the work. Keep RPA where rules and repetition are the right fit. Use agentic AI where the work requires interpretation, coordination, and judgment. Then build the controls needed for those decisions to run safely.  

That reconciliation bot from 2019 is still earning its keep, because nobody ever asked it to make a judgment call.  

References  

• EY-Parthenon Generative AI in Banking survey (2025 edition, 100 retail and commercial banks)  

• Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (Gartner, June 2025)