Banks are becoming comfortable with automation. Software reviews transactions, flags suspicious activity, routes documents, generates reports and helps employees complete work more efficiently. Those systems support decision-making, but they rarely own it.
Agentic artificial intelligence introduces a different proposition. Rather than assisting an employee, the software can carry out a sequence of tasks on its own. It can gather information from multiple systems, review documentation, complete workflow steps, escalate exceptions and move a process toward completion with limited human involvement.
The evolution has been highlighted by news coming in recent weeks from companies including Catena Labs, Primitive and Saris, which are developing infrastructure intended to support AI agents operating within financial institutions. Their focus extends beyond chat interfaces and productivity tools. The objective is to place software deeper inside processes that have historically relied on human judgment and supervision.
The applications are not difficult to identify. Banks spend enormous amounts of time managing lending documentation, conducting compliance reviews, investigating fraud cases, onboarding customers and handling servicing requests. Many of those activities require employees to gather information from several systems, apply established rules and move work between departments.
Agentic AI promises to reduce that burden. But challenges, and ultimately liabilities, lie with delegation.
Automation allows a bank to accelerate work. Delegation requires a bank to decide which responsibilities can be handed to software and under what conditions.
Advertisement: Scroll to Continue
A loan officer may use software to organize a file. An AI agent could be asked to collect missing documents, validate information, identify inconsistencies, request additional materials and prepare the file for final review. A fraud analyst may use technology to identify suspicious activity. An AI agent could assemble account histories, cross-reference customer records, summarize findings and recommend next steps before a human ever enters the process.
The value proposition becomes apparent when viewed through the lens of productivity. The governance challenge becomes apparent when viewed through the lens of accountability.
Control Becomes the Operating Question
Financial institutions are confronting that question at a time when risk management is already becoming more demanding.
According to PYMNTS Intelligence’s “State of Fraud and Financial Crime in the United States,” 46% of financial institutions report increasing sophistication in fraud schemes. Nearly half of the executives surveyed by PYMNTS Intelligence and Block cite regulatory pressures as a big challenge, while 41% point to pressures associated with faster and more diverse payment systems. Meanwhile, 68% have increased spending on fraud detection capabilities.
Those figures help explain why discussions surrounding agentic AI frequently return to governance rather than capability.
As noted in the announcements from the aforementioned AI firms, and by way of one example, Primitive has emphasized controls, measurement and oversight as central components of deploying AI agents within regulated environments. The emphasis reflects a reality familiar to every bank executive: Operational authority cannot simply be transferred to software without establishing clear boundaries around what the software is permitted to do and how those actions are monitored.
Fraud prevention adds another layer of complexity. PYMNTS Intelligence found that unauthorized-party fraud now accounts for 71% of fraud incidents and losses, driven largely by credential theft and account takeover activity. Institutions also report damage to customer loyalty, reputational harm and lost business opportunities stemming from fraud events.
The embrace of agentic AI trains a spotlight on authority. Which actions can an agent initiate? Which decisions require human approval? How are exceptions handled? How are actions documented? Who is accountable when an agent makes an error that affects a customer, a transaction or a regulatory obligation? Those questions sit at the intersection of technology and governance. Financial institutions have spent years deciding what machines can do. The discussion now turns to what authority they are willing to grant them once they can do it.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
