A new wave of agentic AI systems is reshaping banking operations. Unlike typical large language model (LLM) applications that answer prompts, agentic systems execute sequences of actions: querying systems, retrieving documents, transforming data, and producing outputs. Quietly, these autonomous tools are beginning to redefine the banking technology landscape.
The potential impact is sufficiently profound that McKinsey is now framing agentic AI as a structural shift in banking rather than a side bet; the consultant estimates that AI adoption—including agentic AI systems—could reduce banks’ aggregate cost base by 15% to 20%. Bain, in its 2025 report, “State of the Art of Agentic AI Transformation Technology Report,” cites that in the first half of 2025, “tech-forward enterprises” turned their focus from automating tasks to redesigning entire workflows, as early adopters get to grips with how agents—or the AI systems that independently handle multi-step tasks by coordinating tools, data and actions to meet specified objectives—may coexist safely and collaborate productively. Yet progress is limited.
Although agentic AI may hold promise, definitional confusion and implementation hurdles mean very few true use cases exist, cautions Armand Angeli, AI and automation specialist and vice president, Digital Transformation and AI Group, at DFCG, the French network of CFOs.
“Financial institutions still struggle to understand and implement agentic AI properly,” he says, “and are jumping too fast into these new tools without addressing the fundamentals of data quality, clear processes, skillsets, and ROI [return-on-investment]. There’s a high degree of confusion about what agentic AI is, with people equating AI assistants or RPA [robotic process automation] with true agents. Only a very small number are actually building and scaling agentic effectively.”
Angeli also contends that people overuse the word “agentic.”
“GenAI is mistaken for agentic because it seems intelligent or retrieves data,” he says. “But GenAI is relatively simple and doesn’t self-correct, unlike agents with memory and feedback loops for auto-healing and learning. Building these agents requires mapping complex processes and understanding where the data is, which can take months and thousands of euros in costs. It’s a fine line between a simple agent or RPA and true agentic AI.”
Even though the tools themselves are complex, their appeal is straightforward and powerful.
Where Agentic AI Is Actually Being Deployed
Whether LLM-powered information retrieval agents, single-task agentic workflows, cross-system agentic workflow orchestration, or multi-agent constellations, true agentic AI can perform complex tasks independently within defined boundaries, all with limited human intervention.
BBVA Peru’s Blue Buddy agentic AI assistant is an example. The “lightning-fast knowledge synthesizer” autonomously navigates the commercial bank’s vast ecosystem of unstructured data—product manuals, regulations, and complex processes—to deliver precise, contextualized answers in real time and in a risk managed way.
“We’re not just exploring AI; we’re putting it to work on the front lines of our business,” says Benjamín Chávez, head of engineering at BBVA Peru.
UK-based consultant Capco recently deployed an agentic AI assistant at a global investment bank to support junior bankers in producing credit memos, company profiles, and peer benchmarks.
“Previously, analysts could spend five to ten hours a week on a single memo, largely on manual data gathering, formatting, and rewriting,” says Charlotte Byrne, Capco’s UK GenAI lead. “The new workflow allows a banker to request, for example, ‘Draft a credit memo for a corporate client with the latest financials and peers.’ The agent delivers a first draft within minutes.”
The client bank ultimately saw a 50% reduction “in time spent on the mechanical parts of the process.”
Wells Fargo recently announced a collaboration with Google Cloud that will deploy agentic AI at scale via 2,000 employees, with further plans for bank-wide rollout. The tools Google Cloud will supply synthesize information, automate workflows, and boost agility; key applications include triaging foreign exchange post-trade inquiries and navigating guidelines in corporate and investment banking. In Greece, Eurobank is working with EY to develop a scalable, automated system that embeds agentic AI into core banking operations.
In each case, the goal is to replace high-volume, repetitive workflows. But implementation is not without its challenges.
During Capco’s recent rollout, while AI algorithms themselves did not present an issue, the client bank’s internal requirements complicated the process. “We had to use guard-railed, bank-approved models,” says Byrne, “which meant investing heavily in prompt design, retrieval quality, and validation. Governance also added long lead times; simply getting proof-of-concept approvals took nearly two months, by which point the model landscape had already shifted again.”
Engagement was another challenge. Asking already stretched teams to dedicate extra hours to testing is often one of the practical challenges of implementing agentic AI, and adoption suffers if solutions are built too far from the day-to-day workflow. And while banks see the potential of autonomous agents, Byrne observes, few currently have the infrastructure to use them effectively and safely, with poor data and legacy systems the key obstacles.
“Most AI failures in banking have nothing to do with the models themselves,” she says; many banks still lack clean APIs into core systems or struggle with slow, fragmented approval cycles that are incompatible with iterative AI development.
Scaling The Challenge
Scaling GenAI from “lab to regulated banking environment” is no small feat, BBVA’s Chávez concedes. Operationally, BBVA’s major challenge was transforming vast amounts of unstructured data into a clean, corporate-grade knowledge base.
“We had to implement rigorous data governance to ensure the agent’s ‘brain’ was fueled only with accurate, up-to-date information,” he notes.
And while agentic AI has generated significant enthusiasm, there are, as yet, only isolated examples of success, and tangible value across financial services remains limited. Ambiguous strategic objectives, organizational complexity, and the challenge of replicating interpersonal dynamics represent critical barriers, says Chang Li, chief manager, Nippon Life Insurance Company, director of the Fintech Association of Japan, and ambassador for FinCity.Tokyo.
“First, we must understand what we’re looking to achieve, whether that’s better customer communication or cost cutting,” she says. But defining strategy and purpose is difficult for any one division alone; it requires collaboration between departments, Li notes, since bureaucratic structures often prevent meaningful conversations between the correct stakeholders.
Are there concerns about agentic AI taking over from humans in some finance functions? That may no longer be the right question, Li says: “I think it’s more useful to think about the conditions under which the first human ‘channel’ might be taken over by AI and consider how companies should prepare for that.”
The necessary degree of trust is not yet in place for agentic AI to truly replace humans in banking, however. “Currently, agentic AI is only feasible for the information collection step,” says Li, with an agentic contract still “a few years” off.
For BBVA, building trust into agentic AI systems is foundational. “In the financial sector, trust is our most valuable currency,” says Chávez. The bank proactively aligns with demanding emerging standards, including frameworks from Europe and the US, in addition to Peruvian regulations.
“This ethical stance has directly shaped our strategic roadmap,” he notes. “We’ve prioritized decision support use cases over autonomous decision-making. We started where AI assists and humans validate. It’s the most responsible way to deliver immediate value while mitigating risks and building the trust needed for deeper automation.”
In an era of falling revenues, financial institutions may find the productivity gains they need from agentic AI, McKinsey suggests, predicting that early adopters will secure a lasting advantage over slow movers: but not overnight.
McKinsey anticipates a breakout agentic business model will emerge in the next three to five years and is urging bank executives to focus on a small number of high‑value workflows, such as frontline sales, account planning, and financial close processing; define clear guardrails for agent autonomy; and invest early in data quality and risk controls to ensure pilots can scale safely: all with “surgical precision” in identifying the potential earnings impact.
The post AI In Finance: The Power Of Agency appeared first on Global Finance Magazine.