September 25, 2026 — In the rapidly evolving landscape of fintech, a new frontier has emerged: Agentic AI. While generative AI dominated the conversation in previous years with its ability to draft emails and summarize reports, the financial services sector is now pivoting toward "agents"—autonomous systems capable of executing complex, multi-step workflows. However, as the industry moves from pilot programs to enterprise-scale deployment, a critical question remains: Is the current enthusiasm matched by tangible, risk-managed impact?
GlobalData recently convened an expert roundtable, "Agentic AI in Banking & Financial Services: Beyond the Hype," to dissect this shift. Featuring insights from Jeff Veis, Chief Marketing Officer at Impetus Technologies; Deepak Khosla, Chief Growth Officer & Head of AI Business at Impetus; and Stephen Walker, Retail Banking Analyst at GlobalData, the session provided a sobering yet optimistic look at the realities of AI integration in highly regulated environments.
The Main Facts: Defining the Agentic Shift
The transition from "Chatbot" to "Agent" represents a fundamental change in how banks leverage machine intelligence. Unlike traditional AI models that respond to passive queries, Agentic AI is designed to act. It can navigate software ecosystems, initiate transactions, and make autonomous decisions based on pre-defined objectives.
In the banking context, this means moving beyond simple document summarization. An agent might be tasked with the end-to-end management of a loan application, real-time fraud intervention, or personalized financial advisory services that adjust to market volatility. However, the roundtable participants agreed that the leap from "cool experiment" to "production-grade system" is fraught with architectural hurdles that many institutions are only just beginning to map out.
Chronology: The Evolution of AI in Finance
To understand why the industry is currently at an inflection point, it is necessary to look at the timeline of AI adoption in the financial sector:
- 2022–2023 (The Generative Infancy): Financial institutions experimented with Large Language Models (LLMs) primarily for marketing copy, basic customer service chatbots, and internal knowledge management.
- 2024 (The Sandbox Phase): Banks began building internal sandboxes, testing AI’s ability to analyze unstructured data and support decision-making for human employees.
- 2025 (The Governance Awakening): As AI models hallucinated or failed to adhere to strict financial compliance, firms pivoted toward "Human-in-the-loop" systems, slowing down widespread deployment.
- 2026 (The Agentic Realization): The current year marks the shift toward autonomous agents. The industry has realized that successful implementation requires a robust, data-centric infrastructure—a shift from "model-first" to "data-first" development.
Supporting Data: Why Scaling Remains Elusive
Despite the buzz, a significant "deployment gap" exists. While over 80% of top-tier financial institutions report having active AI initiatives, fewer than 15% have successfully moved agentic workflows into full-scale production.
The roundtable identified five critical friction points that prevent scaling:
- Fragmented Enterprise Context: Bank data is often siloed in legacy systems (mainframes vs. cloud), making it difficult for an agent to gain a "holistic view" of a customer.
- Governance Gaps: Standard AI models lack the rigorous audit trails required for financial regulatory compliance (e.g., Basel III, CCPA, GDPR).
- Safety Risks: Without strict guardrails, autonomous agents could inadvertently approve unauthorized credit or provide non-compliant financial advice.
- Operational Realities: AI models often struggle to understand the nuances of specific institutional business logic that hasn’t been codified into digital form.
- Regulatory Expectations: Regulators are increasingly demanding that banks explain the "reasoning" behind an AI’s decision—a hurdle known as the "Explainability Problem."
Official Responses: The Impetus Perspective
During the roundtable, Deepak Khosla offered a sobering perspective on the stakes involved in this technological shift. "An agent in banking is not just summarizing a document," Khosla noted. "It could influence and impact credit, fraud, payments, customer treatment, reporting, or advice. The bar for production is therefore much higher in the banking and financial services sector."
Khosla’s argument centers on the concept of "Context Engineering." He posits that AI agents fail not because the underlying models (like GPT-4 or Claude) are insufficient, but because they lack the proprietary "context" of the bank.
"If that context—your business processes, your historical interactions, your risk thresholds—is fragmented, stale, or poorly governed, the agents will fail in production," Khosla explained. He emphasized that the winning banks will be those that prioritize building a "semantic layer" that translates complex, messy enterprise data into a structured format that AI can reason over.
Implications: The Road Ahead for Financial Institutions
The roundtable concluded with a clear roadmap for financial leaders looking to operationalize Agentic AI. The transition from hype to reality will likely hinge on three strategic pillars:
1. The Construction of Semantic Layers and Knowledge Graphs
Rather than simply feeding raw data into an LLM, banks must build knowledge graphs that map the relationships between customers, assets, risks, and regulations. This creates a "source of truth" that agents can reference, ensuring that their autonomous actions are grounded in current institutional facts rather than hallucinations.
2. Prioritizing High-Impact, Manageable Use Cases
Institutions should avoid "boil the ocean" approaches. Instead, they should focus on use cases where the ROI is measurable and the risk is contained. For example, using agents to automate the reconciliation of cross-border payments is a high-value, high-complexity task where the rules are rigid enough for AI to master, but the cost of failure is manageable through oversight.
3. Redefining Governance as "Continuous Compliance"
Traditional governance is periodic; AI governance must be continuous. Financial institutions are moving toward real-time monitoring of agentic output. This involves building "AI Observability" platforms that can detect when an agent is veering outside of its defined policy constraints and intervene before a transaction is finalized.
4. The Human-Agent Partnership
The ultimate implication of this technological shift is not the displacement of human bankers, but the transformation of their roles. In a post-agentic world, the banker becomes an "orchestrator." They will define the goals for the agents, audit their reasoning processes, and step in for the high-empathy, high-complexity interactions that define modern private banking and wealth management.
Conclusion
The transition to agentic AI in banking is no longer a question of "if," but "how." The industry has moved past the initial infatuation with AI’s creative writing capabilities and is now grappling with the difficult, unglamorous work of data architecture and governance.
As Stephen Walker of GlobalData noted during the session, the institutions that succeed will be those that recognize that Agentic AI is an infrastructure play, not just a software play. By investing in the underlying semantic layers and grounding their AI in the reality of their enterprise context, banks can move beyond the hype and begin to harness the true potential of autonomous systems.
For the financial sector, the message is clear: The agents are coming, but they will only be as smart, compliant, and effective as the data foundations they are built upon. As the industry looks toward the next 24 months, the winners will be determined by who can most effectively bridge the gap between technical potential and institutional trust.
For further exploration of the strategies discussed in this roundtable, including a deep dive into the technical implementation of knowledge graphs in banking, viewers can watch the full session [here].
