Marco Montes de Oca is the AI lead at EnFi and a lecturer at Northeastern University.

Banks are pouring billions into artificial intelligence, making AI spending nearly a third of customer experience budgets, according to the Publicis Sapient 2024 Global Banking Benchmark Study. Yet major players like Bank of America, Citigroup, and Goldman Sachs still block access to ChatGPT and other large language models (LLMs).

That caution reveals an important truth. Banks desperately need AI to meet rising customer demands, but standalone LLMs like OpenAI's models have limits when applied to the complexities of lending. When integrated as part of agentic AI systems, however, LLMs can become a pivotal tool in transforming the sector.

The limitations of LLMs

Privacy and security are imperative in a sector as heavily regulated as finance. Samsung's high-profile privacy breach using ChatGPT served as a warning sign for financial institutions. While security concerns have been tightened up by OpenAI and others, LLMs – when used in isolation — have limited ability to meaningfully impact lending workflows.

Having said that, agentic AI architecture often employs multiple LLM calls under the hood, combining their strengths with additional layers of intelligence. Rather than relying on a single LLM to handle all tasks, agentic systems coordinate many interactive LLM calls, integrating them with private datasets and other data streams. This shift from "one call to rule them all" to a dynamic and purpose-built framework is where the real innovation is happening.

LLMs excel at processing public knowledge like general accounting principles or market trends, but they require integration with proprietary datasets, including private company financials, credit memos and internal risk frameworks that inform lending decisions. The gap between publicly accessible training data and the sensitive, context-rich information lenders rely on highlights the need for tailored AI solutions in finance.

There’s also a lack of access to real-time private financial data. Lending decisions rely on live data, as does portfolio management. Once a loan has been issued, there’s a huge risk sitting there, which in an ideal world needs constant monitoring. Periodic compliance reviews that take place currently just aren’t regular enough, making them woefully inadequate.

Lending workflows are also complex. They often require multiple negotiations, document revisions, stakeholder coordination and regulatory compliance actions. When used in isolation, LLMs lack the ability to manage dynamic decision-making and intricate coordination needed for successful lending operations.

Agentic AI advantages

While standalone LLMs can’t revolutionize lending or the banking industry, their role within agentic AI architecture is transformative. Agentic AI architecture goes beyond the static, general-purpose capabilities of LLMs. LLMs excel at processing and generating human-like text based on pre-trained patterns, but they lack the ability to take independent actions, manage workflows or adapt dynamically to specific contexts.

Agentic AI, on the other hand, refers to systems designed to act with purpose and autonomy within defined parameters. These systems can be used to handle complex, multi-step processes by integrating data from multiple sources, dynamically responding to changing information and collaborating with other agents or human operators.

By incorporating “planning” into their workflows, agentic systems can develop step-by-step approaches to problem-solving, identify the tools needed to fetch additional data and even invoke other agents to complete subtasks and report back for consolidation. In the context of lending, this kind of system can enable the monitoring of real-time financial data, flagging of risks, and actionable solutions to bridge the gap between static predictions and active decision-making.

Data, decisions, and dynamism

Agentic AI architecture builds on the foundational capabilities of LLMs, addressing critical gaps that standalone implementations cannot bridge and offering transformative benefits in lending. 

These systems can enable real-time monitoring and proactive risk management and can continuously analyze live financial data including borrowing base certificates, covenant compliance and cash flows. Unlike periodic manual reviews, agentic AI architectures can flag risks and suggest actionable solutions if designed to do so and supported by the appropriate organizational workflows. The big benefit of such an automated workflow is that it lets lenders intervene early to mitigate potential issues. By integrating private datasets and adapting to evolving borrower performance and market conditions, these systems reduce risk and provide tailored lending strategies that go beyond one-size-fits-all solutions.

Equally important is the ability of agentic AI to streamline complex lending workflows. From managing stakeholder coordination to automating repetitive tasks, these systems can ensure efficiency and compliance through seamless orchestration of approvals and documentation. They can enable credit-risk analysts to leverage deeper insights, where human expertise complements AI’s precision and scale.

Surpassing the shortfalls of LLMs

The collaboration between human expertise and AI enables lenders to make more accurate and strategic decisions by combining deep industry knowledge with the efficiency of real-time, data-informed insights. In an environment where traditional methods often fall short, and standalone LLMs face inherent limitations, agentic AI can leverage the capabilities of LLMs in a coordinated, purpose-driven manner.

Moving beyond the general capabilities of large language models, agentic AI provides the precision and adaptability necessary to reshape how loans are evaluated, structured and monitored. Agentic AI, however, is not a plug-and-play solution. Financial institutions must reimagine their workflows to take full advantage of this technology.

As financial institutions look to the future, understanding the synergy between LLMs and agentic AI will be critical. It's not about dismissing LLMs ,but about unlocking their full potential through strategic, multi-layered applications. Those leveraging specialized AI systems tailored to the complexity of lending will redefine success, reshape the industry and reclaim a competitive edge.