The performance of AI tools such as ChatGPT and Claude has created enormous expectations for AI agents. Yet when lenders move from impressive demonstrations to real production workflows, the results are often underwhelming.
The problem is usually not the underlying model. It is the gap between general-purpose intelligence and the specialized operational knowledge required to perform a particular lending task reliably. Based on our experience with AI agents in SMB lending, five lessons stand out.
1. It is all about Vertical
General-purpose AI models possess broad knowledge and increasingly impressive reasoning capabilities. What they typically lack is the operational intelligence of an experienced SMB lending team.
That intelligence extends far beyond credit policy. It includes product-specific requirements, customer segmentation, documentation standards, escalation rules, exception handling, regulatory constraints, and the judgment required when a case does not fit neatly into a standard workflow.
Much of this knowledge is not captured in a policy manual. It lives in meeting notes, informal checklists, email threads and, most importantly, in the heads of employees who have spent years navigating the intricacies of SMB lending. An agent cannot reliably apply knowledge it has never been given.
Building an effective vertical agent therefore involves much more than connecting a large language model to a user interface. It requires extracting institutional knowledge, translating that knowledge into usable instructions and workflows, connecting the agent to the appropriate systems and data, and establishing controls for monitoring and exception management.
Before evaluating an AI agent, lenders should ask two questions:
- Has the agent been designed for this specific task, including the necessary tools, workflows, controls and exception paths?
- Does the team building it possess meaningful operational expertise in SMB lending?
If the answer to either question is no, the product may be little more than a general-purpose model with a thin industry wrapper. Its performance is likely to reflect that limitation.
2. The right question is not “What can AI replace?”—it’s “What ROI can AI deliver?”
More than three and a half years after the release of ChatGPT, the capabilities of large language models have advanced considerably. When lenders consider adopting AI agents, they often begin by asking, “Can an AI agent do this?”
That is not necessarily the right question.
Given the right data, tools, operational knowledge and guardrails, the technical answer will often be yes for a wide variety of AI agents for SMB lending such as Opening agents, Document collection agents, Underwriting agents and Closing agents. But technical feasibility is not the same as a compelling business case.
Developing, integrating, monitoring and governing an AI agent requires investment. The more useful question is, “Which agent will generate the greatest return on investment?”
That assessment should consider more than labor savings. Depending on the use case, value may come from faster response times, higher application-completion rates, fewer documentation errors, shorter time to funding, greater operating capacity or improved customer retention. Those benefits must be weighed against implementation and integration costs, ongoing supervision, compliance requirements and the potential cost of errors.
The best starting point is often a high-volume, repeatable workflow with measurable outcomes, or high touch work flow like SBA where borrowers and lenders interact on the average 20+ times. A narrowly defined agent that solves a costly operational problem can create more value than a technically ambitious agent without a clear business case.
3. Perfection is the enemy of progress
Lenders often evaluate AI agents as if they were conventional software. They expect near-perfect performance out of the box and hesitate to deploy an agent until it can handle every possible scenario.
A more useful operating analogy is to think of an AI agent as a digital employee. Human employees are not perfect, yet lenders hire them because they can perform valuable work within defined responsibilities and with appropriate training, supervision and escalation procedures.
This does not mean that agents should be governed like people or held to a lower standard. The lender remains responsible for their outputs. The analogy is useful because it shifts the focus from perfection to effective delegation and oversight.
The relevant question is not whether an agent will ever make a mistake. It is whether the agent can perform a well-defined task at an acceptable level of accuracy and risk—and whether uncertain, unusual or high-consequence cases are routed to the appropriate person.
That operating model requires clear boundaries: what the agent may do autonomously, what requires human approval, when it must escalate, and how its performance will be monitored. It may also require staged deployment, beginning with recommendations or employee assistance before progressing to more autonomous execution.
Human review will remain essential for regulated or high-consequence decisions. But demanding perfection across every possible scenario can prevent lenders from capturing value in the many cases that agents can handle safely and effectively today.
4. Do not focus only on internal operations—the bigger AI opportunity Is customer-facing
Most AI investment in lending has focused on internal processes such as underwriting , AML and fraud detection
These are important applications—but we believe they represent only about 20% of the opportunity.
The other 80% is customer-facing: how lenders acquire, engage, convert and move borrowers through the lending journey. These activities frequently determine whether an applicant completes the journey.
Customer-facing agents can support borrowers around the clock, respond quickly, provide consistent guidance and conduct persistent—but appropriately calibrated—follow-up. They can also summarize conversations, record customer information and update internal systems, reducing administrative work for relationship managers and operations teams.
The goal should not be to remove people from customer relationships. It should be to use agents for routine coordination, communication and information gathering so employees can spend more time on situations that require judgment, empathy, negotiation or relationship development.
For many lenders, this is where some of the most immediate value may be found. Faster and more consistent customer engagement can improve both operating efficiency and the borrower experience.
5. Compliance isn’t a feature—it’s the foundation
Compliance is fundamental to SMB lending. It cannot be added to an AI agent after the agent has been designed; it must shape the architecture from the beginning.
Unlike agents used in many other industries, SMB lending agents operate in a highly regulated environment and may interact directly with borrowers, sensitive data and consequential workflows. They should therefore be designed around four principles:
- Controlled and predictable. Critical business outcomes should not depend on unconstrained model responses. Deterministic workflows, approved tools, explicit business rules and human-approval requirements should govern consequential actions. Generative AI should operate only within clearly defined boundaries.
- Auditable. Relevant customer interactions, agent responses, data sources, tool calls, actions, escalations and human overrides should be recorded and retained in accordance with the lender’s policies. The lender must be able to reconstruct what happened and understand why.
- Grounded. Customer-facing responses should be based on approved, current sources rather than the model’s general knowledge. When the available information is insufficient or uncertain, the agent should ask for clarification, decline to answer or escalate the matter to a qualified employee. Fully grounded AI agents will eliminate hallucinations, which is a significant problem in using Generative AI.
- Focused. An agent should respond only within the scope of the task it has been authorized to perform. It should politely decline unrelated requests—such as attempts to discuss political events—and resist efforts to redirect it beyond its approved role. Narrow scope and limited system access are critical to containing what could otherwise become an unbounded compliance risk.
Compliance playbooks for AI agents in SMB lending are still emerging. One useful starting point is to treat an agent as a digital employee: it should be subject to the same policies, supervision, quality controls and escalation requirements as a human employee performing the same task.
The analogy should not stop there, however. Because agents can operate continuously and at far greater scale, they also require controls designed specifically for AI, including access restrictions, version management, systematic testing and ongoing performance monitoring. Automation should never lower the compliance standard applied to the underlying activity.
The takeaway
Our conviction is simple:
The future of SMB lending will not be won by the lender with the most broker relationships, the largest marketing budget, or the best decisioning or loan management system. It will be won by the lender that turns AI into measurable execution—and measurable business outcomes.