Teaching Compliance With AI Still Requires Human Judgment

Banks looking at AI for training and development should not mistake scale for autonomy. In a recent ProSight article, Steven Ramirez, CEO of financial services consultancy Beyond the Arc, says the case for AI in compliance learning is not that it can replace people—it is that, if built and governed carefully, it can help institutions create and update training more efficiently while keeping human oversight firmly in place. 

That distinction matters. Ramirez’s argument is that AI belongs inside existing risk management and compliance frameworks, not outside them. The same standards that govern accuracy, documentation, escalation, and review still apply. In some ways, they become even more important. 

Here are a few takeaways: 

Keep humans in the loop—and around the loop. AI for coursework and training materials still requires people to train the system, supervise outputs, and step in when needed. Ramirez describes several layers of oversight, including “human-in-the-loop” for complex escalations, “human-on-the-loop” for a less hands-on monitoring role, and “human-out-of-the-loop” for downstream audits. The message: AI can support learning, but governance does not disappear. 

Train the model on the right sources. One of the clearest practical points is that bank-specific AI needs bank-specific inputs. Training requirements can include sources such as industry bulletins, legal memos, approved archived compliance guidance, and prior training manuals. Teams are also encouraged to use structured instructional design, validated scenarios, and subject matter expert review to improve the quality of AI training and implementation. 

Build regulatory nuance into the system. Ramirez’s approach is not just about feeding more information into a model. It is about feeding the right kinds of information into it. Data inputs must account for state-level regulatory differences, the status of federal proposed legislation, approvals, final rulemaking, executive orders, and legal challenges. He also notes that banks may need models aligned not only with evolving regulations, but with internal policies that go beyond supervisory requirements. 

Accuracy and trust need active management. The risks are familiar: data bias, human bias, hallucinations, and outdated information. The recommended response is equally clear: structured review, validation, explainability, logging, source verification, and ongoing training education. 

The broader takeaway: AI can help compliance learning move faster—but only if institutions are equally serious about control, transparency, and human judgment.