.png)
Regulators broadly agree on one thing when it comes to generative AI in finance: Humans should be, need to be, in the loop. Models make mistakes. Someone needs to catch them.
The human-in-the-loop requirement presumes that finance professionals are focused enough to spot AI-generated mistakes. Increasingly, they aren't. They're toggling between seven windows, glancing over AI outputs before pasting them into another application.
Generative AI was supposed to free up human judgment for more critical work in automating the mundane, eye-glazing-over tasks like data entry. But by piling on a new attention-demanding workload—overseeing and reviewing the AI's output and shuttling it between applications—AI compromises the very judgment it promised to reserve.
The human-in-the-loop rule assumes time and attention AI has consumed
Governing bodies like FINRA, the European Commission and the CFPB mandate that finance professionals exercise the necessary authority and time to supervise, explain and correct AI-generated content. They reason that human review, even if constrained, still provides more accountability than a fully automated process.
That may still be true. But the loan officer who, in his five minutes between calls, rubber-stamps the AI-drafted loan narrative without checking if he's entered the wrong customer's data into the LLM is exercising neither time nor authority.
Part of the problem is that finance professionals are under pressure to perform at an even higher standard now that they've been handed AI tools—the same tools adding to their workloads.
The majority of finance professionals I've spoken with recently have said productivity expectations at their organizations have increased. And about half reported that they're handling a higher volume of tasks because they're now reviewing the AI's output and decisions.
Then there's the impact on their cognitive loads. Finance professionals are more apt to make mistakes when they're processing too many pieces of information at once. In talking with finance professionals about how AI has changed their day-to-day work lives, many described spending more time distracted by alerts, searching for information and re-entering data between systems. They're worried about slipping up because they're moving too fast; worried about the downstream impact of their mistakes on customers and the wider public.
Finance workstreams required immense mental energy before AI and working in fragmented, legacy systems only exacerbates the issue. I once observed retail bank employees copy-pasting information between no fewer than eight apps during a loan application interview, including one used just to strip formatting before pasting the text elsewhere. On a 30-minute call, agents used this workaround more than once per minute and made three keying errors.
I saw then how deeply the finance industry runs on stacks of specialized software. Trading terminals, market data feeds, CRMs, analytics platforms and whatever homegrown software IT built a decade ago. None of it was built to talk to the rest. Generative AI adoption makes sense for the industry when it eases the friction of re-entering data from one system to another (eliminating the chance for keystroke errors and data leakage in the process). However, an LLM that drafts a loan narrative but still needs its output copied across seven other applications just adds more work.
Integrating AI compliantly and productively
The financial institutions getting AI implementation right are building interoperable digital workspaces that share context between systems automatically. Their employees get to direct their attention to making judgment calls rather than copy-pasting data into different windows. They can truly do their jobs faster and better without abandoning the guardrails regulators set in place for good reason. Isn't that the whole point of introducing AI to a regulated industry?
Human-in-the-loop calls for a professional with enough time, context and mental bandwidth to catch what a model gets wrong. It’s not enough to guarantee the signature on the form. You have to guarantee the sound judgment behind it. Until financial institutions address the disconnected and cumbersome interfaces their employees work in every day, putting an overloaded human in the loop is a compliance fallacy.
