Who Is Training the Human in the Loop

A few days ago, I posted on LinkedIn that I had only recently learned the acronym HITL after hearing it on an industry podcast. Since then, my antenna has been up. I keep seeing “human in the loop” in posts about AI, often offered as reassurance that someone is overseeing the technology.

The pessimist in me wonders whether the phrase is becoming a marketing tool before we have asked enough questions about the human. What preparation sits behind that reassurance? How does a company know its employees understand the limitations of the AI they are expected to oversee?

Then a friend, Kevin Conte sent me an article about PwC’s investment in critical thinking. A human friend, for the record. PwC has announced a program for more than 85,000 partners and employees, including practical simulations relevant to their work. Bloomberg Tax reports an estimated investment exceeding $100 million, including development costs and university funding. The rollout is scheduled to begin in December 2026, so its results are still ahead of us. I think the commitment to investing in people deserves our attention.

Much of what I hear in mortgage centers on AI governance at the 30,000-foot level, and rightfully so. I am hearing less about oversight on the ground, where an employee has to decide whether an output is reliable enough to use. That is where I believe we need a deeper conversation.

My emerging concern is employees’ knowledge of the AI tools themselves. A mortgage professional can understand guidelines, programs and processes thoroughly while still needing preparation to supervise AI. Does that person understand hallucinations, where generative AI produces convincing but incorrect or fabricated information? Do they know how to check the underlying source, including whether a cited source actually exists and supports the answer?

They should also understand why earlier performance cannot settle the question of future reliability. Changing data and operating conditions can affect results, which is why drift monitoring matters. Frontline employees need to recognize concerns and know where to report them. Technical and governance teams need to own systematic monitoring and investigation. Oversight requires support beyond the person reviewing an individual output.

Reading about PwC, I found myself putting my lender C-suite hat back on. It has been several years since I wore it, but it still fits well. I thought about what I would do if I were leading a lender today.

I would start with an assessment of every teammate’s AI acumen. That would mean getting beyond whether someone uses AI and understanding how they use it. I would want to know which tools they rely on, what work those tools perform, and how they decide whether an output can be trusted.

I would use the AI System Inventory in FRAME, MISMO’s AI governance toolkit, to identify the systems involved. I would pair that inventory with questions about employees’ understanding and validation practices. Each person should be able to explain the review process appropriate to their use of AI and when they need help. Those expectations should be part of the organization’s overall AI governance program.

I would also want to see that understanding demonstrated. Consider a training exercise in which an AI-generated explanation cites a plausible but nonexistent program requirement. An employee could walk through how they would investigate it, check authoritative information and handle the discrepancy. That exercise would reveal something a question about how often they use AI would miss.

The gaps we discover should shape the curriculum. Some employees may need foundational knowledge about AI’s limitations. Others may need practice validating outputs from a particular tool. Critical thinking brings that knowledge into the work: examining the evidence, questioning assumptions and deciding what further review is warranted. Leadership should provide consistent validation procedures and enough time to follow them, with review matched to the risk of the task.

I would expect this to be a looped process. Training should be followed by opportunities to practice and reassessment of what employees understand. As tools change, new uses emerge and responsibilities shift, we should revisit the preparation those changes require. Problems encountered in daily work should help inform the next round of training. The curriculum needs to stay current with the organization’s adoption of AI.

As I work through this issue, I see an opportunity for MISMO to help. Our new AI Governance Forum provides a place to share experiences and help shape the evolution of FRAME. Together, the Forum and FRAME give us a foundation for collaboration, consensus and adoption around the practical challenges our industry is facing.

I believe preparing the humans in the loop belongs in that conversation. We could work together on common assessment questions, realistic training scenarios and shared learning resources that companies adapt to their tools and workflows. Lenders could bring the gaps they encounter, while technology providers help explain the limitations and oversight needs of their products. That is an opportunity to explore through the Forum.

When a company says it keeps a human in the loop, it should be able to explain how that human is prepared for the responsibility. The investment in that preparation should continue for as long as we rely on them to provide oversight.

#VieauxPoint