For most credit unions, lending has always been a balance between structure and discretion.On one side, you have your underwriting criteria: Credit scores, income thresholds, debt ratios, policy guardrails. On the other, you have human judgment: Context, member history, edge cases that don’t fit neatly into a model.
That balance has held for a long time, and AI doesn’t remove it—but it is starting to shift where and how decisions get made.
Lending has never been purely rules-based
Even in highly standardized lending environments, real, human discretion has always played a role.
A borrower might fall just outside a threshold but still get approved based on compensating factors. A long-time member relationship might influence how risk is interpreted. A loan officer might spot something in an application that doesn’t show up cleanly in a score. That flexibility is part of what has historically differentiated credit unions from purely algorithmic lenders.
At the same time, it creates variability. Two similar applications can produce different outcomes depending on who reviews them, how they interpret the information, and how comfortable they are making exceptions.
That tension—between consistency and discretion—is not new. But it is becoming more visible within the advent of AI.
AI reshapes the workflow around judgementA
A lot of the early conversation around AI in lending focused on automation: faster approvals, lower costs, reduced manual effort. Some of that is happening, but for now, the more immediate shift is AI getting better at structuring information before a decision is made.
Instead of replacing the loan officer, it changes what they’re looking at and how quickly they can interpret it.
For example:
- Summarizing borrower profiles across multiple data sources.
- Highlighting potential risks or inconsistencies.
- Surfacing comparable cases or historical outcomes.
- Flagging edge cases that fall outside typical patterns.
None of this makes the decision on its own, but it compresses the time and effort required to get to a clear view of the situation. The role of the human decision-maker starts to shift from gathering and organizing information to evaluating and approving it.
The real shift is consistency, not just speed
Speed is the obvious benefit here, sure, but consistency is arguably the more important one. When AI is used to structure inputs and surface key factors, it reduces the variability in how information is presented and interpreted.That actually makes human judgement more visible, and more comparable across decisions, too. Instead of each loan officer building their own mental model of a borrower from scratch, they’re starting from a more standardized view.
That makes it easier to: Apply policy more consistently.
Identify when an exception is actually warranted.
Explain decisions internally and externally.
In other words, AI doesn’t remove discretion—it gives it a more stable foundation.
Where things start to break
The downside is: This is also where the risk shows up. I.e, if AI is introduced as a layer on top of existing processes without clear ownership, it can create more confusion than clarity. Who is responsible for the output?
How much weight should be given to what the system surfaces?
When should a recommendation be overridden—and by whom?
Without clear answers, institutions end up in an awkward middle ground:
- Decisions are influenced by AI, but not fully understood.
- Accountability becomes harder to trace.
- Exceptions become harder to justify.
It looks like progress in theory, but in practice, it can increase operational and compliance risk. This is the same pattern showing up across other areas of AI adoption: activity increases before alignment does.
Human judgment becomes more visible, not less important
One of the more interesting effects of AI in lending is that it, ironically enough, actually makes human judgment more explicit.
When decisions were slower and more manual for example, a lot of judgment lived in the process itself, so it was harder to isolate. But now, as AI standardizes inputs and accelerates analysis, the remaining points of discretion stand out more clearly.
That raises the bar for lending decisions. If a decision deviates from what the system suggests, there needs to be a reason. Not necessarily a rigid justification, but a clear line of thinking.
That’s not…a bad thing, at all. It pushes institutions toward more intentional decision-making. But it does require a shift in how lending teams think about their role.
What this means for credit unions
The question is not whether AI will replace human judgment in lending, because it very likely won’t ever do that. The more relevant question is how that judgment is structured, supported, and governed as AI becomes more embedded in the process.
A few practical implications:
Make decision ownership explicit: As AI becomes part of the workflow, it needs to be clear who is ultimately responsible for the outcome. Not the system, not the model—the institution.
Define where discretion lives: Not every part of the process needs the same level of flexibility. Being explicit about where judgment is expected—and where it isn’t—reduces confusion.
Focus on explainability early: If a decision can’t be explained clearly, it becomes harder to defend. That applies to both AI-supported recommendations and human overrides.
Treat AI as a structuring layer, not a replacement layer: The most effective use cases right now are not about removing people from the loop. They’re about improving how information is organized and decisions are framed.
Lending is not becoming fully automated overnight — probably never will be — and for credit unions, it probably shouldn’t.
Instead, the shape of the decision-making process is changing in real time. Less time gathering information, less variation in how cases are presented, and more focus on interpretation, judgment, and accountability. That may not be as dramatic as “AI replacing underwriters.” But it’s more realistic—and, in the long run, more important.
The institutions that navigate this well will not be the ones that remove humans from the process. They’ll be the ones that make human judgment more consistent, more visible, and better supported.
If this feels familiar, this is exactly the kind of work we’ve been doing with leadership teams through AI BASECAMP—helping them get oriented before decisions start compounding.
