Where AI Should and Shouldn’t Make Lending Decisions

I have spent the last several months working on a book about artificial intelligence in lending, and one question keeps surfacing in almost every conversation I have with COOs, Heads of Lending, and digital transformation leaders at specialty and commercial lending institutions. It is not whether AI can do more inside the lending process. At this point, it clearly can, and the pace of what it can do is accelerating quickly. The real question, the one that separates organizations that will get stronger from those that will get burned, is where exactly the line falls between AI that augments the judgment of an experienced lender and AI that quietly replaces that judgment in ways nobody planned for.

This is not an abstract question. It is showing up right now in how lenders are configuring their loan origination software, how they are training staff to interact with AI-enabled tools, and how they are thinking about risk governance as they modernize operations that were built on spreadsheets and manual review. Getting the answer right determines whether AI becomes a durable operational advantage or a liability that does not reveal itself until a bad decision has already happened.

The lending process is not one workflow, it is many

The mistake I see most often is treating the lending lifecycle as a single, uniform workflow and applying the same posture toward AI across all of it. In practice, the lending process is a sequence of decisions with very different characteristics, and those differences matter enormously when you are deciding where AI belongs.

Some decisions in the lending lifecycle are high-volume, repetitive, and governed by clear, definable rules. Others are low-volume, complex, and dependent on the kind of contextual judgment that a credit officer builds only after evaluating hundreds of deals across different market cycles. AI performs very differently in these two environments. Applying it uniformly, as if a rules-based screening step and a complex credit approval carry the same risk profile, is one of the most common and most consequential errors lending organizations make when they start deploying AI tools inside their operations.

The organizations that get this right are not the ones asking whether to adopt AI in lending. They are the ones doing the harder work of mapping their process step by step and asking, honestly, what kind of decision is actually happening at each point.

Where AI is genuinely transformative

On the repetitive, rule-based end of the spectrum, AI is not incremental improvement. It is transformative, and the case for using it is straightforward. Document extraction and data validation are an obvious example. Lending operations have historically absorbed enormous amounts of manual labor simply moving information from borrower-submitted documents into origination systems, checking it for completeness, and flagging discrepancies. AI-enabled document extraction does that work faster and more consistently than manual review, and it does it without the fatigue-driven error rate that creeps into any high-volume manual process.

Initial application screening against defined credit criteria is another clear fit. If a lender has established thresholds, debt service coverage minimums, loan-to-value caps, or industry exclusions, AI can apply those consistently across every application without the variability that inevitably shows up when different underwriters interpret the same policy slightly differently. Payment processing and ACH management fall into the same category. These are operational tasks with clear rules and low tolerance for inconsistency, which is exactly the environment where automation performs best.

Portfolio monitoring is where I think AI is proving to be the most quietly valuable. Flagging loans that show statistical patterns associated with future default, patterns a human reviewing a portfolio manually might not catch until they are already showing up in delinquency reports, gives lending teams an early warning capability that did not exist at scale before. And generating first-draft credit summaries from structured data inputs frees experienced underwriters from repetitive writing so they can spend their time on analysis rather than data assembly.

What all of these use cases have in common is that they reduce manual work, increase consistency, and free experienced lending professionals to spend their time on the decisions that actually require their judgment. That is the correct framing, and it is worth repeating because it gets lost in a lot of AI marketing. The goal is not to remove people from the process. The goal is to remove low-judgment work from their day so their judgment gets applied where it is worth the most.

Where the picture changes

On the complex judgment end of the spectrum, the picture changes considerably, and lending executives need to be honest with themselves about why. Consider a thirty million dollar commercial real estate bridge loan. That evaluation involves the experience and track record of the borrower, the specific market dynamics of the asset’s location, the quality and credibility of the proposed exit strategy, and often the relationship history between the lender and the borrower built over multiple deals. None of that reduces cleanly to a rule or a statistical pattern, because each of those factors is context-dependent in ways that resist standardization.

An experienced credit officer brings years of pattern recognition to that kind of evaluation. They have seen exit strategies that looked sound on paper fail because of market timing. They have seen borrowers with thin track records outperform because of factors that do not show up cleanly in a credit file. That pattern recognition is not something AI can replicate, because it is built from lived experience across ambiguous, non-repeating situations, not from labeled data. AI can support that evaluation. It can surface relevant comparable transactions, summarize borrower history, and flag inconsistencies across documents. But it should not be rendering the final judgment.

The risk of over-relying on AI in that kind of decision is not just the possibility of one bad loan, though that is real. The deeper risk is organizational. Credit judgment capability takes years to build inside a lending organization, and it erodes quietly if it stops being exercised. If an organization leans on AI outputs for complex decisions long enough, the muscle of contextual credit judgment atrophies, and by the time a market cycle turns and that judgment is needed most, it may not be there anymore. That is a risk that does not show up on a dashboard. It shows up years later, and by then it is expensive to rebuild.

Augmentation versus replacement is the right frame

The frame I keep coming back to, and the one I think every lending executive evaluating AI adoption should adopt, is augmentation versus replacement. AI should be making experienced lenders faster, better-informed, and more consistent. It should be handling the work that does not require judgment so that judgment can be applied where it actually matters. It should be surfacing information and flagging patterns a human might miss when managing a large and complex portfolio. What it should not be doing is serving as the final word on a complex credit decision.

Organizations that blur that line are not necessarily making a decision they will regret immediately. That is what makes the risk dangerous. They are taking on risk that will not be visible until something goes wrong, often well into a credit cycle, when a concentration of loans approved with insufficient human judgment starts underperforming at the same time. By the time the pattern is visible in delinquency data, the exposure has already been built.

This is also why governance conversations around AI in lending need to happen earlier than most organizations are having them. It is not enough to deploy AI tools and see what happens. Lending organizations that are thoughtful about this are defining, in advance, which categories of decisions AI is permitted to influence, which categories it can only inform, and which categories require a human decision-maker with clear accountability. That distinction needs to be built into workflow design, not left to individual discretion in the moment a loan is being reviewed.

Why this matters for how lending platforms are built

This is directly relevant to how lending software gets architected, and it is part of why I think the platform question matters more than most lenders initially assume. A lending platform built on Salesforce has an advantage here, because the underlying architecture supports configurable workflows where AI-enabled steps, document extraction, initial screening, portfolio monitoring, can be embedded directly into the process while structurally routing complex credit decisions to human reviewers with full visibility into the supporting data.

That is a meaningfully different design philosophy than treating AI as a bolt-on tool sitting outside the core system. When AI capabilities are embedded natively into the same platform that manages origination, underwriting workflows, and loan servicing, the organization has one place to define where automation applies and where human judgment is required, rather than trying to enforce that distinction across a patchwork of disconnected point solutions. An alternative lending platform that cannot make that distinction configurable at the workflow level is going to struggle to give lending executives the control they actually need as AI capability keeps expanding.

This also matters for auditability. Regulators and internal risk committees are going to want to know, for any given loan, what role AI played in the decision and what role a human played. A platform where that distinction is baked into the workflow, rather than reconstructed after the fact from disconnected systems, makes that conversation far easier. It also makes it easier to demonstrate that the organization has thought seriously about where automation belongs, which increasingly matters for both regulatory conversations and for institutional investors evaluating a lender’s operational maturity.

The practical exercise every lending executive should run

The practical exercise I would recommend to any lending executive thinking seriously about AI adoption right now is straightforward, even if it takes real effort to execute well. Map the lending process step by step, from initial application through servicing, and for each step ask a single question. Does the decision at this step benefit primarily from speed and consistency, or does it benefit primarily from experience and contextual judgment.

Steps that benefit from speed and consistency are strong candidates for AI augmentation, and the case for moving quickly there is strong. Steps that benefit from experience and contextual judgment are steps where AI can inform the decision but should not be making it, and where human oversight is not a compliance checkbox. It is the product. It is the thing the organization is actually selling when it tells a borrower, an investor, or a regulator that it exercises sound credit judgment.

That mapping exercise will look different for every lending organization depending on the complexity of their portfolio, the diversity of their loan products, and how much of their process is already standardized versus still dependent on individual underwriter discretion. But the exercise itself is not optional if an organization wants to adopt AI responsibly rather than reactively. The lenders who do this mapping deliberately, rather than letting AI adoption happen piecemeal across departments, are the ones who will end up with AI that makes their organization stronger. The ones who skip that step are the ones who will eventually discover, usually at the worst possible moment in a credit cycle, exactly where the line should have been.