By Michael Crockett, Chief Operating Officer
The AI conversation in mortgage lending has gotten loud. Every conference has a panel on it. Every vendor has a pitch. And underneath all of that, most lenders are asking the same question: how do we actually do this?
Not the theory. Not the marketing. The actual implementation: how do you bring AI into a heavily regulated, high-stakes environment in a way that works, holds up to scrutiny, and does not create more problems than it solves?
I have been in this industry long enough to watch a lot of technology waves move through it. The ones that left damage behind were rarely bad ideas. They were good ideas deployed before the organization was ready to support them. AI is no different. The question is not whether the technology works. It is whether organizations are ready to deploy it responsibly, consistently, and at scale.
The Gap Between Potential and Practice
The capabilities of AI in mortgage are real. From analyzing self-employed income and processing documents to detecting fraud, automating verification workflows, and identifying risk patterns across large datasets, the technology can meaningfully improve speed and consistency. The gap is not what AI can do. It is in what it takes to deploy it responsibly in an environment like ours.
Mortgage lending is not a forgiving context for trial and error. Every decision affects people’s ability to buy homes, and every decision must withstand regulatory scrutiny. An AI model that performs brilliantly in a controlled environment can still create serious problems in production if governance and explainability were not built in from the start.
Most AI pilots look promising. The use case is well defined, the data is clean, the team is motivated. Then production happens. The data is messier. Edge cases that never came up in testing start appearing constantly. Governance questions nobody answered during the pilot because there was not time suddenly matter a great deal. McKinsey puts a number on it: 88% of organizations now use AI in at least one function, but only 39% see any measurable financial impact. The gap between deployment and value is exactly where this conversation needs to happen.
What Compliance Actually Demands
The regulatory framework around AI in mortgage is not ambiguous, even if it is still evolving. Fair lending law applies to AI-driven decisions the same way it applies to human ones. If an algorithm produces outcomes that have a disparate impact on protected classes, that is a fair lending issue regardless of whether the model intended it. Adverse action requirements do not disappear because a model made the call. When a credit decision goes against a consumer, they are entitled to specific reasons that a human can understand, and a regulator can evaluate. “The model said no” does not satisfy ECOA or the FCRA.
The consequences when organizations fall short are real. In 2025, Massachusetts reached a $2.5 million settlement with an AI-driven lender over practices found to cause disparate harm to protected groups. The model made the decision. The organization bore the liability. That dynamic is not going away.
Compliance requirements are not a filter you run finished work through at the end of a project. They are design constraints that shape what you build from the beginning. Explainability requirements affect which model architectures are viable. Fair lending obligations affect what features can be used in training. The strongest AI programs involve compliance, legal, and risk from day one of model design, not waiting at the end to review what got built. It slows down the early stages and speeds up everything after.
The Part That Gets Skipped
Beyond compliance, there is a set of operational questions that are less obvious but just as consequential. Who owns the model’s outputs? That sounds simple until a model and a human underwriter disagree, which they will. What is the escalation path? Who has authority to override, and under what circumstances? How do you know when the model starts drifting, performing differently in production than it did during validation? If you do not have monitoring that surfaces drift before it becomes a pattern, you are flying without instruments.
And then there is the people side, which is consistently the most underestimated part of any AI deployment. Experienced loan officers, underwriters, and operations teams bring years of professional judgment to every decision. Asking them to trust a model they do not understand produces one of two outcomes: they ignore the tool entirely, or they defer to it without applying the judgment that catches the cases where the model is wrong. Neither gets fixed by better technology. It gets fixed by investment in training, transparency about what the model does and does not do, and a clear understanding of AI’s role in the decision-making process. The goal is for AI to enhance professional expertise, not replace it.
The data on this is consistent: roughly 70% of AI implementation challenges trace back to people and process issues, not the algorithms. Technology is rarely the problem. The organization around it usually is.
What Good Actually Looks Like
When AI is deployed with the right governance and operational infrastructure, what changes is not just speed. Decision quality improves because the model surfaces patterns that human review at scale cannot catch. Consistency improves because the same logic applies to every file, regardless of who is handling it. Risk becomes more manageable because potential issues are flagged earlier, before they become problems downstream.
The lenders I have seen get this right share a few things in common. They started with a narrow, well-defined problem rather than trying to transform everything at once. They invested as much in governance, training, and operational processes as they did in the technology. And they built monitoring in from the beginning so they could continuously evaluate performance as markets and consumer behavior evolve. None of that is glamorous, but it is what creates a durable competitive advantage.
The Conversation Worth Having
The opportunity in AI is real, and the lenders who figure this out will have genuine competitive advantages in speed, consistency, and consumer experience. The conversation just needs to go deeper than it typically does. Not just what AI can do, but what it takes to deploy it in a way that holds up. Not just the pilots, but what happens when you run it at scale with real data, real teams, and real operational complexity.
That is the conversation I am looking forward to having at the HousingWire AI Summit. If you are working through these questions in your own organization, I hope to see you there.
Sources
- AI Adoption in Financial Services 2025 Report, Caspian One
- Why AI Projects Fail: The 63% Human Factor Problem, Bosio Digital
- Why 90% of Enterprise AI Implementations Fail, Talyx AI
- CFPB AI Fair Lending Rollback: What Still Binds Lenders in 2026, The Leveraged Years
- AI Governance in Mortgage Banking: Moving From Policy Discussion to Regulatory Reality, National Mortgage Professional
Author
Read previous blog: From Static Processes to Confident Decisions