In mortgage operations, an AI error is not an inconvenient suggestion. It is a repurchase demand, a fair lending flag, or a compliance finding that lands on a desk weeks or months after a loan closed. The stakes attached to every document extraction, income calculation, and underwriting decision mean that deploying AI in a mortgage environment is a fundamentally different problem than deploying it anywhere else.
That distinction shaped every conversation at a recent California MBA Technology Roundtable, where four leaders who build, deploy, and govern mortgage AI sat down to talk candidly about what responsible adoption actually looks like in production.
The panel included Rajan Nair, CEO of Indecomm, who has spent more than two decades embedding mortgage operational expertise into AI systems; Andrew Komaromi, Senior Director of AI and Automation at Stewart, who builds custom LLM solutions for mortgage and title workflows; and Alicia Gazotti, an enterprise risk and compliance leader focused on how AI is reshaping mortgage risk management — with Bryan Jackson, CTO at Gateless and chair of the CMBA Technology Committee, moderating.
The efficiency gains from mortgage AI are real and measurable, the panel agreed. So are the risks. And the gap between the two tends to show up in the same place every time: not in the technology itself, but in everything that was supposed to surround it. The testing that should have happened before an AI model touched a live loan file, the mortgage compliance team that should have been in the room during AI system design, and the governance framework that is still trying to keep pace with agency guidelines and regulations that have not finished being written.
As Rajan Nair put it plainly: every AI solution’s accuracy looks highest in the sales deck. The real test is what happens after it touches a live loan.
Mortgage AI Vendor Promises Vs. Deliverables
A year ago, a mortgage AI vendor could close a meeting with a clean demo. Now the first questions to come from mortgage lenders are about failure of a mortgage AI solution. With intelligent documents: Where does the automated document extraction break on a self-employed borrower’s file? In underwriting: What happens when the income calculation model encounters an edge case it was never trained on? On QC feedback loops: Who is accountable when an AI-driven underwriting decision contributes to a repurchase demand or a regulatory finding?
Lenders asking those questions are not slowing adoption. They are forcing mortgage AI vendors to answer for the part of the pitch that never survives a live loan environment. Real mortgage operations are full of nuance: the self-employed borrower whose income spans four document types, the property that falls outside the model’s training data, the exception a human underwriter flags in thirty seconds that the AI passes without comment. Those are the files that reveal whether a mortgage AI system actually works.
Rajan Nair pointed to the reason those questions keep surfacing. Most off-the-shelf mortgage AI was built on generic loan data with no exposure to a specific lender’s document mix, borrower population, or exception history. When that model meets a real production environment, the gaps show up fast: productivity gains that looked certain on paper disappear, and compliance exposure accumulates in the places nobody thought to test.
The institutional knowledge that fills those gaps cannot come from a vendor. It lives in years of actual loan decisions made at that institution : the edge cases, the document variations, the exceptions that only show up in a specific lender’s market. A mortgage AI system that has never processed those loans has no basis for handling them correctly. That knowledge has to be built in, decision by decision, before the system can be trusted with production volume.
Andrew Komaromi echoed that point and added a practical corollary: the way to manage the gap between promise and reality is to make the rollout small and visible before it is large and assumed.
His approach is to carve out a subset of loan orders and run the mortgage AI process in parallel with the existing human process, measuring the actual gain before expanding. The goal is not to generate a headline number for a business case. It is to show the team exactly what the tool does and does not do, in their own environment, on their own files.
Why mortgage AI fails after deployment:The trust tax
When it comes to mortgage AI solutions, trust is critical. There is productivity loss when a loan reviewer or mortgage underwriter does not trust the mortgage AI system’s output enough to act on it, redoes the work manually, and the organization ends up paying for both the AI process and the human process while capturing the benefit of neither.The panel assigned this a name: The Trust Tax
Rajan Nair identified the direct cause of mistrust with mortgage AI. When a mortgage AI system such as an AI-powered AUS returns information such as an underwriting decision on a loan file without any explanation of how it reached that conclusion, the mortgage underwriter has no basis for calibrating when to accept the output and when to push back. The natural response is to verify everything, which eliminates the productivity gain entirely.
The solution he described is what Indecomm calls the glass box: a mortgage AI system that shows its reasoning, surfaces the data points it drew on, and is willing to flag its own uncertainty rather than returning a false confidence.
Andrew Komaromi added a second cause of the trust tax that has nothing to do with model transparency: excluding the end users from the build process. Loan processors, auditors, and mortgage underwriters who had no involvement in designing or testing the system have no stake in its success. They did not see how it was built, they were not part of the testing, and they have no reason to advocate for it with their team. The result is a slow-motion adoption failure that does not show up in any single report; it shows up as a system nobody uses quite the way it was designed.
His counter to this is direct involvement: bring the mortgage subject matter experts into the process early, make them part of the iterative testing, and give them ownership of the outcome. A mortgage AI system that the people using it helped build gets used differently than one handed to them after launch.Governance, validation, and who owns the liability when mortgage AI gets it wrong
The most sobering voice on the panel belonged to Alcia Gazzotti, highlighting the risk, compliance and governance aspects of AI. Gazzotti made the clear point that it goes beyond mortgage trust. Mortgage teams that were never part of building mortgage AI solutions are now being asked to govern them, inside a regulatory landscape that has not finished being written.
If a model is trained in a way that starts picking up on demographic signals, that becomes a fair lending risk the organization did not design for and may not catch until an exam finds it. Add to that the consumer opt-out laws some states have enacted, which let borrowers decline AI-based fraud detection entirely and leave the lender with no real substitute for those transactions, active CFPB proposals on adverse action notices tied to AI-driven decisions, and a state like California cycling endlessly through legislation proposed, enacted, failed, and proposed again.
Deployment is outrunning governance almost everywhere. The lenders managing that gap well are not waiting for regulatory clarity before they act. They are building governance in parallel with deployment, so that by the time a regulator asks the question, the answer already exists.
AI-enabled mortgage fraud: the offense is ahead of the defense
Alicia Gazotti also brought the fraud conversation into focus. The threat landscape for mortgage AI fraud has expanded well beyond document forgery, and the consequences of being behind it are direct and quantifiable.
AI-generated pay stubs, W-2s, and bank statements are now clean enough to pass initial loan review. The GSEs are actively screening submitted documents on the back end for exactly this reason. A lender that is on the defensive rather than the offensive has already created repurchase exposure for itself.
Beyond document fabrication: synthetic identities assembled from real personal data belonging to deceased individuals, children, or elderly borrowers. Deepfake voice and video convincing enough to impersonate a title agent, an escrow officer, or a company executive on a video call. Wire interception schemes that insert themselves into a legitimate email chain and ride it all the way to closing before redirecting funds. Once a wire moves through three or four layered accounts, the recovery window closes fast.
Rajan Nair argued that fighting AI-generated mortgage fraud purely with AI detection is a losing strategy on its own. Treating it as an arms race by deploying fraud-detection AI to chase fraud-generating AI means always being reactive. The more durable defense is using AI to dramatically expand the number of verification steps applied to every loan file: cross-checking values across documents, comparing data points that a human reviewer would never have time to reconcile manually, and flagging inconsistencies for human review rather than trying to make a binary call.
Andrew Komaromi’s point on the human side of fraud defense was direct: the vast majority of successful fraud in mortgage involves social engineering, not technical infiltration. Someone receives convincing communication, trusts it, and acts on it. The defense is education, not just firewalls: sharing specific real-world examples of how fraud schemes succeeded, the way the NTSB releases findings after an aviation incident, so that teams understand exactly how the deception worked rather than receiving abstract warnings about emerging risk.
Operational controls matter as much as detection. Requiring dual approval on wire transfers and designing systems so a single compromised credential cannot initiate a funds movement alone will not eliminate fraud risk, but they contain the damage when a scheme succeeds.
What the Panel Agreed On
None of the panelists argued against mortgage AI. They are running it in production. What they argued for, collectively, was a more honest accounting of the gap between what mortgage AI looks like in a vendor demonstration and what it requires to work reliably in a live loan environment.
Rajan Nair’s closing advice was direct: validate before you automate. Test the system against your hardest loan files. Build the escalation path before the exceptions arrive. Make sure the people using the system have visibility into every decision it makes. The demo will always look better than the deployment. It is critical for all parties to work together toclose that gap before a mortgage AI solution goes live, not after.
Alicia Gazotti’s closing point was a caution against letting efficiency goals outrun risk appetite. No agency or GSE has told lenders to remove the human from the underwriting decision. A DE underwriter still has to sign off. The organizations creating exposure for themselves are the ones whose mortgage AI deployment has moved past what the regulatory framework has actually authorized.
Andrew Komaromi’s closing was a reset on expectations: mortgage AI is a capable tool that augments the people using it. It is not a system that replaces human judgment, runs itself, and produces perfect outputs. Treating it like one is, as he put it, a recipe for disaster.
Deployment in mortgage AI will always move faster than governance. The organizations managing this well are the ones who decided not to let that gap become a liability before anyone noticed.
