The F-16 fighter jet is inherently designed to be unstable, relying on advanced instrumentation and computer-driven feedback loops to remain in flight. Without these systems, it quickly transforms into a costly projectile. This analogy often crosses my mind when I see AI governance reduced to mere compliance checkboxes. Effective risk management must be dynamic—not a dusty artifact that is revisited only after a crisis. It should commence early, implement substantial controls, and continually adjust as circumstances evolve.
The Compliance Narrative Has Limitations
On August 6th, Fannie Mae introduced new AI/ML governance requirements similar to those of Freddie Mac. While meeting the deadline was straightforward, the real challenge lies in demonstrating to an examiner that governance is an evolving entity.
AI governance is often justified through apprehension—fear of fair lending violations, model risk, or failing examinations. However, fear has its limits. A compliance-focused program merely functions as a safety net, preventing losses but not conferring a competitive advantage.
Moreover, it’s insufficient. The critical question isn’t just “How can we prevent AI from causing harm?” but rather, “Do we have real-time insight into what AI is doing?” Too frequently, the response is “No,” stemming from a gap in feedback loops where real-world developments outpace governance mechanisms.
It Begins with Effective Instrumentation
Consider the primary risk associated with any mortgage AI governance framework: fair lending.
The traditional compliance approach involves quarterly testing for disparate impacts based on samples after loans have already been finalized. In contrast, an instrumentation-focused strategy monitors this risk in real-time, identifying discrepancies before decisions involving protected-class borrowers are made.
It’s the same legal requirement and risk, but with different operational timelines.
This isn’t a problem confined to the mortgage sector; it’s pervasive. In a piece for the Organisation for Economic Co-operation and Development (OECD), Duke’s Lee Tiedrich referred to the pervasive “evaluation gap,” illustrating the disparity between controlled lab performance and unpredictable real-world applications. Our industry happens to be particularly vulnerable, with regulators and lawsuits already in play.
Episodic Reviews: A Relic of the Past
Even with instrumentation, an F-16 requires between 15 to 20 hours of maintenance for each hour of flight. Continuous telemetry doesn’t eliminate the need for routine inspections; instead, it enhances their reliability.
Episodic audits and reviews assume that system behavior remains static between evaluations. Given the rapid pace of change, this assumption is outdated. Foundation models are updated on their vendors’ timelines, not yours, while teams integrate new AI swiftly, often outpacing inventory efforts.
As a result, examinations reflect an organization that may have ceased to exist weeks prior. That’s not risk management—it’s merely posturing. While periodic records remain essential, they cannot be the sole focal point of oversight.
I’ve Witnessed This in Action Twice
I have seen the story of leveraging instrumentation as a competitive advantage unfold twice in different sectors.
- Bloomberg, early 1990s. While helping design the firm’s SOC II precursor, I observed the founder operate with a trader’s urgency—an acute need for immediate data. The culture dictated the infrastructure, ensuring that leadership had the visibility needed to make informed decisions. This resulted in a company that could track its own performance in real time, while the rest of the industry relied on nightly data batches.
- Goldman Sachs, late 2000s. Sitting across from Goldman at Morgan Stanley, I noticed that its competitive advantage was not rooted in a superior model but rather in a streamlined feedback loop between monitoring positions in its SecDB and taking action. Morgan Stanley had sufficient instrumentation but still incurred approximately $9 billion in losses during mortgage trades. Alerts were present; those monitoring them simply failed to act swiftly.
This latter observation underscores a crucial point: having instrumentation is inadequate unless there are individuals ready to act promptly on the insights provided—rather than waiting for the next week’s meeting.
The Barriers to Entry Have Lowered
Both Bloomberg and Goldman utilized proprietary, costly systems that once defined the standard for entry. This is no longer the case—those barriers have diminished.
Mortgage companies don’t need to build foundational models from scratch. AI capabilities are integrated within existing loan origination systems, point-of-sale platforms, CRMs, and fraud detection tools they already employ. These off-the-shelf, AI-enhanced solutions can achieve much of what Bloomberg developed initially. The task at hand isn’t about constructing a SecDB; it’s about demanding visibility from already-existing platforms and actually leveraging that insight.
Returning to the Aircraft Analogy
AI that updates per a vendor’s schedule and proliferates faster than it can be recorded creates unpredictability. The instinct to prioritize compliance is to treat that unpredictability as an adversary—limiting adoption until outdated checks can regain control.
The F-16’s developers, however, chose to embrace this instability, designing instrumentation that translates it into agility. The successful institutions won’t force AI into predictability; they’ll establish the right systems to enable optimal performance.
Importantly, this instrumentation isn’t inherently advantageous; it’s become a standard requirement. What distinguishes winning entities in the field is how swiftly and effectively their operators interpret that data in real-time scenarios.
This differentiation applies in AI governance as well. The instrumentation layer can and should be standardized across the industry, as emphasized in the OECD discussion. However, the competitive edge is still earned at the individual lender level, dependent on how effectively they interpret and apply that intelligence.
You need not instrument every aspect from day one. Focus on understanding where risks are concentrated and start from there.
A Familiar Demand in a Modern Context
Reconceptualized, AI governance is not a new discipline; it represents a longstanding demand—essentially asking, “Does your organization maintain live visibility into its operations?” Bloomberg had it; Goldman acted on it promptly; Morgan Stanley possessed it but was slow to respond. AI governance is this necessity redefined for a landscape that evolves faster than routine evaluations.
This shift in perspective alters who secures funding. A compliance-centered approach merely garners a defensive budget; a focus on instrumentation attracts executive support as a means to gain a competitive edge rather than just risk mitigation. This proactive stance compounds benefits: institutions monitoring AI in real-time are more agile, approving use cases that competitors are still deliberating and venturing into adjacent markets before others catch up.
Industry discussions shouldn’t stop at minimizing harm. They should address how we can establish a common instrumentation layer that empowers each lender to monitor their AI operations at machine speed—similar to the innovation Bloomberg enforced in market data, refocused towards AI oversight. Standardize the instruments while allowing individual mortgage companies to maintain their advantages in execution.
The music is playing. The real concern is whether we monitor the instruments closely enough to detect tempo changes or if we learn the hard way, as Goldman’s peers did.
Marvin Chang is the Executive in Residence at Duke University Pratt School of Engineering and Principal at Mercer Knoll Strategies.
This commentary does not necessarily represent the views of HousingWire’s editorial team or its ownership. To reach the editor responsible for this article: [emailprotected].
