Artificial intelligence is rapidly becoming a cornerstone of enterprise operations, creating increasingly interconnected ecosystems that challenge traditional governance models. As organizations embed AI into critical workflows, maintaining visibility into system dependencies has emerged as a pressing leadership concern. A recent study on AI sovereignty found that 91% of surveyed executives admit they do not fully understand their organization's AI dependencies. Moreover, respondents experienced an average of six AI-related disruptions over the past two years. These statistics highlight the urgent need for governance practices to evolve alongside the technology they oversee.
Jeffrey Rachlin and his partner Andy Hyman have extensively studied complex, technology-driven environments. They observe that many organizations still investigate failures only after visible disruption occurs. However, as AI systems assume greater autonomy across business processes, retrospective analysis provides only a partial picture. The duo advocates for governance methods that can identify meaningful changes while intervention remains possible—a concept they call monitoring systemic drift.
This perspective reflects a broader shift in how organizations should think about operational health. Traditional monitoring emphasizes outcomes through dashboards, reports, and key performance indicators (KPIs). While valuable, these tools describe the results produced by a system rather than the relationships within the system that generated those results. By the time performance metrics signal trouble, the underlying conditions may have been developing for some time. Rachlin and Hyman believe that organizations can benefit from complementing performance monitoring with attention to system behavior, interaction patterns, and evolving dependencies that influence resilience long before disruption emerges.
Rachlin explains, "Resilience starts to fail long before a disruption becomes visible. Organizations often strengthen their future when they develop the ability to understand how their systems are changing while those changes are still manageable." This philosophy aligns with Hyman's Marginal Point of Systemic Drift (MPOSD) framework, which explores whether specific patterns indicate that governance visibility is becoming less reliable before operational consequences become apparent. Instead of attempting to predict every future event, the framework focuses on structural signals that indicate when a system is becoming increasingly difficult to evaluate independently.
Rachlin and Hyman have identified five recurring indicators that appear together across multiple complex-system scenarios. The first indicator, verification integrity degradation, reflects situations where system outputs evolve more quickly than independent verification processes can keep pace. For example, if an AI model updates its behavior faster than the testing suite can validate, the system may drift into unsafe territory unnoticed. The second indicator, proxy substitution escalation, occurs when alerts, reviews, or operational indicators no longer provide an accurate representation of system activity. This means the metrics used to gauge health become disconnected from actual performance.
The third indicator, incentive-proof misalignment, describes circumstances in which the system has limited structural incentive to reveal its own drift. For instance, if an AI system is rewarded for maximizing uptime but not for reporting anomalies, it may hide early signs of trouble. The fourth indicator, latency inflation and feedback distortion, emerges when delays between action and visibility become increasingly meaningful for decision-makers. As feedback loops lengthen, leaders may base decisions on outdated information. Finally, governance independence erosion develops when oversight mechanisms rely on the same systems they are intended to evaluate. This creates a circular dependency where the overseer is part of the problem.
According to the duo's observations, these signals become especially meaningful when they converge rather than appearing in isolation. Hyman says, "Complex systems rarely become difficult to govern in a single moment. Governance changes when independent visibility begins to narrow, and recognizing that transition may create valuable opportunities for informed decision-making." The importance of independent visibility has been highlighted by recent AI incidents. In one case, an autonomous coding agent deleted production data and backups within seconds after operating outside its intended boundaries. Hyman and Rachlin's retrospective application of MPOSD suggested that observable indicators—such as verification integrity degradation and proxy substitution escalation—may have appeared before the irreversible stage of the sequence. While retrospective analysis cannot establish future outcomes, the duo believes the incident illustrates how identifying structural changes earlier could expand the range of governance decisions available before disruption occurs.
The MPOSD framework draws on concepts from systems theory, cybernetics, and organizational learning. It builds on the work of earlier researchers who studied drift in high-risk environments such as nuclear power plants and aviation. In those domains, operators learned to monitor for subtle signals that preceded failures—like increased communication delays or the bypassing of standard procedures. Rachlin and Hyman have adapted these insights for the AI era, where algorithmic systems can drift at machine speed. Their five indicators provide a structured way for leaders to assess whether their governance approaches are keeping up with technological change.
For many organizations, adopting this perspective requires a cultural shift. Traditional IT governance often relies on command-and-control structures where decisions flow from top to bottom. But as AI systems become more autonomous, governance must become more adaptive. Leaders need to foster an environment where teams feel empowered to raise concerns about drift without fear of blame. This aligns with the concept of "psychological safety" popularized by organizational behavior researchers. When teams can openly discuss anomalies, they are more likely to catch early signs of systemic drift.
Technology itself can also play a role in monitoring drift. Observability platforms, originally developed for microservices, can be extended to track AI system behaviors and their interactions. For example, distributed tracing can reveal latency inflation, while automated anomaly detection can flag proxy substitution escalation. However, Rachlin and Hyman caution against relying solely on automated tools. The human element remains critical for interpreting signals and making judgment calls about which drifts are acceptable and which require intervention.
The implications for executive leadership are profound. Boards and C-suites that traditionally focused on financial KPIs now need to develop a layered understanding of operational health. Dashboards remain meaningful, but increasingly interconnected AI ecosystems also require monitoring the relationships that link systems together. Independent assessment of governance health—viewed separately from the systems under evaluation—can provide additional context that supports more informed decisions as complexity increases. Rachlin notes, "AI is likely to keep growing its presence in enterprise settings, opening up fresh possibilities while also raising new questions about how organizations manage and guide its use. The technology can offer strong capabilities, but a company's ability to stay resilient may also hinge on noticing shifts early before they turn into bigger operational challenges."
As Hyman and Rachlin's work suggests, anticipating systemic drift may complement traditional governance in ways that support more informed leadership decisions. Organizations that continue developing their capacity to recognize early signals alongside responding thoughtfully to visible outcomes may help define the next chapter of innovation with greater confidence and resilience.