Sneak Peek Q&A: Why AI governance breaks down in production -- and what comes next

As enterprises accelerate AI deployment, governance frameworks have not kept pace with the systems they are meant to control.

Following his session at the BrightTALK Cloud Convergence Summit, thought leader Varun Raj dives deeper into why existing AI governance models are fundamentally misaligned with modern systems and proposes a control-plane approach to restore accountability, reliability, and system-level oversight in production environments. While organizations have made significant progress in model development and deployment, the mechanisms used to govern these systems remain largely static, fragmented, and extrinsic.

Drawing on experience designing and operating large-scale enterprise AI platforms, Varun Raj argues that governance must evolve from a policy-layer construct into an integrated runtime capability -- one that operates continuously and directly within the execution path of the system.

Read on for a featured expert Q&A from Varun Raj’s session.

Viewers can register for the full webinar to learn more.

AI capabilities are advancing rapidly. Where are current governance approaches falling short?
Varun Raj: The core issue is that governance models have remained static while the systems they are meant to govern have become dynamic. Most existing approaches assume predictable behavior and infrequent change. AI systems violate both assumptions. They continuously adapt, interact, and evolve within broader system contexts. That mismatch creates blind spots that governance frameworks were never designed to address.

You’ve described a structural gap between AI systems and governance mechanisms. How does that manifest in production environments?
Raj: In production, systems often meet all observable health indicators, uptime, latency, throughput, yet still produce outcomes that are inconsistent, misaligned, or contextually incorrect. This is not a failure of infrastructure; it is a failure of control. The system is functioning, but it is not behaving as intended. That distinction is critical and largely unaddressed in current governance models.

Why are traditional governance models fundamentally insufficient for modern AI systems?
Raj: Traditional models rely on defining policies externally and validating compliance periodically. That approach assumes behavior can be bounded ahead of time. AI systems, especially those embedded in workflows, operate under changing conditions and inputs. Governance that is external and retrospective cannot keep pace with that level of dynamism.

What kinds of risks emerge from this misalignment?
Raj: The most significant risk is not system failure, it is undetected behavioral deviation. Systems can produce outputs that are subtly incorrect, non-compliant, or misaligned with intent without triggering alerts. Over time, these deviations accumulate, leading to systemic risk that is difficult to trace or remediate.

How does this differ from risk in traditional distributed systems?
Raj: In traditional systems, failure is observable and often localized. In AI systems, risk is distributed and behavioral. The system continues to operate, but the integrity of its decisions degrades. This creates a class of failure that is both persistent and difficult to detect using conventional observability tools.

You’ve introduced the Control Plane AI Governance Model (CPAGM). What is the core idea behind it?
Raj: The core idea is to treat governance as a system capability rather than an external function. CPAGM embeds control mechanisms directly into the execution path, enabling systems to evaluate behavior continuously, enforce policies in real time, and intervene when deviations occur. It shifts governance from observation to active control.

Why is the control plane the right place to enforce governance?
Raj: The control plane is where decisions about execution happen. Placing governance there ensures that control is applied before outcomes propagate through the system. External monitoring can tell you what happened; the control plane determines what is allowed to happen.

What are the foundational capabilities required to implement this model?
Raj: Three capabilities are essential: continuous visibility into system behavior, real-time evaluation against expected outcomes, and deterministic intervention mechanisms. These must operate as part of the system itself, not as add-ons.

What are the primary challenges organizations face when adopting this approach?
Raj: The main challenge is that existing infrastructure is not designed for this level of control. Additionally, governance responsibilities are fragmented across teams, which makes coordinated implementation difficult. Addressing this requires both architectural changes and organizational alignment.

What trade-offs come with embedding governance into the execution layer?
Raj: There is increased system complexity and some additional operational overhead. However, these costs are offset by improved reliability, stronger risk control, and the ability to operate AI systems with confidence at scale. Without these capabilities, organizations are effectively operating without full control.

How does this approach impact compliance in regulated industries?
Raj: It enables organizations to demonstrate that governance is not only defined but enforced in real time. This closes the gap between policy and execution, which is a critical requirement for auditability and regulatory assurance.

As AI systems become more autonomous, what role do platform teams play?
Raj: Platform teams become the primary stewards of system behavior. Their responsibility extends beyond infrastructure to ensuring that systems operate within defined boundaries. Governance becomes an inherent property of the platform, not an external overlay.

Varun Raj is an enterprise AI and cloud platforms leader whose work focuses on system-level governance, runtime control, and the operational integrity of large-scale AI systems. His contributions center on embedding accountability and enforcement mechanisms directly into distributed AI architectures, addressing a critical gap between model capability and system control. His perspectives on AI governance and cloud architecture have been widely featured across TechTarget, CIO.com, and VentureBeat, and he is a frequent speaker on the evolving challenges of operating AI systems in production environments.

Alicia Landsberg is senior managing editor on the BrightTALK summits team. She previously worked on TechTarget's networking and security group and served as senior editor for product buyer's guides.