The Enterprise AI Governance Gap Is a Data Problem First

/ AI, AI Governance, Application Development, Data Management

The AI Governance Gap Is a Data Problem First

Quest Software’s Global Field CTO Susan (Sue) Laine joined ECI Research’s Paul Nashawaty on the AppDevANGLE podcast to discuss why so many enterprise AI initiatives stall between proof of concept and production. The conversation centers on a core tension: organizations are accelerating AI adoption while simultaneously neglecting the data foundations those systems depend on. Laine argues that without trust, visibility, and a strong control plane around enterprise data, AI doesn’t just underperform; it fails in ways that are difficult to detect and expensive to reverse.

The Bigger Picture

The Governance Gap Is Wider Than Most Organizations Admit

The headline statistic from this conversation is a damning one. According to ECI Research, 50.7% of organizations rely on public AI tools such as ChatGPT and Copilot, while only 20.2% report enterprise-wide AI deployments built on a governed framework. That isn’t a maturity gap waiting to close naturally over time. It’s an active risk posture. Organizations are running consequential AI workloads on consumer-grade tools while their governance infrastructure lags by years, not quarters.

Laine’s observation from a recent conference makes this concrete: in a room of thousands of practitioners, fewer than ten could identify who within their organization was responsible for AI. Ownership without accountability is one of the most reliable predictors of governance failure, and that pattern is playing out across industries in real time.

The implications for ITDMs are direct. An AI initiative running on ungoverned data isn’t just inefficient; it’s a liability. As agentic AI systems scale the volume and velocity of data consumption far beyond what any human analyst could produce, the small discrepancies that once got caught in conversation now propagate at machine speed. Laine’s framing is precise: humans working through BI reports apply judgment, ask colleagues, and resolve discrepancies organically. Agents don’t. They process orders of magnitude more requests without that corrective feedback loop, which means bad data scales as fast as good data.

What Data Readiness Actually Requires

Laine draws a sharp distinction between organizations that treat data governance as a compliance checkbox and those that treat it as an operational capability. The mature organizations, typically large financial institutions that have been managing structured data under regulatory scrutiny for decades, are now using AI to accelerate governance itself. JPMorgan Chase’s situation is instructive: with over 31,000 data models and a rule requiring everything to be modeled before reaching production, the traditional process became a bottleneck. Their response was to use AI to generate data specs, identify the best source data, and attach a regulatory control plane to each new data product automatically.

That’s the model worth studying. AI isn’t replacing governance; it’s being used to make governance fast enough to keep pace with development velocity. Organizations still treating data management as a back-office function will find that their AI ambitions are effectively capped by the quality and traceability of their underlying data.

For developers, the semantic layer discussion is particularly relevant. Laine cites a study showing that answering a business question like “what is our most profitable customer segment” without a semantic layer requires 5,000 to 15,000 tokens across three or four attempts. With a well-constructed semantic layer including defined metrics, trust-scored data sources, and clear dimension definitions, the same question resolves in one or two attempts at 500 to 2,000 tokens. At enterprise query volumes, the cost and latency difference compounds quickly. This is a concrete argument for investing in data modeling and semantic infrastructure before scaling agentic systems, not after.

The Operational Drag No One Is Measuring

ECI Research’s 2025 AI Builder Summit survey found that 44% of enterprise AI leaders have only moderate confidence that AI agents can act autonomously without human intervention. That finding pairs uncomfortably with the operational reality Laine describes: models that perform well in test environments frequently degrade in production because the production data is messier, less governed, and less consistent than the controlled data used during development.

ECI Research data reinforces the scale of this problem. According to our research, 43.8% of AI/ML teams lose one to two weeks per project annually to compute efficiency challenges, while 28.4% lose two to four weeks, and 6.1% lose more than two months. These aren’t tooling failures. They’re strategy failures. Teams that haven’t invested in data infrastructure pay for it in velocity, and the cost is invisible to the managers reviewing outcomes rather than experiencing the day-to-day friction.

The practitioner-manager perception gap that surfaces in the conversation is real and worth naming explicitly. ECI Research found that 28% of practitioners report that production AI models require daily retraining, compared to approximately 14% of managers. That gap exists because managers track outcomes; practitioners live inside the operational loop where model instability is a continuous condition, not an episodic event. Closing that gap requires organizations to surface practitioner data upward into leadership visibility, not just report on business KPIs that lag the underlying problems by weeks or months.

Governance Frameworks Built for Change, Not Just Compliance

Most enterprise governance frameworks are designed around today’s regulatory requirements, which means they’re already behind the curve on EU AI Act provisions, U.S. executive orders, and evolving NIST guidance. A framework architected only for compliance will break every time the regulatory environment shifts, which is frequently.

The durable alternative is a control plane that wraps individual data products rather than the entire data estate as a monolith. Quest Software’s approach, using AI to generate logical data definitions, score data quality, and attach regulatory metadata at the product level, makes governance portable and composable. Each data product carries its own governance context rather than depending on centralized human review that doesn’t scale. That architecture can adapt to new regulatory requirements by updating the control definitions rather than rebuilding the entire governance layer.

What’s Next

AI Readiness Becomes a Procurement Criterion

This conversation covered a shift that ITDMs should anticipate: AI readiness assessment is becoming a vendor selection and procurement criterion, not just an internal capability audit. As organizations recognize that data quality and governance directly determine whether AI investments deliver measurable returns, vendors that can demonstrate governed data pipelines, semantic layer support, and regulatory traceability will have a structural advantage over those offering raw model access without infrastructure guardrails.

The Semantic Layer Moves from Nice-to-Have to Infrastructure

For development teams, the semantic layer is likely to move from an optional architectural component to a baseline infrastructure requirement as agentic AI systems scale. The token efficiency argument Laine presents is compelling, but the larger argument is about reliability: agents working against well-defined, trust-scored data products produce consistent, auditable outputs. Agents working against raw, unstructured data produce outputs that are difficult to validate and impossible to govern at scale. Organizations that build semantic layers now are positioning themselves ahead of what will become a standard requirement as AI workloads mature from pilot to production at scale.

Author