Share with your CDO
Agentic AI systems, those that don’t just answer questions but plan, retrieve data, and trigger actions autonomously, inherit every data quality problem already living in your enterprise stack. Anjali Garg’s case in TDWI is that data quality must shift from a reporting hygiene function to an active control layer built directly into agentic workflows, complete with real-time checks, confidence-based routing, and lineage that’s visible at decision time rather than catalogued and forgotten.
What this means for your business
Where you sit on this question depends almost entirely on how far your organization has moved from AI assistants toward AI agents with write-back authority. If your agents can only read and recommend, bad data produces a wrong answer a human catches. If they can route a case, approve a transaction, or escalate a compliance flag, bad data produces an operational error that may not surface until after it compounds. The gap between those two states is where data governance programs are dangerously underbuilt right now.
The argument holds, but Garg, writing for an analytics training and advisory organization with a natural interest in elevating data management’s strategic standing, frames the solution almost entirely through data product design and stewardship workflows. That’s not wrong, it’s just incomplete. The harder organizational problem is that data teams rarely own the agentic workflow layer, AI and engineering teams do, and those teams make retrieval and orchestration decisions daily without a data quality gate in sight. The control plane metaphor is apt precisely because control planes are infrastructure-level, not departmental. Treating this as a CDO mandate without a corresponding architecture mandate at the CTO level leaves the frame without the enforcement mechanism.
The leading indicator worth watching is escalation rate by data-quality reason, the metric Garg names but most AI programs aren’t tracking yet. Organizations that instrument that signal first will know whether their next AI investment dollar belongs in model tuning or master data management. Those that don’t will keep attributing operational failures to the model when the data environment is the actual constraint, a misdiagnosis that’s expensive and self-reinforcing. If your AI program’s postmortems don’t yet distinguish between model errors and data errors, that’s the budget conversation worth having before the next agent goes to production.
Concept deep-dive: Confidence routing
Confidence routing is a decision rule that governs what an AI agent is allowed to do based on the quality of the data it’s acting on, think of it as a graduated permission system tied to data reliability rather than a binary on-off switch. High-quality data in a low-stakes workflow allows full automation. Degraded data or high-stakes outcomes trigger human review or block execution entirely. It replaces the common but brittle design where agents either act with false certainty or halt completely, and it’s the mechanism that makes partial automation safe to scale.
Based on reporting from Data Quality Is the Control Plane for Enterprise Agentic AI, originally published 2026-08-05 08:29:00.
