Share with your CDO
Data ambiguity, not just data dirtiness, is quietly killing enterprise AI deployments before they reach production, according to Anand Ramamoorthy, Director of APAC Data Governance and Quality at Informatica. Speaking after the company’s Data and AI Summit in Mumbai, Ramamoorthy cited one enterprise that flagged 73 percent of its data as bad, a figure that crystallizes why so many AI agent pilots stall at the production threshold. His proposed fix centers on machine-readable metadata, AI-ready data products, and AI-assisted data stewards to carry governance context into agentic workflows.
What this means for your business
The 73 percent figure is striking, but the more dangerous insight is subtler. Ramamoorthy’s distinction between incorrect data and ambiguous data matters enormously for anyone running or funding AI agents. An agent trained on ambiguous data doesn’t fail loudly with an error message; it produces plausible-sounding outputs that diverge from intent in ways that can take weeks to trace. If your organization has invested in agent development without first auditing how context-dependent your core datasets are, you’re almost certainly measuring the wrong failure mode.
Ramamoorthy’s three-pillar framework, machine-readable metadata with trusted context, AI-ready data products, and AI-assisted stewardship, is reasonable on its face, though Informatica’s commercial incentive to make governance feel infrastructurally mandatory nudges the prescription toward platform investment rather than process redesign. That tilt is worth noting because the same outcome, stewards focused on judgment rather than manual tagging, is achievable by retraining existing teams and tooling without a full platform overhaul. The real constraint most enterprises face isn’t a missing product category; it’s organizational patience. Governance programs erode when business leaders don’t see value inside a quarter, which is exactly the dynamic Ramamoorthy names and then proposes to solve with more tooling.
The CDO whose data stewards are still doing manual, row-level remediation is the one with the most exposure here. AI amplifies whatever interpretation is baked into the data pipeline, so ambiguity that a human analyst once caught through domain knowledge becomes a systematic hallucination at scale. The decision this reframes isn’t whether to buy a governance platform; it’s whether your current stewardship model can encode business context in a form an agent can consume. If the answer is no, that’s a workforce design problem as much as a technology one, and it won’t be solved by renewal of a data catalog license alone.
Concept deep-dive: Machine-readable metadata
Metadata is data about data, the label on the jar rather than the contents. Machine-readable metadata means those labels are structured so a software agent can interpret them autonomously, without a human explaining what “revenue” means in a given regional context. Think of it as the difference between a handwritten sticky note and a standardized field in a database. For AI agents that must make decisions across multiple data sources, the absence of this structured context is where ambiguity enters and hallucination follows.
Based on reporting from Poor data quality is breaking AI ambitions: Anand Ramamoorthy, Director APAC Data Governance and Quality, Informatica, Salesforce, originally published 2026-06-16 11:26:00.
