AI Agents Don't Fail in the Model. They Fail in the Data.

At an executive AI steering committee, the agent answered with total confidence. The revenue figure was wrong. Nobody caught it until a downstream workflow adjusted pricing, a report went to the board, or a customer received an incorrect answer. The large language model did not malfunction. The context it was given—definitions, lineage, timeliness, and master data consistency—was not fit for autonomous action.

Organizations are racing to deploy agentic AI: systems that analyse live data, trigger workflows, and act with limited human oversight. Yet most investment still flows to model selection, prompt engineering, and retrieval infrastructure. Far less flows to the data management disciplines that determine whether an agent can be trusted at scale: Metadata Management, Reference and Master Data, Data Quality, and Data Governance.

Part 1 establishes why data—not the model—is the control plane for AI agents, using terminology from DAMA-DMBOK 2 and IBM's data management practice. Part 2 (publishing July 25, 2026) covers real-world news examples, the governed context layer beyond RAG, and a ten-question readiness checklist.

The Shift to Agentic AI Changes the Data Stakes

Traditional analytics could tolerate imperfect data because a human interpreted the output. Agentic AI compresses the window between data consumption and action. IBM's data management guidance notes that AI-powered agents are designed to act autonomously—analysing live data, triggering automation workflows, and responding faster than human review allows [2]. IDC projects that a large majority of agentic AI use cases will require real-time, contextual data access [2].

When inputs are stale, siloed, or inconsistently defined, agents do not pause to interrogate them. IBM warns that "data pipelines and agentic AI systems are built to act on data, not interrogate it"—data can arrive correctly formatted and still reflect inaccurate conditions [3]. The automation executes on a premise that may no longer be true.

DAMA-DMBOK 2 frames the underlying risk plainly: low-quality data represents risk not only because information is wrong, but because it can be misunderstood and misused [4]. Agents multiply that misuse across every connected workflow.

AI-Ready Data Is Not the Same as "Good Enough" Data

IBM defines AI-ready data as high-quality, accessible, and trusted information that organizations can confidently use for AI training and initiatives—accurate, complete, and consistent, with governance and privacy controls embedded in pipelines [5]. Yet IBM's own 2024 survey found only 29% of technology leaders strongly agreed their enterprise data meets the quality, accessibility, and security standards needed to scale generative AI efficiently [5].

IBM also cites a Gartner analyst observation that AI-ready data must be representative of the use case—including patterns, errors, outliers, and unexpected emergence needed to train or run the model for a specific purpose [6]. Data refined for traditional reporting (outliers removed, values conformed to human expectations) may fail that standard for AI, even when it scores well on conventional quality dashboards.

The DAMA Data Stack Agents Depend On

DMBOK 2 describes interconnected data management functions that together enable trustworthy automation. For AI agents, four areas matter most:

1. Metadata Management

Metadata includes business, technical, and operational descriptions of data—definitions, lineage, ownership, and quality rules. Without reliable Metadata, an organization "does not know what data it has, what the data represents, where it originates, how it moves through systems, who has access to it, or what it means for the data to be of high quality" [1]. Agents need this context at inference time—not buried in tribal knowledge.

Agent implication: Retrieval-augmented generation (RAG) alone cannot fix inconsistent business definitions. Agents need certified definitions, lineage, and impact analysis—the deliverables Metadata Management is designed to produce [1].

2. Reference Data and Master Data

DMBOK 2 classifies Reference Data and Master Data as distinct types with distinct lifecycle requirements [7]. When Finance and Marketing maintain different definitions of "active customer," or when product codes diverge across ERP and CRM, agents inherit the inconsistency and automate conflicting outcomes.

Agent implication: Master Data Management (MDM) and governed Reference Data are prerequisites for agents that execute cross-system workflows—not optional master-data cleanup after go-live.

3. Data Quality Dimensions

IBM measures data quality against criteria including accuracy, completeness, validity, consistency, uniqueness, timeliness, and fitness for purpose [8]. DAMA defines a Data Quality dimension as a measurable characteristic that provides vocabulary for requirements and ongoing measurement—directly connected to risk in critical processes [9].

For agents, timeliness and consistency are especially stressed: IBM notes that staleness is harder to detect than structural schema errors, yet equally dangerous when agents act on outdated customer, inventory, or market data [3].

4. Data Governance

Data Governance defines policies, standards, and accountability for data collection, access, use, and lifecycle management. IBM positions governance as the mechanism that transforms enterprise data into trustworthy AI-ready assets—with lineage, provenance documentation, and privacy controls [5]. DMBOK 2 ties governance to regulatory expectations that information appear on the risk register with appropriate mitigations [4].

Agent implication: Autonomous agents acting on PII or regulated data without governed access, consent metadata, and audit trails amplify compliance and reputational risk.

Where Part 1 Leaves Off

The framework is clear: agents need Metadata, Reference and Master Data, measured Data Quality, and Governance before autonomy scales. Knowing the theory, however, is not the same as seeing the consequences—or building the architecture to prevent them.

In Part 2, we examine what 2026 headlines are already proving, why a governed context layer must sit beyond RAG, and ten questions every CDO should ask before the next agent goes to production.