The CDO's new role is curating context for data governance

The CDO and agentic AI

When AI models were primarily driving chatbots or text generation, errors and hallucinations were largely mitigated by human users who interacted with the system. But today, the pressure on CDOs originates in what agentic systems can do. For teams used to delivering enterprise data to human decision-makers through reports and dashboards, the autonomy of AI agents is a different problem entirely.

An AI agent given a business goal -- rebalancing the quarterly marketing budget, drafting a compliance report, routing a vendor contract -- breaks that goal into subtasks, retrieves information from across enterprise systems and executes actions without waiting for a human to evaluate each step. As a result, errors in an agentic workflow compound across steps and may not surface until the action is already complete.

The CDO's traditional toolkit was not built for this. Conventional data governance answered specific questions: what data exists, where does it live, is it accurate, and who is authorized to see it? Data catalogs and master data management  programs answered those questions well for structured data in reporting environments, where a human analyst received the output and applied judgment.

Context architecture for agentic AI must answer a different question: which data, in what combination, with what business definitions, should an AI system retrieve at the time of query to produce a correct result? Context is the defining challenge here.

Context and the CDO

A data catalog shows users that policy documents exist in a given folder. It does not tell an AI agent which of several versions of that document is authoritative, or that the finance team's definition of revenue excludes deferred recognition while the sales team's does not. Increasingly, CDOs must deliver and maintain a semantic layer to capture those distinctions.

For example, an agent reviewing a supplier evaluation may read from an enterprise knowledge base that has versions from both 2023 and 2026. How does it know that the 2026 version is authoritative, especially if it has not been given a rule to prefer the most recent version? The agent may use both versions to draft a report that's clear, well-written and professionally formatted. The output contains no hallucinations: the agent executed perfectly, but on unvalidated context. The distinction is important because these two kinds of issues point to different parts of the organization.

Accountability for AI outputs depends on someone owning the context layer. In too many enterprises, no single executive owns that authority. If marketing spins up customer sentiment agents while finance builds forecasting agents, then each may apply different definitions of the same business terms, different standards for what counts as an authoritative source and different thresholds for data freshness.

The CDO's concern is governing what agents retrieve, a business-critical mandate defined by the following four capabilities:

• Inventory. An enterprise ontology or knowledge graph maps every data asset while recording relationships between documents, databases, agents, and the sources that feed them. This may include vector databases or unstructured data that traditional data catalogs did not reach. Agents should not retrieve what they cannot understand.
• Access control. Agents can accidentally combine information that humans would never be permitted to view in a single session. Application-layer security, such as the permissions Salesforce enforces on CRM records or SharePoint enforces on documents, governs what a human user can open when they log into a specific system. But an AI agent may simultaneously retrieve and assemble data from multiple systems. Context-layer access control prevents an agent from combining data that no individual human would have been permitted to combine.
• Governed delivery. Context must reach agents without centralizing every system. Model Context Protocol (MCP) -- an open standard with rapidly growing adoption across AI vendors -- enables an agent to query enterprise data sources and receive governed, logged and auditable responses in real time, without the organization first moving all that data into a centralized store.
• Observability. Continuous monitoring of how agents behave, what they retrieve and why they make specific choices. Of the four capabilities, this one most deserves a close look.

Observability

For both auditability and continuous development, organizations must understand why an agent made particular choices. Traditional data observability, tracking data quality as it moves through pipelines, and standard LLM observability, tracking prompts and token usage, are no longer sufficient on their own. The table below shows the differences between these forms of observability,  illustrating the changing demands on the CDO.

The semantic layer

The semantic layer translates governed data into AI-readable context and is critical to all four capabilities. It tells the agent retrieving revenue data whether that term means gross revenue, net revenue or revenue recognized under the organization's specific accounting policy. The semantic layer explicitly records these distinctions, makes them consistently available across data systems and delivers data to the agent with that precision intact. Without semantics, agents guess the meaning of business terms based on their model's general training, rather than on specific business vocabulary.

As Fern Halper, senior research director of TDWI, put it, effective organizations "formalize an enterprise business vocabulary and explicitly encode business rules and logic so that models operate within defined guardrails rather than inferred assumptions." Human validation of those definitions remains part of that process, because the semantic layer is only as reliable as the business logic people have approved and recorded in it. If an agent measures its performance against KPIs that influence operational or strategic decisions, someone needs to verify that those KPIs are properly defined and managed.

The new CDO

CDOs who can build and manage this infrastructure of context transform their businesses. An enterprise with a mature context layer dramatically shortens the time required to deploy new agents safely. The CDO who can own this infrastructure becomes the executive who manages what AI knows about the business. That sets the boundary of what the enterprise is capable of.

Donald Farmer is a data strategist with 30+ years of experience, including as a product team leader at Microsoft and Qlik. He advises global clients on data, analytics, AI and innovation strategy, with expertise spanning from tech giants to startups. He lives in an experimental woodland home near Seattle.