High-Quality Datasets: The New Data Governance Battlefield After DCMM 2.0

AI Doesn't Need More Data—It Needs Better Data

The previous article analyzed changes in DCMM 2.0. While the standard's clauses emphasize data assets, data application circulation, data compliance management, and data culture, placing it in the "AI+" context reveals a more critical direction: data governance is shifting from "governing data assets" to "building high-quality datasets."

Traditional data governance focused on data standards, quality, metadata, master data, indicator definitions, and security. These remain important, but the AI era demands more stable, trustworthy, explainable, and machine-friendly data supply for model training, knowledge augmentation, agent execution, and industry LLM applications. This aligns with DCMM 2.0's "intelligent" and "trustworthy" directions: AI needs not just more data, but data that models can stably use, businesses can continuously verify, and compliance can constrain.

Many enterprises have abundant data from years of digital transformation—business systems, data platforms, warehouses, lakes—but most cannot be directly used for AI. Business system data serves processes; warehouse data serves analytics; AI requires a different set of criteria:

Clear provenance

Consistent definitions

Accurate labels

Explicit semantics

Compliant permissions

Assessable quality

Traceable versions

Support for training, evaluation, retrieval, and inference

The real scarcity in AI adoption is not data existence but high-quality datasets that models can stably consume.

High-Quality Datasets Are Data Products, Not Just Table Collections

A common misconception equates high-quality datasets with cleaned and labeled data. That is only part of the story. A true AI-facing dataset must answer five questions:

First, is the source trustworthy? The originating system, business step, time range, and processing steps must be traceable; otherwise model issues cannot be root-caused.

Second, is the semantics clear? The same field, metric, or business object may carry different meanings across systems. Without unified semantics, models learn historical system noise instead of business knowledge.

Third, is quality measurable? Completeness, accuracy, consistency, timeliness remain vital. For AI, additional dimensions include annotation quality, sample coverage, class balance, noise ratio, and bias risk.

Fourth, is usage compliant? Training, knowledge-base construction, Q&A, and agent invocation all involve data-use boundaries. Rules must define what data can be used, what cannot, and what requires de-identification.

Fifth, is versioning traceable? Model performance shifts often correlate with training, evaluation, or knowledge-base data versions. Without dataset versioning, it is hard to explain why a model improved or degraded.

Together, these capabilities turn a dataset into a manageable, assessable, traceable, operable data product—requiring continuous build, release, use, and optimization cycles, not one-off preparation.

Multimodal Data Makes Governance More Complex

Traditional governance centered on structured data: tables, fields, metrics, code values, master and reference data. AI applications need far more: text, images, audio, video, scans, drawings, sensor time-series, logs, knowledge documents, process records.

This raises new governance questions:

How to link transcribed text to original audio?

How to extract key information from contracts, medical records, judgments, transcripts, equipment reports?

How to represent objects, events, states, and relationships behind images, video, sensor streams?

Field-level governance cannot fully solve these. Multimodal governance requires managing not just formats but the objects, events, states, relationships, rules, and context behind the data—driving the growing importance of ontologies, knowledge graphs, and industry semantic platforms.

Data Governance Must Evolve into AI Data Engineering

Classic data governance revolved around:

<ol>
<li><code>How to define data standards?</code></li>
<li><code>How to check data quality?</code></li>
<li><code>How to manage metadata?</code></li>
<li><code>How to unify metric definitions?</code></li>
<li><code>How to control data permissions?</code></li>
</ol>

These remain foundational. For AI, the questions extend further:

<ol>
<li><code>Which data suits training?</code></li>
<li><code>Which data suits evaluation?</code></li>
<li><code>Which data belongs in the knowledge base?</code></li>
<li><code>Which data can be invoked by Agent?</code></li>
<li><code>Which data carries bias, noise, or compliance risk?</code></li>
</ol>

This means data governance must advance from traditional data management to AI data engineering. AI data engineering is not mere cleaning or a standalone labeling platform; it is a continuous data supply system built around model training, knowledge augmentation, agent execution, and business applications. Its goal is not cleaner-looking data but data that models can actually use, businesses can verify, and risks can control.

High-Quality Datasets Also Need a Semantic Foundation

High-quality datasets cannot rely solely on manual annotation and rule-based cleaning. Without business semantics, datasets remain mere sample collections. In courts, public security, manufacturing, healthcare, energy, data records are interconnected:

A record maps to a business object.

A metric maps to a statistical definition.

A document maps to a process node.

An alert maps to equipment state, process rules, risk thresholds.

If these objects, rules, relationships, and contexts are not modeled, datasets cannot truly express domain knowledge. Hence, dataset construction requires a semantic foundation—ontologies and industry semantic platforms that define business objects, unify concepts and definitions, express relationships, embed business rules, and connect structured, document, and multimodal data so models and agents understand the business meaning behind the data. From this view, high-quality datasets are not a byproduct of governance but a core asset for industry AI capability.

Summary

DCMM 2.0 pushes data governance toward assetization, business alignment, intelligence, and trustworthiness. In AI implementation, this direction converges on a single critical question:

<ol>
<li><code>Can the enterprise continuously build and operate high-quality datasets?</code></li>
</ol>

Without high-quality datasets, AI stalls at demos, Q&A experiences, and limited pilots. With them, model training, knowledge augmentation, intelligent Q&A, agent execution, and industry LLM deployment gain a stable data foundation. Therefore, beyond DCMM 2.0, the new battlefield is not just data-asset accounting, data-product circulation, or platform upgrades—the longer-term, more decisive battlefield is building high-quality datasets for AI.