Data Classification and Grading: The Decision Chain from General Data to Important Data

Data classification and grading is often treated as documentary work: find a template, list the table names in the systems, mark them ‘sensitive’ or ‘general’, and file it away. Its actual function is to fix the starting point of every downstream obligation — whether data may leave China, whether an impact assessment is required, what an audit examines, and which MLPS level a system takes. Get the grading wrong and the later procedures are either wasted or abandoned halfway.

Three concepts, not one

The commonest confusion in practice is to treat three concepts as a single one.

Personal information and important data are not parallel categories but intersecting ones. A single dataset may be both personal information and important data; conversely, ordinary business data may fall within important data by reason of scale or sector. Grading applies to all data, not to personal information alone.

The four-step decision chain

First, inventory. List by system and business action rather than by department. Registration, content generation, payment, customer service, logs, model training and data export each correspond to a group of data. A departmental listing omits the intermediate data that flows across departments — and that omitted part is often precisely the carrier of outbound transfer and sharing.

Second, classify. Group by business attribute. The usual four groups are user data, operational data, technical data and public data. Classification allows one set of rules to govern a whole group, avoiding table-by-table judgments.

Third, grade. Grade by affected object — individual rights, enterprise operations, public interest and national security — and then by degree of impact into general, important and core. Determining the affected object is harder than determining the degree, because the same data means different things to different objects.

Fourth, determine important data. Compare against regional and departmental catalogues; where no catalogue applies, apply the identification guidance and record the reasoning in writing. What matters at this step is not the conclusion but a basis capable of review.

Four categories AI companies overlook

Training corpora. Purchased, crawled, user-generated and open-source datasets carry different authorisation chains. They are often graded simply as ’text’, overlooking the personal information and third-party copyright content they may contain.

User input and uploads. Product design may collect only account and log data, but users paste medical records, identity cards and contracts. The test is actual processing, not design intent.

Logs and tracking. A single log entry is not sensitive, but aggregated behavioural trails can identify an individual. Grade by the identification capability after aggregation, not by the individual field.

Model outputs, caches and vector stores. A fine-tuned model may reproduce training data, and a retrieval-augmented vector store retains user documents for long periods. These two are often absent from the asset inventory, yet they are among the first points raised in regulatory enquiries.

Where the results must land

First, the data inventory, which should carry at least seven fields: category, level, source, storage location, processing purpose, retention period and sharing recipients.

Second, the outbound transfer decision. Whether data constitutes important data, and the volume and sensitivity of personal information, determine whether the security assessment, the standard contract or certification applies; see the three routes for data export.

Third, impact assessment and compliance audit. The grading result is the working paper that identifies the trigger and scope; see personal information protection impact assessment and the compliance audit measures.

Fourth, risk assessment and MLPS. The annual risk assessment of important data processors and the MLPS grading of systems both take the grading result as input.

Fifth, suppliers and contracts. Only by writing the data level into procurement contracts and data processing agreements can downstream parties be bound.

Three common mistakes

Grading personal information only. Operational and technical data are left without a level, so outbound transfer and sharing cannot be judged and the identification of important data loses its foundation.

Equating important data with confidential data. The test for important data is impact on national security, not confidentiality classification. Publicly available industry statistics may constitute important data through aggregation effects.

Treating it as a one-off exercise. Product iterations, changes of data source and overseas expansion all alter the conclusion. Without a review mechanism the inventory becomes stale within six months.

How long it takes

For a team of fewer than ten people, the first inventory and grading typically take two to four weeks, provided that a system list and data-flow diagrams already exist. Where there is no data-flow diagram, draw one first; otherwise the output is a list of table names rather than a map of data assets. Maintenance thereafter costs far less than the first round, but an owner and a review cycle must be fixed.

This article is based on the Data Security Law, the Regulations on Network Data Security Management and publicly available important data identification rules. It is for reference only and does not constitute legal advice. Grading should be determined by reference to the applicable sectoral catalogues and the company’s actual data processing activities.