What the governing body should require for AI data governance and classification

Authority disclaimer. This page is an executive narrative. It does not establish requirements; the linked corpus governs. It is provided to support understanding, discussion, and decision-making by the governing body and accountable executive leadership (board, ELT, or senior management, as applicable). It creates no obligation, control, or assurance by itself. Its publication approval is an editorial act only and confers no authority over any corpus document. Where it differs from a corpus document, the corpus document prevails.

Every AI system the organization operates is built on data, and the corpus makes the data question one of the governing body's own oversight questions: what data the organization's AI uses, under what authority, and how its quality and provenance are governed (the board oversight guide, its section 5). The AI security and risk standard places classification at the front of the data lifecycle and sets a closed data-governance baseline covering classification, provenance, and lineage (its section 4.2).

The governing body does not classify datasets or run pipelines. Its oversight is a dependency on management operating the data governance layer the corpus defines, and on management holding the evidence that each control exists and operates. This brief points to where the corpus places each control; the division of labour it describes is a composite reading, and the limitations say so.

• Classification comes before use. The AI security and risk standard sets the closed data-governance baseline in its section 4.2. Use that section for its full requirements; the executive point is that management must evidence the baseline before and throughout AI-system use.
• AI data has a handling floor. The data classification and handling standard defines the organization's classification levels (its section 4) and sets the handling floor for AI artefacts by reference to those levels (its section 7). It also assigns oversight of AI dataset classification to the governance roles it names (its section 3). This brief points to the standard for the levels, the floor, and the named roles, and does not reproduce them.
• Classification travels with the data. The data classification and handling standard sets the classification-inheritance rule for AI-derived data in its section 5, and the data architecture standard governs classification-tag propagation and the marking of AI-unsuitable data in its section 5.
• Training data has its own ten-step lifecycle procedure. The training data governance procedure owns the ordered steps and their contents; use it for the dataset source and lawful-basis, sensitive-content, approval-gate, lineage, and deletion-propagation requirements it defines.
• Policy backs the procedure. The privacy and data governance policy requires that all data assets are classified (its section 4.2), that AI training datasets meet the conditions its section 4.7 sets, that retention schedules cover AI training artifacts (its section 4.4), and that privacy impact assessments cover AI model training activities (its section 4.5).
• Quality is gated, not asserted. The AI data quality and readiness validation standard records a readiness sign-off before a dataset trains or operates an AI system at scale; a dataset that has not passed the gate is not used at scale, and for a dataset that misses its targets, passing the gate has a dependency on the executed improvement plan or the formally accepted residual risk (its sections 3.2 and 3.7).
• Every dataset is documented. The dataset datasheet template owns the datasheet's applicability and update boundaries in its Purpose and Operating expectations sections; use those for when a datasheet is required, produced, and refreshed.
• Retention and cross-border movement are governed by registers. The data retention schedule carries the retention rows for AI governance records, including training data provenance and AI decision logs (its section 7), composed with the domain minimums the records retention and destruction standard sets (its section 5); the applicable period resolves to those corpus documents rather than to this page. The cross-border data flow register documents transfers of personal data across jurisdictions, and maintaining it is a policy requirement (the privacy and data governance policy, its section 4.6).
• Access rules follow the data into retrieval. The AI access and agent permissions standard sets the AI-to-data access-control requirements for retrieval and context injection (its section 7); use that section for the specific controls. This is a prevention against an AI system becoming a route around the organization's data access controls.
• The assessments are routed, not improvised. The integrated AI and privacy assessment procedure determines, from a system's regime triggers, which data protection and AI assessments are required and how they compose (its Step 3).

For the governing body and accountable executive leadership, the corpus position reads as follows. (This reading is a composite of the sources above; see the limitations.)

• The governing body requires the layer, management operates it. The board oversight guide has the governing body set appetite and policy direction while management operates within it (its section 4). Applied to data, the governing body's contribution is to require that classification, lineage, quality gating, retention, and access control exist as the corpus defines them, and to ask for the evidence that they operate.
• Dataset controls apply before the data is used. The corpus places its dataset-level controls (the approval to train, the readiness sign-off, the datasheet) at the boundaries its sources set, before a dataset is put into AI use at scale. Any later approval that relies on that data has the data governance layer as a dependency.
• Some answers are weak answers. A quality claim without the recorded readiness sign-off the AI data quality and readiness validation standard defines is an assertion, not evidence. A retention answer that does not resolve to the data retention schedule does not demonstrate that it uses the applicable period. And a claim that retrieval "only sees what the user can see" is checkable against the specific controls in the AI access and agent permissions standard (its section 7).
• The values stay in the corpus. Classification levels, retention periods, approval roles, and assessment triggers are corpus values. The governing body should expect management's answers to point into those documents rather than restate them from memory, for the same reason this brief does.

Each item below is an evidence class the governing body can call for from management, and each is evidence for a named control or outcome. The board oversight guide frames the questions (its section 5); the guide's section 7 sets what the governing body should expect to receive on a regular cadence.

• The data classification record for each AI system's data, as evidence for the classification-before-use control in the AI security and risk standard (its sections 4.2 and 5).
• The dataset datasheet, as evidence that each dataset in the AI System Register is documented per the dataset datasheet template.
• Provenance and lineage records, as evidence for the lineage controls in the AI security and risk standard (its section 4.2) and the training data governance procedure (its Step 5).
• The approval-to-train record, as evidence that the dataset approval gate operated with the approvers the training data governance procedure names (its Step 4).
• The readiness sign-off record, as evidence for the data-quality gate in the AI data quality and readiness validation standard (its sections 3.7 and 4).
• The retention mapping for AI records, as evidence that AI governance records are held to the rows in the data retention schedule (its section 7).
• The cross-border transfer entries, as evidence for transfer governance per the cross-border data flow register and the privacy and data governance policy (its section 4.6).
• The retrieval access-control design, as evidence for the AI-to-data controls in the AI access and agent permissions standard (its section 7).
• Deletion propagation records, as evidence for the deletion-propagation controls the training data governance procedure defines (its Step 6), and that retirement handles datasets and retrieval stores per the AI security and risk standard (its section 4.11).
• The assessment routing record, as evidence that the required data protection and AI assessments were routed and completed per the integrated AI and privacy assessment procedure (its Step 3 and Required outputs).

• This page creates no compliance and imposes no requirement. The linked corpus documents govern; where this page differs from a corpus document, the corpus document prevails.
• This page carries composite claims. The division of labour it describes (the governing body requires the data governance layer and asks for evidence; management operates the controls; the dataset-level controls apply at the boundaries their sources set) is synthesized across several corpus sources, and the adopting organization validates that reading against its own governance and committee structure.
• The corpus documents this page points to carry the classification levels, handling floors, retention periods, approval roles, assessment triggers, and named oversight bodies. This page routes to them and does not reproduce them; every value resolves to the corpus documents, not to this page.
• This page makes no claim about how likely any failure is, and none of its statements should be read as a probability or a frequency.