Data Governance for AI: Ownership and Stewardship

1. Why This Topic Matters

Artificial intelligence systems are only as trustworthy as the data that feeds them. Every stage of the AI life cycle — from problem framing and data collection, through training, validation, deployment and monitoring — depends on data that is accurate, complete, representative, lawful to use, and properly protected. When no one clearly owns a dataset and no one is actively stewarding it, predictable failures follow: biased models trained on unrepresentative samples, privacy violations from data used beyond its original purpose, drift that nobody notices, and regulatory penalties because the organisation cannot demonstrate lineage or consent.

ISACA places data governance for AI inside the AI Life Cycle Risk Management domain of the Advanced in AI Risk (AAIR) body of knowledge because ownership and stewardship are the control mechanisms that convert abstract data principles into day-to-day accountability. Without named owners and stewards, risk treatment for data is nobody's job. For the exam, this topic tests whether you understand that data governance is a structure of roles, decision rights and accountabilities, not merely a set of technical tools.

Regulatory pressure reinforces this. The EU AI Act requires data governance and management practices for high-risk systems (Article 10), covering design choices, data collection, preparation, assumptions, bias examination and gap identification. GDPR demands an identifiable controller accountable for lawful processing. NIST AI RMF (Govern and Map functions) and ISO/IEC 42001 both expect documented roles for data throughout the AI system life cycle. An AAIR candidate must be able to connect these expectations to the practical roles of owner, steward and custodian.

2. What It Is: Core Definitions

Data governance is the system of decision rights, policies, standards, processes and accountabilities that ensures data is managed as a strategic asset and that its use aligns with organisational objectives, risk appetite, ethics and legal obligations. For AI, governance extends to training data, validation and test data, inference inputs, model outputs, feedback data, synthetic data and labels/annotations.

Data ownership is the formal assignment of accountability for a defined data asset or domain to a business role (not IT by default). The data owner:
• Is accountable for the quality, classification, protection and appropriate use of the data.
• Approves access, sharing, secondary use (such as re-use of customer data to train a model) and retention decisions.
• Accepts or escalates risk related to the data within the enterprise risk framework.
• Is typically a senior business leader for the domain (for example, the Head of Lending owns loan application data).

Data stewardship is the operational responsibility for executing the owner's decisions and maintaining the data day to day. The data steward:
• Defines and maintains business metadata, definitions, data dictionaries and glossaries.
• Monitors and remediates data quality issues (accuracy, completeness, consistency, timeliness, validity, uniqueness).
• Documents lineage and provenance, and maintains datasheets or data cards for AI datasets.
• Enforces classification labels and handling rules, coordinates with privacy, security and legal.
• Acts as the subject-matter liaison between business, data science and engineering teams.

Data custodian (frequently tested as a distractor) is the technical role — often IT or platform engineering — responsible for the technical environment: storage, backup, access control implementation, encryption, availability. Custodians implement controls; they do not decide who should have access or what the data means.

Key distinction to memorise: Owner = accountable (decides, approves, accepts risk). Steward = responsible (executes, maintains, monitors quality and metadata). Custodian = technical safeguarding (implements infrastructure and security controls). In RACI terms, the owner is the A, the steward is the R, and the custodian is often a C or an R for technical tasks.

Related roles in an AI context:
• Chief Data Officer (CDO) — enterprise-level accountability for the data governance programme.
• Data governance council / board — cross-functional body that sets policy, resolves conflicts between domains and approves standards.
• Model owner — accountable for the model as a whole; must coordinate with data owners whose data the model consumes. A model may draw on many data domains, so one model owner may depend on several data owners.
• AI ethics / responsible AI committee — reviews high-impact uses of data (sensitive attributes, inferred characteristics).
• Data Protection Officer (DPO) — oversees personal data compliance; advises owners and stewards on lawful basis and data subject rights.

3. How It Works Across the AI Life Cycle

Stage 1 — Planning and Problem Definition
Before any data is gathered, the organisation identifies which data domains the AI use case will require and confirms that each has a named owner. The owner assesses whether the proposed purpose is compatible with the original collection purpose (purpose limitation). A Data Protection Impact Assessment or AI impact assessment is initiated where personal or sensitive data is involved. Absence of an identifiable owner at this stage is itself a red flag and a reason to halt or escalate.

Stage 2 — Data Acquisition and Collection
Owners approve sourcing decisions: internal extracts, third-party purchases, web scraping, open datasets, synthetic generation. Stewards document provenance (where did it come from, under what licence or consent, when, by whom) and record it in a data catalogue or datasheet. Contract terms with external providers should define ownership, permitted uses, quality obligations and audit rights. Scraped or purchased data without clear provenance is a significant legal and reputational risk.

Stage 3 — Data Preparation, Labelling and Feature Engineering
Stewards monitor quality dimensions and representativeness, run bias and gap analyses, and ensure labelling guidelines are documented and annotator quality is measured. Any transformation (cleansing, imputation, de-identification, enrichment) is recorded in lineage so the training set can be reproduced and audited. Owners approve the use of sensitive attributes (race, gender, health) even when they are used solely for fairness testing.

Stage 4 — Model Training and Validation
Training, validation and test datasets are versioned and their composition documented. Stewards verify that there is no leakage between sets and that holdout data reflects the deployment population. Owners confirm that the data used matches what was approved. Governance evidence — approvals, datasheets, quality reports — is retained for audit and regulatory inquiry.

Stage 5 — Deployment and Operation
Inference data is governed too: who owns real-time inputs, logs and generated outputs? Outputs of a model frequently become new data assets (scores, classifications, embeddings) and must be assigned an owner and classification. Stewards monitor for data drift and schema changes upstream that would silently degrade the model, and have escalation paths to the model owner.

Stage 6 — Monitoring, Feedback and Retraining
Feedback loops (human corrections, user behaviour) create new training data. Governance determines whether that data may be used for retraining, whether consent covers it, and how long it is retained. Stewards update datasheets with each new version.

Stage 7 — Retirement and Decommissioning
Owners decide retention and secure disposal of training data, model artefacts and logs in line with legal holds and retention schedules. Data that was de-identified may need to be re-evaluated for re-identification risk before any further use or archival.

Enabling mechanisms that the exam expects you to recognise:
• Data catalogue and metadata management — the system of record for owners, stewards, definitions, classification and lineage.
• Data classification scheme — public, internal, confidential, restricted; drives handling and access controls.
• Datasheets for datasets / data cards / model cards — structured documentation of motivation, composition, collection process, preprocessing, uses and limitations.
• Data quality metrics and SLAs — measurable thresholds that trigger remediation.
• Access governance — role-based or attribute-based access, approved by owners, implemented by custodians, reviewed periodically.
• Policies and standards — acceptable use, secondary use, synthetic data, third-party data, retention.
• Data governance council escalation — resolution of cross-domain disputes (for example, Marketing wants to train on Finance data).

4. Common Risks When Ownership and Stewardship Are Weak

• Orphaned data — datasets with no owner, so nobody approves use or fixes quality; models silently inherit defects.
• Purpose creep — data collected for one purpose reused for AI training without reassessing lawful basis or consent.
• Shadow data and shadow AI — teams copying production data into unmanaged environments or external tools.
• Unknown provenance — inability to prove licence, consent or source; exposure to copyright and privacy claims.
• Bias and unrepresentativeness — no steward checking demographic coverage or label quality.
• Lineage gaps — inability to reproduce a training set or explain a model decision to a regulator.
• Misaligned accountability — IT custodians treated as owners, so business risk decisions are made by people without authority or context.
• Third-party dependency — vendor datasets or foundation-model training data with no contractual quality or audit rights.

5. Governance Design Principles ISACA Favours

• Ownership sits with the business, because the business understands meaning, value and acceptable use.
• Accountability must be singular and explicit — one accountable owner per data asset or domain, documented in the catalogue.
• Stewardship is continuous, not a one-time project; it operates across the entire life cycle including post-deployment.
• Controls are risk-based — more rigorous for sensitive, high-impact or regulated data and for high-risk AI systems.
• Governance produces evidence — approvals, lineage, quality reports and impact assessments that support audit, assurance and regulatory response.
• Governance is integrated with enterprise risk management, privacy, security and the AI governance framework; it is not a standalone silo.
• Segregation of duties — those who build models should not unilaterally approve the data they use.

6. Exam Tips: Answering Questions on Data Governance for AI: Ownership and Stewardship

Tip 1 — Know the role vocabulary cold. Many questions hinge on choosing the correct role. If the question asks who is accountable, who approves access, who accepts risk, or who decides on secondary use, the answer is the data owner. If it asks who maintains metadata, monitors data quality, documents lineage or defines business terms, the answer is the data steward. If it asks who implements backups, encryption or access control technically, the answer is the data custodian. Do not let the words IT or database administrator lure you into picking a custodian for an accountability question.

Tip 2 — Business over IT for ownership. When options include a business executive and an IT leader as potential data owners, ISACA almost always expects the business role. IT is custodian unless the data is genuinely IT's own (for example, system logs).

Tip 3 — Look for the word FIRST, BEST or MOST. AAIR questions often ask what should be done first. In data governance scenarios, the first step is usually to identify and assign an accountable owner or classify the data before any technical control, tool purchase or model development. You cannot govern what has no owner and no classification.

Tip 4 — Governance before technology. If one option is to buy a data quality tool or catalogue and another is to establish roles, policies and accountabilities, choose the governance structure. Tools enable governance; they do not substitute for it.

Tip 5 — Connect to the life cycle. Questions in this domain are framed around life-cycle stages. Ask yourself: at which stage is the scenario, and what governance activity belongs there? Early stages emphasise purpose compatibility, owner identification and provenance; middle stages emphasise quality, bias checks, lineage and versioning; late stages emphasise drift monitoring, output ownership, retention and disposal.

Tip 6 — Secondary use and purpose limitation are favourite scenarios. When a scenario describes reusing existing customer data to train a new model, the correct response involves the data owner (with privacy/legal input) assessing compatibility of purpose, lawful basis and consent before use. Anonymisation or de-identification may be a supporting control but does not replace the owner's decision.

Tip 7 — Watch for the model owner versus data owner trap. A model owner is accountable for the model, but does not automatically have rights over the data it consumes. If the scenario is about whether data may be used, the data owner decides; if it is about model performance or deployment, the model owner is accountable. Good answers often involve coordination between them through the governance council.

Tip 8 — Third-party and foundation-model data. When data or pre-trained models come from vendors, the best answers emphasise contractual clarity on ownership, permitted use, provenance documentation, quality obligations and audit rights, plus internal assignment of an owner for the ingested asset. The organisation remains accountable even when the data is external.

Tip 9 — Evidence and auditability. If an option mentions documenting datasheets, lineage, approvals or decisions, it is frequently the best answer because AAIR values demonstrable accountability that supports assurance and regulatory response.

Tip 10 — Eliminate absolutes and shortcuts. Options that say the data science team may use any internal data, that ownership is unnecessary for anonymised data, or that stewardship ends at deployment are wrong. Governance is continuous and risk-based; anonymised and synthetic data still need owners and quality oversight.

Tip 11 — Read for the risk being asked about. If the stem emphasises bias, point toward stewardship activities (representativeness analysis, label quality, documentation of composition). If it emphasises privacy, point toward owner approval, lawful basis and DPO involvement. If it emphasises reproducibility or explainability, point toward lineage and versioning. If it emphasises accountability gaps, point toward assigning owners and updating the catalogue.

Tip 12 — Use the council for conflicts. When two business units disagree about data use, the correct escalation is the data governance council or equivalent body, not the CIO unilaterally and not the data scientists.

7. Quick Revision Summary

• Data governance for AI = roles + decision rights + policies + processes + evidence, spanning the entire AI life cycle.
• Owner is accountable, business-based, approves use and access, accepts risk.
• Steward is responsible for metadata, quality, lineage, documentation and day-to-day enforcement.
• Custodian implements technical safeguards.
• Key artefacts: data catalogue, classification scheme, datasheets/data cards, lineage records, quality metrics, impact assessments.
• Key risks of weak governance: orphaned data, purpose creep, unknown provenance, bias, lineage gaps, shadow AI, third-party exposure.
• Exam heuristics: assign ownership first, business over IT, governance before tools, owner decides on secondary use, steward monitors quality and bias, document everything, governance never stops at deployment.

Master these distinctions and you will be able to handle any AAIR question that tests Data Governance for AI: Ownership and Stewardship with confidence.