12. Data Governance, Security, Privacy and Responsible AIIntroduction and contextData science creates value by converting data into evidence, predictions, recommendations, and automated decisions. Yet the technical ability to collect and analyse data does not establish the right to use it, guarantee its quality, or make the resulting decisions legitimate. A model trained on inaccurate, unlawfully obtained, or poorly governed data may be statistically sophisticated but organizationally unsafe. Governance therefore provides the institutional conditions under which data platforms, analytics, machine learning, and artificial intelligence can be used responsibly. Data governance is a strategic discipline concerned with authority, accountability, policies, standards, and controls over data assets. It connects executive priorities with operational practices such as metadata management, access control, data quality, lineage, retention, and audit. Security protects data and systems from unauthorized access, alteration, disruption, or destruction. Privacy concerns the appropriate handling of information about identifiable or identifiable individuals. Responsible AI extends these concerns to systems that infer, recommend, classify, or act, requiring attention to fairness, transparency, human oversight, safety, and accountability. For data professionals, these subjects are not separate compliance topics. They are connected parts of the data lifecycle. Weak ownership produces unclear access decisions; poor metadata obscures sensitive information; inadequate lineage makes impact analysis difficult; and insufficient monitoring allows both cyber threats and model harms to persist. This chapter examines how organizations can design an integrated governance capability for data platforms and AI-enabled operations. Learning objectivesBy the end of this chapter, the learner should be able to: 1. Explain the relationship between data governance, security, privacy, and responsible AI. 2. Distinguish ownership, stewardship, custodianship, and accountability. 3. Design governance arrangements for metadata, lineage, quality, access, and retention. 4. Apply risk-based security and privacy controls to data science workflows. 5. Evaluate fairness, transparency, explainability, auditability, and human oversight in AI systems. 6. Analyse organizational trade-offs between data utility, protection, compliance, and innovation. 7. Recommend a practical governance framework for a data platform or AI use case. Governance as an operating model for data and AIA governance framework answers five basic questions: What data exists? Who is responsible for it? Who may use it and for what purpose? How can its use be demonstrated as lawful, safe, and appropriate? What happens when something goes wrong? These questions convert abstract principles into operating mechanisms. Data governance should be distinguished from data management. Data management includes the technical and operational activities used to store, integrate, process, protect, and analyse data. Governance defines the policies, decision rights, standards, and accountability that guide those activities. A data catalogue, for example, is a management tool; rules defining who must register a dataset, which metadata are mandatory, and who approves access are governance arrangements. A mature model distributes responsibility rather than placing all responsibility on an information technology department. Senior leaders establish risk appetite and priorities. Data owners make decisions about a business domain, such as customer or finance data. Data stewards maintain definitions, quality rules, and appropriate use practices. Platform custodians operate storage, networks, identity services, and controls. Data scientists and analysts remain accountable for how they use data and communicate results. Privacy, legal, security, risk, and audit specialists provide independent challenge and assurance. A useful responsibility model is a tailored RACI matrix. The person or group that is accountable owns the final decision; those responsible perform the work; those consulted provide expertise; and those informed receive relevant information. For example, a customer-data owner may be accountable for access policy, a steward responsible for data definitions, security consulted on technical controls, and the privacy office consulted on lawful purpose. The exact allocation varies by organization, but ambiguity should never be treated as a governance strategy. Data ownership, stewardship, and accountabilityOwnership is decision authority, not necessarily possession. A business owner may be accountable for a dataset even when the data is stored in a cloud platform operated by another team. Stewardship is the continuing practice of maintaining meaning, quality, classification, and appropriate use. Custodianship concerns the technical operation and protection of the data environment. Accountability is broader: it includes being able to justify decisions, demonstrate compliance, and remediate harm. Governance councils are often useful for cross-domain decisions. A council can approve enterprise standards, resolve conflicts between business units, prioritize remediation, and review high-risk analytical initiatives. It should not become a committee that delays ordinary work. Decision thresholds, delegated authority, service levels, and escalation routes are necessary to make governance operational. The information foundation: metadata, quality, and lineageMetadata is data about data. Technical metadata describes formats, schemas, tables, columns, interfaces, and storage locations. Business metadata records definitions, owners, permitted uses, and business context. Operational metadata includes refresh times, job status, access events, and quality results. Security and privacy metadata may include classification, sensitivity, retention period, legal basis, and restrictions on use. A catalogue makes metadata discoverable, but discovery alone does not create trust. Users need to know whether a dataset is current, complete, representative, approved, and fit for a particular purpose. A strong catalogue links assets to owners, policies, quality indicators, lineage, and usage history. Automated scanning can identify schemas and sensitive patterns, while human stewardship is needed to validate meaning and context. Data quality is multidimensional. Common dimensions include accuracy, completeness, consistency, timeliness, validity, uniqueness, and conformity to accepted definitions. Quality thresholds should be connected to business consequences. A small error rate may be tolerable for exploratory analysis but unacceptable for regulatory reporting or clinical decision support. Quality rules should be monitored at ingestion, transformation, and consumption points rather than checked only after a failure. Data lineage records the origins, transformations, movement, and destinations of data. It may be represented as technical lineage between tables and pipelines or as business lineage showing how a source measure contributes to a report, feature, or model output. Lineage supports impact analysis, incident response, reproducibility, regulatory explanation, and model validation. If a source field changes, lineage helps identify affected dashboards and models. If a prediction is challenged, lineage helps establish which data and transformations contributed to it. Classification, access, and security architectureClassification assigns protection requirements according to sensitivity and impact. A simple scheme might distinguish public, internal, confidential, and restricted information, although categories should reflect the organization’s risks. Classification is meaningful only when connected to controls. Restricted information may require encryption, stronger authentication, limited processing locations, enhanced logging, and shorter retention. Access should follow least privilege: users receive only the permissions necessary for an approved task and for an appropriate period. Role-based access control assigns permissions to job roles, while attribute-based access control can consider attributes such as project, location, data sensitivity, purpose, and time. Just-in-time access and periodic recertification reduce standing privileges. Service accounts and machine identities require the same discipline as human accounts. A secure data platform uses defence in depth. Typical layers include identity and access management, network segmentation, encryption in transit and at rest, secrets management, endpoint protection, vulnerability management, secure configuration, backup, recovery, and centralized logging. The platform should be designed on the assumption that a control may fail. Zero-trust principles reinforce this approach by requiring explicit verification and minimizing implicit trust based on network location. Security is also a lifecycle activity. Threat modelling should occur before deploying a data product or model. It asks what must be protected, who may attack it, which attack paths exist, and what controls reduce the risk. Secure development practices include code review, dependency scanning, infrastructure-as-code controls, test-data management, and separation of development, testing, and production environments. Incident response plans must identify containment, investigation, communication, recovery, and lessons learned. Data exfiltration, ransomware, credential theft, insider misuse, insecure APIs, misconfigured storage, and supply-chain compromise are common categories of risk. AI systems add threats such as prompt injection, training-data poisoning, model theft, membership inference, and adversarial manipulation. Controls should be proportionate to the sensitivity and consequences of the use case rather than based only on the novelty of the technology. Privacy by design and compliance reasoningPrivacy is not equivalent to secrecy. It includes appropriate collection, use, disclosure, retention, accuracy, access, correction, and deletion of personal information. Personal data may include direct identifiers, such as names, and indirect combinations, such as location, timestamps, and behavioural patterns, that can make individuals identifiable. Sensitive categories often require enhanced safeguards, but the precise legal definitions depend on the applicable jurisdiction. Privacy by design incorporates protection into architecture and process from the beginning. Important practices include purpose specification, data minimization, default confidentiality, retention limitation, controlled sharing, privacy impact assessment, and mechanisms for individual rights. A privacy impact assessment should identify the people affected, intended purposes, data flows, risks, controls, residual risks, and approval conditions. Several technical approaches can reduce privacy risk. Pseudonymization replaces identifiers with tokens but does not make data anonymous if re-identification remains possible. Aggregation reduces detail; suppression removes records or values; generalization replaces precise values with broader categories. Differential privacy adds calibrated statistical noise and provides a formal privacy guarantee under specified assumptions. Federated learning keeps some training data in local environments, although model updates and shared infrastructure still require protection. Synthetic data can support development, but it may preserve sensitive patterns or fail to represent important populations; it therefore requires validation rather than automatic trust. Privacy decisions require a lawful and ethical purpose. Consent is one possible basis for processing, but it should not be treated as a universal solution, especially where people have limited choice or cannot understand future uses. Organizations must consider purpose compatibility, transparency, proportionality, retention, cross-border transfers, contractual obligations, and sector-specific requirements. Compliance is jurisdiction-dependent; therefore, a governance program should map applicable obligations to controls and maintain evidence rather than relying on general claims of compliance. Responsible AI: from principles to controlsResponsible AI applies governance to systems that learn from data or generate outputs with social, economic, or operational consequences. Common principles include fairness, transparency, explainability, privacy, safety, robustness, security, human oversight, contestability, and accountability. Principles become credible only when translated into measurable requirements, documented decisions, testing procedures, and monitoring. Fairness is context-dependent. A model may perform differently across demographic or operational groups because of historical disadvantage, sampling imbalance, measurement error, or differences in base rates. Relevant measures may include group-specific error rates, false-positive and false-negative rates, calibration, selection rates, and equal opportunity. These measures can conflict; improving one notion of fairness may worsen another. Consequently, teams must identify the decision context, affected groups, acceptable harms, and legal or organizational constraints rather than selecting a metric mechanically. Transparency means making relevant information available about a system, its purpose, data, limitations, decision authority, and routes for challenge. Explainability concerns the ability to provide understandable reasons for a particular output or overall model behaviour. A highly interpretable model is not automatically fair, and a complex model is not automatically unacceptable. Explanations should be accurate, stable, meaningful to the audience, and suitable for the decision. Feature-importance displays can be misleading when variables are correlated or when an explanation is mistaken for causation. Human oversight must be substantive. A human who merely approves automated outputs under time pressure may provide little real control. Effective oversight defines when a person must review, what evidence they can access, when they may override a system, and how overrides are recorded. High-impact decisions should include channels for notification, correction, appeal, and remediation. Human involvement does not eliminate responsibility; it can instead create a false sense of safety if roles are unclear. Model governance typically includes an inventory of models, documented intended use, data and feature descriptions, validation results, fairness analysis, approval status, version control, performance thresholds, monitoring plans, and retirement conditions. A model card or system card can communicate purpose, limitations, evaluation populations, and known risks. A decision log records assumptions, trade-offs, approvals, exceptions, and changes. A lifecycle framework for governed analytics and AIAn integrated lifecycle can be organized into seven connected stages. 1. Define the purpose and risk. State the business or public purpose, affected people, decision being supported, alternatives, expected benefits, and foreseeable harms. Classify the use case by impact and sensitivity. 2. Establish authority and data provenance. Identify the owner, steward, processor, users, sources, permissions, legal or organizational basis, retention requirements, and lineage. Reject data whose provenance cannot be adequately established for the intended use. 3. Design controls before development. Specify access, encryption, minimization, quality thresholds, privacy safeguards, security tests, human review, documentation, and incident procedures. Treat controls as design requirements rather than paperwork added after deployment. 4. Build and validate. Separate environments, restrict development data, test data quality, evaluate model performance across relevant populations, conduct threat modelling, and document limitations. Independent review is valuable for high-impact systems. 5. Approve and deploy conditionally. Define an approval authority and deployment conditions. Use staged release, rollback capability, monitoring, and clear communication to users and affected stakeholders. 6. Monitor and respond. Monitor data drift, concept drift, performance, fairness indicators, access events, security alerts, privacy incidents, complaints, and human overrides. Establish thresholds that trigger investigation, retraining, restriction, or suspension. 7. Retire and learn. Remove obsolete models and data according to retention policy, revoke access, preserve necessary records, assess outcomes, and feed lessons into standards and future designs. This lifecycle is iterative. A change in purpose, population, data source, model, or operating environment may require renewed assessment rather than routine change approval. Strategic implementation in organizationsGovernance implementation is a change program as much as a technology program. Organizations should begin with a risk-based baseline rather than attempting to govern every asset identically. High-value and high-risk domains—such as identity, financial decisions, employment, health, safety, or critical infrastructure—usually deserve earlier investment. A practical roadmap may start by inventorying critical data products and models, assigning owners, defining a minimum metadata standard, implementing identity controls, and establishing incident and exception processes. The next stage can integrate catalogues with lineage, data-quality monitoring, policy enforcement, model registries, and audit evidence. Mature organizations connect governance metrics to business outcomes, such as reduced incident exposure, faster impact analysis, improved model stability, or fewer inappropriate access events. Technology can automate discovery, classification, policy checks, access reviews, lineage capture, quality tests, model monitoring, and evidence collection. Automation should be treated as assistance, not authority. A classifier may incorrectly identify sensitive data; an automated fairness test may omit a relevant population; and a policy engine may enforce an outdated rule. Human review is necessary for ambiguous and high-impact cases. Governance can create friction, but ungoverned speed creates hidden costs: repeated data preparation, unreliable decisions, security incidents, regulatory action, loss of public trust, and expensive remediation. The strategic objective is not maximum restriction. It is controlled reuse—making trustworthy data available to authorized users while limiting misuse and making consequences visible. Illustrative case study: a lending platform and automated affordability assessmentBackgroundA fictional digital lender, Northstar Finance, uses transaction history, application information, and external indicators to assess affordability and offer loan terms. It operates a cloud data platform containing customer records, feature pipelines, model artefacts, decision logs, and support data. Situation and problemThe organization discovers that applicants from one region receive substantially more manual reviews and have higher rejection rates. The model team cannot quickly explain the difference because feature definitions are inconsistent, the training dataset has incomplete lineage, and several analysts share broad access to production extracts. A security review also identifies long-lived service credentials and inadequate logging of data exports. AnalysisThe problem is not merely a model-accuracy issue. It combines weak governance, privacy exposure, security weakness, and possible unfairness. The region variable may act as a proxy for protected characteristics or may reflect uneven data quality. The absence of reliable lineage prevents the team from determining which source changes affected the model. Broad access increases the consequences of a breach, while poor decision logging limits contestability. Approach and strategyNorthstar establishes an accountable data owner and a model owner, supported by privacy, security, risk, and customer-service representatives. It creates a catalogue entry for each source and feature, records definitions and permitted uses, and reconstructs lineage from ingestion to decision. Production extracts are replaced with controlled views and pseudonymized development data. Role- and attribute-based access is introduced, service credentials are rotated, and export events are logged. The model is revalidated using representative evaluation data. The team compares selection rates, false-positive rates, false-negative rates, calibration, and manual-review rates across relevant groups. It investigates whether regional differences arise from legitimate affordability signals, missing data, or proxy effects. A simpler model is considered where it offers comparable performance with better interpretability. Applicants receive understandable information about the decision and a route to request review or correction. Monitoring thresholds are established for drift, group disparities, data quality, and unusual access. OutcomeThe revised process does not guarantee identical outcomes across groups; rather, it documents the reasons for differences and tests whether they are justified, accurate, and proportionate. Access exposure is reduced, audit evidence improves, and the organization can investigate complaints more quickly. Some applications take longer because of additional review, but the lender gains a defensible process and clearer accountability. Learning pointsThe case demonstrates that responsible AI cannot be achieved by adding an explanation tool after deployment. It depends on ownership, provenance, quality, privacy, security, evaluation, human review, and continuous monitoring. It also illustrates that fairness is a socio-technical question: model metrics must be interpreted alongside data practices, institutional policy, and the effects on people. Challenges, limitations, and common mistakesA frequent mistake is treating governance as a central approval bottleneck. Excessive centralization encourages workarounds, while complete decentralization produces incompatible definitions and inconsistent controls. Federated governance—common standards with delegated domain responsibility—often provides a practical balance. Another mistake is assuming that de-identification eliminates privacy risk. Combining datasets, auxiliary information, or repeated releases can enable re-identification. Similarly, encryption does not solve inappropriate authorization, and access control does not guarantee that an authorized user will use data ethically. Organizations may also measure model performance only on average. Aggregate accuracy can conceal severe disparities, poor calibration, or failure on rare but important cases. Fairness metrics themselves have limitations and cannot determine which social outcomes are acceptable. Documentation can become a ritual if it is not connected to decisions, monitoring, and consequences. Data and model drift create further difficulty. A model that was valid at deployment may become unreliable when behaviour, policy, markets, or populations change. Monitoring thresholds can generate false alarms or fail to detect gradual harm. Finally, vendor dependence can obscure training data, subcontractors, retention, security practices, and model updates. Procurement must therefore include audit rights, incident notification, data-use restrictions, portability, service levels, and exit arrangements. Current developments and emerging directionsOrganizations are increasingly integrating data catalogues, policy engines, data-quality platforms, privacy-enhancing technologies, security information systems, and model registries into governed data platforms. Machine-readable policies can support automated enforcement, while lineage standards improve impact analysis across distributed environments. Generative AI introduces additional governance questions: provenance and copyright of training material, confidential information in prompts, hallucinated outputs, evaluation of generated content, prompt and tool security, and responsibility for downstream use. Retrieval-augmented systems require controls over document access and citation quality. AI agents require particular attention to delegated authority, tool permissions, action limits, and human confirmation for consequential operations. Regulatory approaches increasingly emphasize risk classification, transparency, documentation, accountability, and controls across the AI lifecycle. Because legal requirements differ and continue to develop, organizations should maintain a living obligations register and seek qualified legal interpretation for specific deployments. International standards and frameworks can provide useful structures, but adoption does not itself prove compliance or ethical adequacy. Best-practice checklistA credible governance capability should be able to demonstrate that: - critical data and models have named owners and stewards;
- definitions, classifications, quality rules, lineage, and retention are documented;
- access is least-privilege, periodically reviewed, and auditable;
- sensitive data is minimized, protected, and used for an identified purpose;
- threats, privacy risks, and potential harms are assessed before deployment;
- models are tested for performance, robustness, fairness, and relevant limitations;
- decisions, approvals, exceptions, changes, and incidents are recorded;
- affected people have appropriate notice, correction, review, or appeal mechanisms;
- monitoring can trigger investigation, rollback, restriction, or retirement; and
- governance responsibilities extend to vendors and downstream users.
Chapter SummaryData governance establishes decision rights, accountability, standards, and controls for trustworthy data use. Ownership and stewardship clarify responsibility, while metadata, quality management, and lineage make data understandable, traceable, and reusable. Security protects systems through layered controls, least privilege, identity management, encryption, monitoring, and incident response. Privacy by design limits inappropriate collection and use through purpose specification, minimization, retention controls, impact assessment, and privacy-enhancing techniques. Responsible AI applies these ideas to predictive and generative systems. Fairness, transparency, explainability, safety, human oversight, contestability, and accountability must be translated into lifecycle controls, evidence, and monitoring. Effective governance is risk-based, federated where appropriate, supported by automation but not surrendered to automation, and extended across vendors and the full life of a data product or model. Its strategic value lies in enabling trustworthy reuse while reducing operational, legal, ethical, and security risks. Key TermsData governance: The system of policies, decision rights, standards, and accountability governing data assets. Data owner: The person or function accountable for decisions about a data domain or asset. Data steward: The role responsible for maintaining data meaning, quality, classification, and appropriate use. Data custodian: The technical operator responsible for storing, processing, and protecting data. Metadata: Information that describes data’s structure, meaning, origin, use, or operational state. Data lineage: A record of data origins, transformations, movement, and destinations. Data quality: The degree to which data is accurate, complete, consistent, timely, valid, and fit for purpose. Least privilege: A security principle granting only the access necessary for an authorized task. Role-based access control: Access assignment based primarily on organizational roles. Attribute-based access control: Access decisions based on attributes such as user, resource, purpose, or context. Privacy by design: Embedding privacy safeguards into systems and processes from their inception. Pseudonymization: Replacing identifying values with substitutes while retaining a possibility of re-identification. Differential privacy: A formal technique that limits what can be inferred about an individual from statistical outputs. Model drift: A change in data or relationships that reduces a model’s continuing validity. Fairness: The context-dependent assessment of whether an AI system produces unjustified or discriminatory differences. Explainability: The ability to provide understandable and faithful accounts of model behaviour or outputs. Auditability: The capacity to reconstruct, examine, and evidence decisions, actions, and controls. Responsible AI: The design and use of AI with attention to fairness, safety, privacy, transparency, accountability, and human welfare. Review QuestionsShort Answer Questions1. Distinguish data governance from data management. 2. What is the purpose of data lineage? 3. Explain the principle of least privilege. 4. Why does pseudonymization not necessarily make data anonymous? 5. What is meant by substantive human oversight? Long Answer Questions1. Explain how ownership, stewardship, custodianship, and accountability can be distributed in a governed data platform. 2. Discuss the role of metadata, data quality, and lineage in building trust in analytical outputs. 3. Describe a defence-in-depth security architecture for a cloud data platform. 4. Explain how privacy by design can be incorporated into a data science lifecycle. 5. Discuss why fairness, transparency, and explainability must be supported by monitoring and appeal mechanisms. Analytical / Application Questions1. An organization has accurate models but no record of source transformations or model changes. Analyse the governance risks and propose priorities for remediation. 2. A university wants to use student activity data to identify learners at risk of withdrawal. Design a risk-based governance process addressing purpose, privacy, fairness, access, human review, and monitoring. 3. A vendor offers a generative AI service that stores prompts and frequently changes its underlying model. Evaluate the contractual, security, privacy, and responsible-AI questions that should be resolved before adoption. 22. Multiple Choice Questions with AnswersQuestion 1Which statement best distinguishes data governance from data management? A. Governance is limited to database backup, while management concerns ethics. B. Governance defines authority, policies, standards, and accountability; management performs operational data activities. C. Governance applies only to public data, while management applies only to personal data. D. Governance is performed by data scientists, while management is performed by regulators. Correct Answer: B. Governance defines authority, policies, standards, and accountability; management performs operational data activities. Question 2A data catalogue shows that a dataset exists but does not record its owner, permitted use, or refresh status. What is the principal weakness? A. The dataset has excessive encryption. B. The catalogue provides discovery without sufficient governance metadata. C. The dataset must be deleted immediately. D. The organization has applied differential privacy incorrectly. Correct Answer: B. The catalogue provides discovery without sufficient governance metadata. Question 3Which control most directly implements least privilege? A. Giving every analyst administrator access for convenience. B. Allowing permanent access to all production tables. C. Granting only the permissions required for an approved task and reviewing them periodically. D. Encrypting data while leaving permissions unchanged. Correct Answer: C. Granting only the permissions required for an approved task and reviewing them periodically. Question 4Why is lineage especially important when a source field changes? A. It automatically proves that the model is fair. B. It identifies downstream reports, features, and models that may be affected. C. It removes the need for data-quality testing. D. It prevents all cybersecurity attacks. Correct Answer: B. It identifies downstream reports, features, and models that may be affected. Question 5Which statement about pseudonymization is correct? A. It always makes personal data anonymous. B. It replaces identifiers but may still permit re-identification. C. It eliminates the need for access control. D. It guarantees that data cannot be linked across releases. Correct Answer: B. It replaces identifiers but may still permit re-identification. Question 6A model has high overall accuracy but a much higher false-negative rate for one population. What is the most appropriate next step? A. Publish the aggregate accuracy and ignore subgroup results. B. Delete all demographic evaluation data immediately. C. Investigate subgroup performance, data quality, context, and possible harm before deployment or continued use. D. Assume the difference is acceptable because the model is complex. Correct Answer: C. Investigate subgroup performance, data quality, context, and possible harm before deployment or continued use. Question 7Which example demonstrates substantive human oversight? A. A reviewer approves every output automatically without seeing evidence. B. A human is nominally assigned responsibility but cannot override the system. C. Reviewers have defined authority, relevant evidence, override procedures, and recorded decisions. D. The system is described as automated so that no person is accountable. Correct Answer: C. Reviewers have defined authority, relevant evidence, override procedures, and recorded decisions. Question 8What is the main purpose of a privacy impact assessment? A. To maximize the amount of personal data collected. B. To identify affected people, data flows, privacy risks, controls, and residual risks before or during processing. C. To replace all security testing. D. To certify that a model will never be biased. Correct Answer: B. To identify affected people, data flows, privacy risks, controls, and residual risks before or during processing. Question 9Which event should normally trigger renewed governance assessment? A. A routine password rotation with no change in access scope. B. A change in the model’s purpose, population, data source, or operating environment. C. A documented correction to a typographical error in internal notes. D. A scheduled backup that follows the approved procedure. Correct Answer: B. A change in the model’s purpose, population, data source, or operating environment. Question 10A generative AI vendor retains confidential prompts and changes its model without notice. Which response is most appropriate? A. Adopt it immediately because the service is popular. B. Ignore retention because encryption solves every privacy issue. C. Assess contractual restrictions, retention, security, change notification, audit rights, and data-use responsibilities before adoption. D. Permit unrestricted prompts provided that users sign an informal agreement. Correct Answer: C. Assess contractual restrictions, retention, security, change notification, audit rights, and data-use responsibilities before adoption.
|
