Skip to content

AIF-C01 Task 5.2: Governance and Compliance Regulations for AI Systems

Overview

Governance and compliance for AI systems means ensuring that AI solutions:

  • Use data appropriately throughout its lifecycle.
  • Meet organizational, legal, regulatory, and contractual requirements.
  • Have traceable activity and change records.
  • Apply appropriate security and privacy controls.
  • Assign responsibility between AWS and the customer.
  • Are reviewed regularly as models, data, risks, and regulations change.

For the exam, focus on recognizing the purpose of AWS governance services and understanding practical data governance processes. You are not expected to design a complex legal or regulatory compliance program.

1. AWS Services and Features That Support Governance

AWS Config

AWS Config records and evaluates the configuration of AWS resources.

It can help organizations:

  • Track resource configuration changes.
  • Determine whether resources comply with organizational rules.
  • Evaluate resources against AWS Config rules.
  • Identify resources that do not meet requirements.
  • Maintain a history of resource configurations for auditing.

An organization might use AWS Config to check whether:

  • An Amazon S3 bucket containing AI training data is encrypted.
  • An IAM policy allows overly broad access.
  • Resources are deployed only in approved AWS Regions.
  • Logging is enabled for an AI workload.
  • Certain resources have required tags, such as data classification or owner.

Important distinction

AWS Config answers:

“Is the resource configured according to the required rule?”

It does not primarily answer:

“Who made the API call?”

That is the role of AWS CloudTrail.

Amazon Inspector

Amazon Inspector automatically discovers and scans certain AWS workloads for software vulnerabilities and unintended network exposure.

It can help identify vulnerabilities in:

  • Amazon EC2 instances.
  • Container images in Amazon ECR.
  • AWS Lambda functions, where supported.

Amazon Inspector can help identify:

  • Vulnerable operating system packages on an EC2 instance hosting an AI application.
  • Vulnerable dependencies in a container running an inference service.
  • Software vulnerabilities in a Lambda-based AI workflow.

Important distinction

Amazon Inspector focuses primarily on vulnerability management.

It does not:

  • Certify that an AI solution complies with a specific regulation.
  • Replace AWS Config for configuration compliance.
  • Replace CloudTrail for API activity auditing.

AWS Artifact

AWS Artifact provides access to AWS security and compliance documentation.

Examples include:

  • AWS audit reports.
  • Certifications.
  • Attestations.
  • Agreements and related compliance documents.

Organizations can use these documents as evidence when evaluating whether AWS services support their compliance obligations.

Exam trap

AWS Artifact does not automatically make a customer’s application compliant.

The customer must still configure and operate the workload correctly under the AWS Shared Responsibility Model.

For example:

  • AWS may provide an applicable compliance report for a service.
  • The customer is still responsible for appropriate IAM permissions, data classification, encryption configuration, retention, and application controls.

AWS CloudTrail

AWS CloudTrail records AWS API activity and helps provide an audit trail.

CloudTrail can record information such as:

  • The identity that made an API call.
  • The requested action.
  • The affected resource.
  • The time of the action.
  • The source IP address.
  • Whether the action succeeded or failed.

CloudTrail can help determine:

  • Who changed an Amazon Bedrock configuration.
  • Who modified an IAM policy used by an AI application.
  • Who created or deleted an S3 bucket containing training data.
  • Who changed logging or encryption settings.
  • Which principal invoked a management API.

Important distinction

CloudTrail is primarily an API and activity audit service. It is not a general-purpose application log or model-quality monitoring service.

Also, API audit logs may not automatically contain the complete prompt and response content. If prompt and response logging is required, it must be configured using the relevant AI service features, such as invocation logging where supported. This must be balanced against privacy and sensitive-data risks.

AWS Trusted Advisor

AWS Trusted Advisor provides recommendations based on AWS best practices.

Recommendations can relate to:

  • Security.
  • Fault tolerance.
  • Performance.
  • Cost optimization.
  • Service limits.

Trusted Advisor may help identify:

  • Publicly accessible resources.
  • Security configuration concerns.
  • Underutilized resources.
  • Service limit risks for an AI workload.
  • Opportunities to improve resiliency or reduce cost.

Important distinction

Trusted Advisor provides recommendations. It does not replace:

  • AWS Config for continuous configuration compliance.
  • CloudTrail for detailed API auditing.
  • AWS Artifact for compliance reports.
  • A formal legal or regulatory assessment.

2. Other AWS Governance Features to Recognize

Although the objective lists specific examples, the following features commonly support AI governance.

AWS Identity and Access Management

AWS IAM controls authentication and authorization.

Use IAM to apply:

  • Least-privilege permissions.
  • Role-based access.
  • Separate permissions for developers, data scientists, operators, and auditors.
  • Temporary credentials where appropriate.
  • Restrictions on access to training data, model artifacts, prompts, and outputs.

For multi-account governance, AWS Organizations and service control policies can provide organization-level guardrails.

AWS Organizations and Service Control Policies

AWS Organizations helps centrally manage multiple AWS accounts.

Service control policies, or SCPs, provide permission guardrails across accounts. An SCP does not grant permissions by itself; it limits the maximum permissions available to IAM principals in member accounts.

Examples:

  • Prevent use of unapproved AWS Regions.
  • Prevent disabling CloudTrail.
  • Restrict access to certain services.
  • Require workloads to remain within approved organizational boundaries.

Amazon S3 and Data Controls

Amazon S3 can support governance through:

  • Encryption.
  • Bucket policies.
  • Versioning.
  • Object Lock, where appropriate.
  • Lifecycle rules.
  • Access logging and monitoring.
  • Block Public Access.
  • Replication controls.

The exact control must match the data’s classification and retention requirements.

AWS Audit Manager

AWS Audit Manager can help collect and organize evidence for audits against supported frameworks.

It can assist with:

  • Evidence collection.
  • Control mapping.
  • Audit preparation.
  • Tracking control effectiveness.

It does not automatically prove that an organization satisfies every legal or regulatory requirement.

Amazon Macie

Amazon Macie can help discover and identify sensitive data in Amazon S3, such as personally identifiable information.

It may support governance by helping organizations:

  • Locate sensitive data.
  • Identify unexpected exposure.
  • Improve data classification.
  • Prioritize security controls.

3. AWS Shared Responsibility Model

Compliance responsibilities are divided between AWS and the customer.

AWS responsibility

AWS is responsible for security of the cloud, including:

  • Physical facilities.
  • Hardware.
  • Networking infrastructure.
  • Core infrastructure services.
  • Certain managed-service components.

Customer responsibility

The customer is responsible for security in the cloud, including controls such as:

  • Data classification.
  • IAM permissions.
  • Encryption choices and key policies.
  • Network configuration.
  • Application security.
  • Logging and monitoring.
  • Data retention and deletion.
  • AI prompts, outputs, and training data governance.
  • Regulatory use of the AI solution.

The exact split depends on the AWS service. A more managed service generally reduces the customer’s infrastructure-management burden, but it does not eliminate the customer’s responsibility for data, access, and application governance.

4. Data Governance Strategies for AI Systems

Data governance defines how data is collected, classified, used, stored, monitored, shared, and deleted.

AI systems require careful governance because their data may include:

  • Training data.
  • Fine-tuning data.
  • Retrieval-augmented generation documents.
  • Prompts and user inputs.
  • Model outputs.
  • Evaluation data.
  • Logs and telemetry.
  • Model artifacts and configuration.

4.1 Data classification

Classify data before using it in an AI system.

Common categories include:

  • Public.
  • Internal.
  • Confidential.
  • Sensitive or regulated.
  • Personally identifiable information.
  • Financial, health, or other restricted data.

Classification should determine:

  • Who may access the data.
  • Where it may be stored.
  • Whether it may be used for training or fine-tuning.
  • How long it may be retained.
  • Whether it must be encrypted or anonymized.
  • Whether it may be sent to an external service.

4.2 Data lifecycle

A typical data lifecycle includes:

  1. Collection
    Collect data for a defined and legitimate purpose.

  2. Ingestion
    Validate, classify, and apply access controls before loading data into an AI pipeline.

  3. Storage
    Store data in approved locations with appropriate encryption and access policies.

  4. Processing and use
    Use data only for approved purposes and monitor how it is accessed.

  5. Sharing or transfer
    Control sharing with teams, accounts, Regions, services, or third parties.

  6. Archiving
    Move data to lower-cost or long-term storage when it is no longer actively used but must be retained.

  7. Deletion or destruction
    Delete data when the retention period ends or when required by policy or regulation.

A governance policy should define the lifecycle for both source data and derived data, including embeddings, indexes, cached data, model artifacts, logs, and backups.

4.3 Data residency and geographic requirements

Data residency requirements specify where data must be stored or processed.

Controls can include:

  • Selecting approved AWS Regions.
  • Restricting cross-Region replication.
  • Applying IAM, SCP, and resource policies.
  • Reviewing service-specific data processing behavior.
  • Validating where logs, backups, and model artifacts are stored.
  • Checking whether a feature uses cross-Region inference or processing.

Exam tip

Choosing an AWS Region alone does not automatically guarantee complete residency compliance. The organization must also understand:

  • Service behavior.
  • Replication.
  • Backups.
  • Support access.
  • Cross-Region features.
  • Third-party integrations.
  • Applicable contractual and regulatory requirements.

4.4 Logging and auditability

Logging helps demonstrate what happened and supports investigations.

Useful logging categories include:

  • AWS API activity.
  • Resource configuration changes.
  • Application events.
  • Data access.
  • Model invocation activity, where supported.
  • Security findings.
  • Administrative actions.
  • Approval and review records.

Logs should be:

  • Protected against unauthorized modification.
  • Accessible only to authorized personnel.
  • Retained for the required period.
  • Monitored for suspicious activity.
  • Stored in an appropriate Region or account.
  • Excluded from unnecessary sensitive content.

Sensitive prompt and response data

Prompt and response logs can contain:

  • Personal information.
  • Confidential business data.
  • Credentials accidentally entered by users.
  • Regulated information.

Therefore, enable detailed invocation logging only when justified, and apply appropriate masking, access, encryption, retention, and monitoring controls.

4.5 Retention

Retention specifies how long data and records must be kept.

Retention schedules may apply to:

  • Training and fine-tuning data.
  • Prompts and outputs.
  • Audit logs.
  • Security findings.
  • Model versions.
  • Evaluation results.
  • Approval records.
  • Backups and replicas.

A good retention policy defines:

  • The retention period.
  • The business or regulatory reason.
  • The data owner.
  • The storage location.
  • Who may access the data.
  • How deletion is verified.

Exam trap

“Keep everything forever” is not automatically good governance. Excessive retention increases exposure, storage cost, and privacy risk.

4.6 Monitoring and observation

Monitoring helps identify security, compliance, and operational issues.

Monitor for:

  • Unauthorized access.
  • Configuration drift.
  • Unexpected data transfers.
  • Unusual invocation patterns.
  • Excessive or abusive use.
  • Failed requests.
  • Security findings.
  • Changes in model or application behavior.
  • Policy violations.

Different tools serve different monitoring needs:

Requirement Suitable AWS capability
AWS API activity AWS CloudTrail
Resource configuration compliance AWS Config
Software vulnerabilities Amazon Inspector
Security and best-practice recommendations AWS Trusted Advisor
Sensitive data discovery in S3 Amazon Macie
Application and service logs Amazon CloudWatch and service-specific logging
Compliance reports from AWS AWS Artifact

5. Governance Processes and Protocols

5.1 Establish policies

AI governance policies should define:

  • Approved AI use cases.
  • Prohibited or restricted uses.
  • Permitted data types.
  • Data classification requirements.
  • Human approval requirements.
  • Security and privacy controls.
  • Model evaluation expectations.
  • Logging and monitoring requirements.
  • Incident response procedures.
  • Retention and deletion requirements.
  • Ownership and accountability.
  • Requirements for third-party or foundation models.

Policies should be understandable, documented, communicated, and enforceable.

5.2 Define roles and accountability

Assign ownership for:

  • Data.
  • The AI application.
  • Model selection and configuration.
  • Security.
  • Privacy.
  • Compliance.
  • Risk acceptance.
  • Monitoring.
  • Incident response.
  • Final business decisions.

Do not assume that the AI model provider owns the customer’s compliance obligations.

5.3 Use a review cadence

AI systems should be reviewed periodically and when significant changes occur.

Review triggers may include:

  • A new model or model version.
  • A new data source.
  • A change in the intended use.
  • A change in geographic deployment.
  • A new regulation or contractual requirement.
  • A security incident.
  • A significant change in model behavior.
  • A change to prompt templates or retrieval sources.
  • A change to data retention or logging.

Reviews should not happen only once at deployment.

5.4 Perform risk-based reviews

Not every AI use case requires the same level of governance.

A low-risk internal summarization use case may require fewer controls than an AI system that affects:

  • Employment decisions.
  • Credit or insurance decisions.
  • Medical decisions.
  • Legal outcomes.
  • Access to essential services.
  • Safety-critical operations.

A risk-based review can consider:

  • The sensitivity of the data.
  • The impact of incorrect output.
  • The degree of automation.
  • The affected population.
  • The explainability requirement.
  • The possibility of discrimination or bias.
  • The ability to provide human review.
  • The consequences of data leakage.

5.5 Apply human oversight

Human oversight is especially important when AI output could materially affect individuals or business operations.

Controls can include:

  • Human approval before taking action.
  • Escalation of uncertain or high-risk cases.
  • Separation between AI recommendation and final decision.
  • Review of model outputs.
  • Ability to override or disable the system.
  • Clear accountability for the final decision.

5.6 Train teams

Personnel should receive training appropriate to their responsibilities.

Training may cover:

  • Approved and prohibited AI uses.
  • Handling confidential and personal data.
  • Prompt security.
  • Avoiding secrets in prompts.
  • Recognizing hallucinations and unreliable output.
  • Reporting incidents.
  • Review and approval procedures.
  • Applicable policies and regulations.
  • Secure use of third-party AI tools.

Training should be refreshed periodically and when policies or risks change.

5.7 Document transparency

Transparency helps users and affected parties understand how an AI system is used.

Documentation may include:

  • The purpose of the AI system.
  • The types of data used.
  • The model or service used.
  • Known limitations.
  • Expected accuracy or performance.
  • Human oversight.
  • How decisions are reviewed or challenged.
  • How personal data is handled.
  • How users can report problems.

Transparency requirements vary by use case and regulation. The exam generally tests the principle that organizations should document and communicate AI use appropriately, not the details of a particular legal standard.

6. Generative AI Security Scoping Matrix

The Generative AI Security Scoping Matrix is an AWS framework that helps organizations understand security responsibilities for different generative AI architectures and usage patterns.

It helps answer questions such as:

  • Who is responsible for securing the application?
  • Who controls the model?
  • Who protects the underlying infrastructure?
  • Who governs prompts, outputs, and customer data?
  • Which controls are managed by AWS and which remain with the customer?
  • How does responsibility change when using a managed model service versus hosting or customizing a model?

How to use it

Use the matrix during architecture and governance reviews to:

  1. Identify how the generative AI solution is being used.
  2. Identify the components controlled by AWS, the customer, or another provider.
  3. Map security responsibilities to each layer.
  4. Determine which controls and evidence are required.
  5. Identify gaps in logging, access control, data protection, monitoring, and incident response.

Exam-level takeaway

The matrix is:

  • A security and responsibility-scoping framework.
  • Useful for understanding control ownership.
  • Helpful when reviewing different generative AI deployment patterns.

It is not:

  • A compliance certification.
  • A replacement for IAM, CloudTrail, Config, or other controls.
  • A guarantee that a workload satisfies a regulation.
  • A single automated security tool.

7. Comparing Common Governance Services

Service Primary purpose Typical exam clue
AWS Config Evaluate and track resource configuration compliance “Determine whether resources comply with configuration rules”
Amazon Inspector Find software vulnerabilities and unintended network exposure “Scan EC2, containers, or Lambda for vulnerabilities”
AWS Artifact Access AWS compliance reports and agreements “Obtain AWS audit reports or certifications”
AWS CloudTrail Record AWS API activity for auditing “Identify who changed or accessed a resource through an API”
AWS Trusted Advisor Provide best-practice recommendations “Recommend security, cost, performance, or resiliency improvements”
IAM Control identity and permissions “Apply least privilege”
AWS Organizations and SCPs Apply multi-account governance guardrails “Restrict Regions or services across accounts”
Amazon Macie Discover sensitive data in Amazon S3 “Find PII or sensitive data in S3”
AWS Audit Manager Collect and organize audit evidence “Prepare evidence for an audit framework”
Amazon CloudWatch Monitor logs, metrics, and alarms “Monitor application or service behavior”

8. Exam Tips and Common Traps

Tip 1: Match the service to the question

  • Configuration compliance → AWS Config.
  • API audit trail → AWS CloudTrail.
  • AWS compliance documentation → AWS Artifact.
  • Software vulnerability scanning → Amazon Inspector.
  • Best-practice recommendations → AWS Trusted Advisor.
  • Sensitive data discovery in S3 → Amazon Macie.
  • Identity and permissions → IAM.

Tip 2: Artifact is evidence, not automatic compliance

AWS Artifact provides AWS documentation. It does not configure the customer’s workload or certify the customer’s application.

Tip 3: CloudTrail is not the same as CloudWatch

  • CloudTrail records AWS API activity.
  • CloudWatch monitors logs, metrics, and operational events.

They are often used together.

Tip 4: Config is not a vulnerability scanner

AWS Config checks configuration against rules. Amazon Inspector scans supported workloads for vulnerabilities.

Tip 5: Compliance is a shared responsibility

Using a compliant AWS service does not automatically make the customer’s entire AI solution compliant.

Tip 6: Data governance includes derived data

Governance should cover more than original datasets. Consider:

  • Embeddings.
  • Vector indexes.
  • Caches.
  • Fine-tuned models.
  • Prompt and response logs.
  • Backups.
  • Evaluation datasets.
  • Generated outputs.

Tip 7: Residency includes processing and replication

Review cross-Region features, backups, logs, and third-party integrations—not just the Region where the primary resource is created.

Tip 8: Logging can create privacy risk

Detailed prompt and response logging improves auditability but may store sensitive information. The correct answer may involve controlled logging, redaction, encryption, restricted access, and retention limits.

Tip 9: Governance is continuous

A governance process should include ongoing monitoring, periodic reviews, change-triggered reviews, training, and incident response.

Tip 10: Do not over-select services

If the question asks for the service that provides AWS compliance reports, choose AWS Artifact, not Config, Inspector, or Trusted Advisor.

Practice Questions and Answers ## Question 1 A company must determine whether all S3 buckets used by an AI application comply with a rule requiring encryption and blocking public access. Which AWS service should the company use? A. AWS CloudTrail B. AWS Config C. Amazon Inspector D. AWS Artifact **Answer: B. AWS Config** **Explanation:** AWS Config evaluates resource configurations against rules and can identify noncompliant S3 buckets. CloudTrail records API activity, Inspector scans supported workloads for vulnerabilities, and Artifact provides AWS compliance documentation. --- ## Question 2 A compliance team needs AWS audit reports and certifications to support an assessment of a generative AI workload running on AWS. Which service should they use? A. AWS Trusted Advisor B. AWS Config C. AWS Artifact D. Amazon CloudWatch **Answer: C. AWS Artifact** **Explanation:** AWS Artifact provides access to AWS security and compliance reports, certifications, and agreements. --- ## Question 3 An organization wants to determine which administrator changed an IAM policy that controls access to an AI training-data bucket. Which AWS service provides this information? A. Amazon Inspector B. AWS CloudTrail C. AWS Trusted Advisor D. Amazon Macie **Answer: B. AWS CloudTrail** **Explanation:** CloudTrail records AWS API activity, including the identity, action, resource, and time associated with an API call. --- ## Question 4 A company is hosting an AI inference application on EC2 instances and wants to identify vulnerable operating system packages. Which service should it use? A. AWS Config B. AWS Artifact C. Amazon Inspector D. AWS Organizations **Answer: C. Amazon Inspector** **Explanation:** Amazon Inspector identifies software vulnerabilities and certain network exposure risks in supported workloads such as EC2 instances. --- ## Question 5 A company wants to ensure that an AI workload is deployed only in approved AWS Regions across all member accounts. Which approach best supports this requirement? A. Use AWS Artifact to download compliance reports. B. Use Amazon Inspector to scan the workload. C. Use AWS Organizations with service control policies and supporting configuration checks. D. Use CloudTrail to automatically prevent deployments in unapproved Regions. **Answer: C. Use AWS Organizations with service control policies and supporting configuration checks.** **Explanation:** SCPs can provide organization-wide permission guardrails, including restrictions on Regions. CloudTrail records activity but does not normally prevent actions. --- ## Question 6 An organization enables detailed logging of prompts and model responses for an AI application. What is the most important governance consideration? A. Prompt and response logs never contain sensitive information. B. Prompt and response logging eliminates the need for IAM. C. Logs may contain sensitive data and require appropriate access control, encryption, monitoring, and retention. D. Logs should always be retained indefinitely. **Answer: C. Logs may contain sensitive data and require appropriate access control, encryption, monitoring, and retention.** **Explanation:** Prompts and outputs may contain personal, confidential, or regulated information. Logging must be designed with privacy and retention requirements in mind. --- ## Question 7 A company is reviewing a generative AI architecture to determine which security controls are the responsibility of AWS and which are the customer’s responsibility. Which AWS resource is most directly relevant? A. Generative AI Security Scoping Matrix B. Amazon Inspector C. AWS Trusted Advisor cost checks D. Amazon Macie **Answer: A. Generative AI Security Scoping Matrix** **Explanation:** The matrix helps organizations understand security responsibilities for different generative AI architectures and usage patterns. It does not replace the implementation of security controls. --- ## Question 8 Which practice best represents effective AI data governance? A. Retain all prompts, outputs, and training data indefinitely. B. Apply the same controls to all data regardless of sensitivity. C. Define data classification, approved use, residency, retention, access, monitoring, and deletion requirements. D. Rely exclusively on the model provider to manage customer data compliance. **Answer: C. Define data classification, approved use, residency, retention, access, monitoring, and deletion requirements.** **Explanation:** Effective data governance manages data throughout its lifecycle and applies controls based on sensitivity, purpose, and applicable requirements.

Final Summary

For Task Statement 5.2, remember these core associations:

  • AWS Config evaluates resource configuration compliance.
  • Amazon Inspector identifies vulnerabilities in supported workloads.
  • AWS Artifact provides AWS compliance reports and agreements.
  • AWS CloudTrail records AWS API activity for auditing.
  • AWS Trusted Advisor gives best-practice recommendations.
  • IAM and Organizations enforce access and organizational guardrails.
  • Data governance covers classification, lifecycle, residency, logging, monitoring, retention, and deletion.
  • AI governance requires documented policies, assigned accountability, periodic and risk-based reviews, transparency, human oversight, and team training.
  • The Generative AI Security Scoping Matrix helps identify responsibility boundaries; it is not itself a certification or automated compliance control.