AWS AI Practitioner Exam Notes

Task Statement 5.2: Recognize Governance and Compliance Regulations for AI Systems

This task statement focuses on how organizations govern AI systems, demonstrate compliance, protect data, monitor risks, and assign responsibilities between AWS and the customer.

1. AWS Shared Responsibility Model

Security and compliance in AWS are shared responsibilities.

AWS responsibilities: Security of the cloud

AWS is responsible for the underlying infrastructure, including:

Physical data centers
Networking infrastructure
Hardware and facilities
Host operating systems
Physical access controls
Environmental controls
Infrastructure certifications and attestations

AWS maintains certifications and third-party audit reports for many global, regional, and industry standards.

Customer responsibilities: Security in the cloud

The customer is responsible for configuring and operating their workloads securely and compliantly, including:

IAM permissions
Data classification
Data encryption
Network configuration
Application security
Model and endpoint security
Logging and monitoring
Data retention and deletion
AI governance policies
Bias and performance monitoring
Compliance of customer-generated data and processes

The exact division depends on the service.

Example

For Amazon S3:

AWS secures the S3 infrastructure.
The customer controls bucket policies, IAM permissions, encryption settings, versioning, lifecycle rules, and the data stored in the bucket.

For Amazon Bedrock:

AWS manages the Bedrock service infrastructure and foundation model service.
The customer is responsible for prompts, retrieved data, application access, guardrail configuration, output handling, and compliance of the overall application.

Exam tip

Do not assume that using a managed AI service makes the customer automatically compliant. AWS provides compliant infrastructure and services, but customers must configure and use them correctly.

2. AWS Artifact

Purpose

AWS Artifact provides on-demand access to AWS security and compliance documentation.

It includes reports and certifications produced by third-party auditors, such as:

SOC reports
ISO certifications
Payment Card Industry-related documentation
Regional and industry compliance reports
Other AWS compliance attestations

Each document generally identifies:

The controls evaluated
The reporting or certification period
The scope of the assessment
The standards covered

How Artifact helps audits

Customers inherit some compliance controls from AWS. For example, an auditor does not need to physically inspect an AWS data center if AWS has already provided an independent audit report covering the relevant physical controls.

The customer’s auditors can use AWS Artifact reports as evidence for the AWS-controlled portion of the environment. They then focus on customer-controlled processes, such as:

IAM configuration
Data protection
Application controls
Logging
Operational procedures
AI governance processes

Exam trap

AWS Artifact is not an auditing service for the customer’s applications.
It does not automatically configure controls.
It provides AWS compliance reports and agreements.
AWS Audit Manager, not Artifact, collects customer evidence and creates assessments.

3. AWS Compliance and AI Regulations

AI-related regulations and standards are still evolving. Organizations may need to comply with:

General information security standards
Privacy regulations
Industry-specific requirements
AI-specific standards and regulations
Internal responsible AI policies

ISO/IEC 42001

ISO/IEC 42001 is an AI management system standard. It provides a structured approach for organizations to manage AI responsibly.

It supports practices such as:

AI governance
Risk management
Documentation
Accountability
Continuous improvement
Responsible AI controls

ISO/IEC 23894

ISO/IEC 23894 provides guidance for managing risks related to AI systems.

It helps organizations:

Identify AI risks
Assess likelihood and impact
Apply risk controls
Monitor residual risk
Support responsible development and deployment

These standards are generally voluntary frameworks or certifications, not automatically applicable laws.

EU AI Act

The EU AI Act uses a risk-based approach to regulating AI systems.

Unacceptable-risk systems

Some AI applications are prohibited or heavily restricted, such as:

Certain social scoring systems
Certain facial-image scraping databases
Certain emotion-inference systems used in workplaces or educational environments

High-risk systems

High-risk applications are permitted but subject to significant requirements. Examples can include AI systems used in areas such as:

Employment and recruitment
Critical infrastructure
Education
Essential services
Law enforcement
Safety-related decisions

Typical requirements include:

Risk management systems
Data governance
Technical documentation
Record keeping
Transparency
Human oversight
Accuracy and robustness
Cybersecurity controls
Post-deployment monitoring

Lower-risk systems

Systems that are not prohibited or classified as high-risk generally have fewer requirements, although transparency obligations may still apply in some cases.

Exam tip

The exam may simplify the EU AI Act into three broad categories:

Unacceptable risk
High risk
Other or largely unregulated systems

Remember the core concept: AI regulation is risk-based, and high-risk systems require stronger governance and documentation.

4. NIST AI Risk Management Framework

The NIST AI Risk Management Framework, or AI RMF, provides voluntary guidance for managing AI risks.

It is designed for organizations that:

Design AI systems
Develop AI systems
Deploy AI systems
Use AI systems

Four AI RMF functions

1. Govern

Establish organizational policies, accountability, roles, and processes.

Examples:

Assign AI governance responsibilities
Create policies
Define risk tolerances
Establish documentation requirements
Train employees
Define review and escalation processes

2. Map

Understand the AI system and its context.

Examples:

Identify the intended use case
Identify stakeholders
Identify affected users
Identify data sources
Identify potential harms
Identify legal and regulatory requirements
Define the system’s operating environment

3. Measure

Evaluate and monitor AI risks.

Examples:

Test model accuracy
Test for bias
Assess security vulnerabilities
Measure explainability
Monitor drift
Evaluate harmful outputs
Review system performance

4. Manage

Prioritize and respond to identified risks.

Examples:

Apply controls
Reduce or eliminate risk
Accept residual risk
Transfer risk
Continuously monitor
Escalate unacceptable outcomes

Risk calculation

A simplified risk calculation is:

Risk = Likelihood × Severity of consequences

For an AI system, identify:

The use case
Relevant stakeholders
Potentially harmful events
Likelihood of each event
Severity of each consequence
Required mitigations
Remaining residual risk

Inherent risk

Risk that exists before controls are applied.

Residual risk

Risk remaining after safeguards and mitigations are implemented.

The overall system risk is often determined by focusing on the highest residual risks, rather than simply averaging all risks.

Exam trap

Do not confuse:

Inherent risk: before mitigation
Residual risk: after mitigation

Controls may reduce likelihood or impact, but they do not always eliminate risk.

5. Explainability, Transparency, and Bias

Transparency

Users should understand when they are interacting with AI rather than a human, especially for generative AI applications.

Transparency can include:

Disclosing that content was generated by AI
Explaining the purpose of the AI system
Describing what data is used
Explaining system limitations
Providing information about automated decision-making
Explaining how users can challenge or appeal decisions

Explainability

Explainability is the ability to understand how an AI system arrived at a result or inference.

Model-agnostic explainability

Treat the model as a black box and analyze its behavior based on inputs and outputs.

This can help determine:

Which input features influence results
How changing an input changes the output
Whether the model behaves consistently

Interpretable models

Some models are easier to understand internally, such as:

Decision trees
Rule-based systems
Linear models

There may be a tradeoff between:

Model performance
Accuracy
Complexity
Interpretability
Regulatory requirements

Bias

AI governance requires testing for inappropriate bias in:

Training data
Validation data
Model outputs
Decision thresholds
Features
Data collection processes

Sensitive or protected attributes may include:

Race
Sex
Gender identity
Religion
Political affiliation
Geographic location
Urban or rural residence

Amazon SageMaker Clarify

SageMaker Clarify helps organizations:

Analyze potential bias in datasets
Analyze bias in model predictions
Understand feature importance
Explain model predictions
Monitor for bias drift
Monitor for feature attribution drift

Exam tip

SageMaker Clarify is associated with bias detection, fairness analysis, and explainability.

It is not primarily a data storage, access-control, or audit-reporting service.

6. AWS Audit Manager

Purpose

AWS Audit Manager helps collect evidence of compliance and create audit-ready assessments.

It maps compliance requirements to AWS usage and configuration data.

Main capabilities

Audit Manager can:

Create assessments
Use built-in compliance frameworks
Create custom frameworks
Collect evidence automatically
Organize evidence by control
Generate auditor-friendly reports
Help demonstrate that controls are operating effectively

A framework is a collection of related controls.

Examples include frameworks for:

SOC 2
Generative AI best practices
Organization-specific requirements
Industry-specific requirements

Typical process

Select or create a framework.
Create an assessment.
Audit Manager evaluates relevant AWS resources.
Evidence is collected.
Evidence is mapped to controls.
Review evidence.
Generate an assessment report for auditors.

Exam distinction

Service	Main purpose
AWS Artifact	Access AWS compliance reports and certifications
AWS Audit Manager	Collect customer evidence and prepare audit assessments
AWS Config	Track resource configuration and evaluate configuration rules
AWS CloudTrail	Record API activity and account actions
Amazon Inspector	Find application and infrastructure vulnerabilities
AWS Trusted Advisor	Recommend improvements based on AWS best practices

7. Amazon Bedrock Guardrails

Purpose

Amazon Bedrock Guardrails applies application-specific safeguards to generative AI workloads.

Guardrails can evaluate both:

User inputs
Foundation model responses

Main controls

Guardrails can help detect or block:

Hate
Insults
Sexual content
Violence
Restricted topics
Personally identifiable information, or PII

Denied topics

You can define topics that the application should avoid using natural-language descriptions.

For example:

Investment advice
Medical diagnosis
Political persuasion
Legal advice
Competitor discussions

You can provide example phrases to improve topic recognition.

PII controls

Guardrails can detect PII in prompts and responses.

Depending on the use case, the application may:

Block input containing PII
Redact PII from model responses
Allow the request but apply a defined response strategy

Configurable thresholds

Content filters can be configured with thresholds for different categories. The application can also define different responses when:

The user prompt is blocked
The model response is blocked

Exam trap

Bedrock Guardrails is an application-level generative AI safety control.

It is not the same as:

IAM, which controls access
AWS Config, which checks resource configuration
Amazon Macie, which discovers sensitive data in S3
SageMaker Clarify, which evaluates model bias and explainability

8. AWS Config

Purpose

AWS Config provides resource-level configuration visibility and compliance evaluation.

It can:

Record AWS resource configurations
Maintain configuration history
Capture configuration changes
Evaluate resources against rules
Identify noncompliant resources
Trigger remediation actions

Config rules

Rules can be:

AWS-managed or prebuilt rules
Custom rules implemented with AWS Lambda

Examples of checks include:

Whether encryption is enabled
Whether public access is allowed
Whether required tags are present
Whether security groups are overly permissive
Whether resources meet organizational policies

Remediation

When a resource becomes noncompliant, AWS Config can initiate remediation using an AWS Systems Manager Automation document.

Conformance packs

Conformance packs group together:

AWS Config rules
Remediation actions
Compliance requirements

Useful examples include packs for:

AI and machine learning operational best practices
Amazon SageMaker security best practices

Exam distinction

AWS Config focuses on resource configuration and compliance state.

It does not primarily:

Scan software packages for vulnerabilities — Amazon Inspector
Record API calls — CloudTrail
Provide AWS compliance reports — Artifact
Generate audit evidence — Audit Manager

9. Amazon Inspector

Purpose

Amazon Inspector performs automated vulnerability assessments.

It can assess:

EC2 instances
Container images
Lambda functions, depending on the configured assessment capabilities
Software packages
Application dependencies
Network exposure

Inspector identifies issues such as:

Vulnerable software versions
Known security vulnerabilities
Open access to resources
Deviations from security best practices

Findings

Inspector provides findings that include:

Severity
Description
Affected resource
Vulnerability information
Recommended remediation

Exam distinction

A useful way to remember the difference:

AWS Config: “Is the resource configured according to policy?”
Amazon Inspector: “Does the resource or application have security vulnerabilities?”

10. AWS CloudTrail

Although not discussed in detail in the transcript, AWS CloudTrail is specifically included in the task objective.

Purpose

CloudTrail records AWS account activity and API actions.

It can capture:

Who performed an action
What action was performed
When it occurred
Which resource was affected
From where the request originated
Whether the request succeeded or failed

Examples:

A user changed an S3 bucket policy.
An administrator modified an IAM role.
A model endpoint was created or deleted.
A security group rule was changed.
A logging configuration was disabled.

Governance uses

CloudTrail supports:

Audit trails
Incident investigation
Accountability
Change tracking
Detection of unauthorized activity
Compliance evidence
Forensic analysis

CloudTrail logs should generally be protected from alteration through controls such as:

Centralized logging accounts
Restricted access
Encryption
Log file validation
S3 Object Lock where appropriate
Retention policies
Monitoring for changes to logging configuration

Exam trap

CloudTrail records API activity.
AWS Config records resource configuration state and changes.

They are complementary, not interchangeable.

11. AWS Trusted Advisor

Purpose

AWS Trusted Advisor evaluates an AWS environment against AWS best practices.

It provides recommendations across categories such as:

Cost optimization
Performance
Security
Fault tolerance or resilience
Operational excellence
Service limits

Governance and compliance relevance

Trusted Advisor can identify issues such as:

Exposed resources
Weak security configurations
Unused or inefficient resources
Service-limit concerns
Resilience weaknesses
Operational best-practice deviations

Exam trap

Trusted Advisor provides recommendations. It is not a complete compliance certification or formal audit service.

12. Data Governance

Definition

Data governance is the combination of people, processes, and technology used to manage:

Availability
Usability
Integrity
Quality
Security
Privacy
Appropriate access
Regulatory compliance

The objective is to make data trustworthy and useful without allowing it to be misused.

Three major areas of data governance

1. Curation

Identify and manage important data assets, including:

Databases
Data lakes
Data warehouses
Training datasets
Evaluation datasets
Prompt and response data
Model artifacts

Curation includes ensuring that data is:

Accurate
Current
Relevant
Properly classified
Free of unnecessary sensitive information
Not duplicated or unnecessarily transformed

2. Discovery and understanding

Users should be able to find and understand data.

This includes:

Data catalogs
Business metadata
Technical metadata
Data dictionaries
Ownership information
Data lineage
Data quality scores
Access request processes

3. Protection

Protect data while allowing appropriate business access.

Protection includes:

Authentication
Authorization
Encryption
Network controls
Data masking
Tokenization
PII detection
Least privilege
Temporary access
Monitoring
Retention and deletion controls

13. Data Governance Roles

Data owner

A data owner is generally accountable for business and policy decisions concerning a data domain.

Responsibilities may include:

Defining access policies
Determining regulatory requirements
Approving who can access data
Defining retention requirements
Establishing classification rules
Making policy decisions for sensitive data

The data owner is often a senior business or executive-level role.

Data steward

A data steward handles the practical, day-to-day management of data.

Responsibilities may include:

Understanding data meaning and context
Identifying data quality issues
Coordinating corrections
Supporting projects
Helping define quality rules
Assisting users with data discovery
Escalating problems to data owners

IT and security teams

These teams may be responsible for:

Deploying governance tools
Managing infrastructure
Implementing security controls
Managing IAM
Monitoring systems
Supporting data catalogs
Enforcing technical policies

Exam tip

Remember the distinction:

Data owner: makes policy and access decisions.
Data steward: performs day-to-day data management.
IT: provides and operates the technical capabilities.

Governance should also consider segregation of duties, ensuring that one person does not have excessive control over the entire process.

14. Data Profiling, Cataloging, Quality, and Lineage

Data profiling

Data profiling examines a dataset to understand its characteristics and identify problems.

It may evaluate:

Missing values
Data types
Value distributions
Duplicates
Invalid values
Outliers
Relationships between fields
Potential sensitive information
Conformance to quality rules

Data catalog

A data catalog helps users discover and understand available datasets.

It typically contains:

Dataset locations
Schemas
Tables
Columns
Data types
Classifications
Ownership
Business descriptions
Access information

Data lineage

Data lineage shows where data came from and how it changed.

It can answer:

What was the original source?
What transformations occurred?
Which systems processed the data?
Where was the data stored?
Which reports or models use it?
What upstream data would be affected by a change?

Lineage is especially important for AI because it helps organizations explain:

Which data trained a model
Which data was used for evaluation
Where an output originated
Whether sensitive data entered a pipeline

AWS Glue DataBrew

AWS Glue DataBrew is a visual data preparation tool.

Relevant governance capabilities include:

Data profiling
Data quality validation
Visual data preparation
Data lineage visualization
Cleaning and normalization without writing code

AWS Glue Data Catalog

The Glue Data Catalog stores metadata about data sources, including:

Locations
Schemas
Table definitions
Data types
Partitions

A Glue crawler can scan data sources and populate the catalog automatically.

AWS Glue Data Quality

AWS Glue Data Quality evaluates data registered in the Glue Data Catalog.

It supports:

Data quality rules
Rule recommendations
Machine learning-based anomaly detection
Data quality jobs
Pass/fail results
Monitoring of data quality over time

Exam trap

DataBrew: visual profiling and data preparation.
Glue Data Catalog: metadata catalog.
Glue Data Quality: evaluates data quality rules.
Lake Formation: fine-grained data lake access control.

15. AWS Lake Formation

Purpose

AWS Lake Formation helps govern and secure data lakes built on Amazon S3 and cataloged with the AWS Glue Data Catalog.

It supports fine-grained permissions at:

Database level
Table level
Column level
Row level
Cell level

Lake Formation can integrate with services such as:

Amazon Athena
AWS Glue
Amazon EMR
Amazon Redshift
Machine learning services

Example

A data scientist may be allowed to:

Access anonymized customer data
Access only selected columns
Access only records for a specific region
Use data for model training but not export raw PII

Exam tip

Lake Formation is primarily about data lake governance and fine-grained access control, not model evaluation or bias detection.

16. Data Lifecycle Management

Data lifecycle management defines what happens to data from creation through deletion.

A typical lifecycle includes:

Collection
Classification
Storage
Active use
Transformation
Archival
Retention
Deletion or destruction

For AI systems, lifecycle policies should address:

Training data
Fine-tuning data
Validation data
Test data
Prompt data
Model outputs
Evaluation results
Model versions
Logs
Audit records
User feedback

Key lifecycle questions

Organizations should define:

How long should data be retained?
Is retention required by law or policy?
Who can access it?
Where can it be stored?
Can it be used for training?
When should it be archived?
When must it be deleted?
Are legal holds required?
How should backups be handled?

17. Amazon S3 Storage Classes and Retention

Training data may need to be retained for compliance even if it is rarely accessed.

S3 Standard

Use for frequently accessed data.

Characteristics:

Immediate access
Higher storage cost than archival classes
Suitable for active datasets and frequently used model inputs

S3 Standard-IA and One Zone-IA

Use for data that is infrequently accessed but must be retrieved quickly.

Characteristics:

Lower storage cost
Retrieval charges
One Zone-IA has less resilience because data is stored in one Availability Zone

S3 Intelligent-Tiering

Use when access patterns are unknown or changing.

S3 automatically moves objects between access tiers based on usage.

S3 Glacier classes

Use for long-term archival and rarely accessed data.

They are useful for:

Regulatory retention
Historical training datasets
Old model versions
Audit records
Infrequently accessed logs

Different Glacier classes provide different retrieval speeds and costs.

S3 Lifecycle rules

Lifecycle rules automatically transition or delete objects based on conditions such as object age.

A rule can:

Move objects from S3 Standard to Standard-IA
Move objects to Glacier
Move objects to Glacier Deep Archive
Delete objects after the retention period

Example

Day 0: S3 Standard
Day 5: S3 Standard-IA
Day 120: S3 Glacier Deep Archive
After 5 years: Delete, if permitted by policy

Important retention considerations

Do not delete data merely because it is not actively used. It may be needed for:

Legal retention
Audit evidence
Reproducibility
Model investigations
Incident response
Regulatory review

Conversely, retaining data indefinitely can create privacy, cost, and security risks. Retention must be based on documented requirements.

18. Data Residency and Sovereignty

Data governance should define where data may be stored and processed.

Data residency

Data residency refers to the geographic location where data is stored.

Requirements may specify:

A particular country
A particular AWS Region
A specific legal jurisdiction
Restrictions on cross-border transfers

AI-specific residency concerns

Consider the location of:

Training data
Fine-tuning data
Prompt content
Model outputs
Logs
Backups
Model artifacts
Vector databases
Third-party APIs
Human review systems

Governance controls

Organizations can use:

Region-specific deployments
AWS Organizations service control policies
IAM and resource policies
S3 bucket policies
Encryption key location and ownership
VPC controls
Centralized logging in approved Regions
Data classification
Cross-Region replication restrictions

Exam tip

AWS Regions can help meet residency requirements, but choosing a Region alone does not guarantee regulatory compliance. The customer must also control replication, backups, logs, access, and external integrations.

19. Logging, Monitoring, and Observability

AI governance requires continuous monitoring, not only an initial assessment.

Important types of logging

API and administrative logs

Use AWS CloudTrail to record:

Configuration changes
IAM activity
Resource creation and deletion
Policy changes
Access to AWS services

Resource configuration history

Use AWS Config to record:

Configuration state
Configuration changes
Compliance against rules

Application and model logs

Depending on the application, monitor:

User prompts
Model responses
Latency
Errors
Token usage
Invocation volume
Feedback
Refusals
Guardrail interventions
Retrieval results
Model version

Be careful with sensitive prompts and responses. Logging data may contain PII, confidential data, or proprietary information.

Monitoring objectives

Organizations should monitor for:

Performance degradation
Data drift
Model drift
Bias drift
Feature attribution drift
Prompt injection
Jailbreak attempts
Toxic or unsafe outputs
PII leakage
Unauthorized access
Unexpected cost increases
Availability and resilience issues
Regulatory violations

Predefined thresholds

Governance policies should define actions for threshold breaches.

For example:

Alert the owner
Block an interaction
Disable a model endpoint
Require human review
Roll back a model
Start an incident response process
Open a compliance investigation

Observation versus monitoring

Monitoring usually means collecting and evaluating metrics continuously.

Observation or review may also include:

Human review of outputs
Periodic sampling
Red-team testing
User feedback analysis
Audit review
Investigation of incidents

Both automated and human oversight may be required.

20. The Generative AI Security Scoping Matrix

The Generative AI Security Scoping Matrix helps determine the customer’s security and governance responsibilities based on how generative AI is consumed.

Scope 1 and Scope 2

These generally involve consuming a third-party AI application.

Examples:

A consumer generative AI application
An enterprise AI application managed by a third party

The customer has relatively less responsibility for:

Model training
Model infrastructure
Model resilience
Model security implementation

However, the customer still must govern:

Which employees may use the application
What data may be entered
Data classification
Privacy and legal requirements
Acceptable use
Human oversight
Output handling
Vendor risk

Scope 3, 4, and 5

These involve building or operating more of the AI solution.

The customer may be responsible for:

Training or fine-tuning data
Model selection
Data classification
Threat modeling
Access control
Security controls
Model evaluation
Endpoint resilience
Bias testing
Output monitoring
Incident response
Compliance documentation

As scope increases, customer responsibility increases.

Scope minimization strategy

When possible, begin with the least operationally burdensome solution:

Fully trained AI services
Examples: Amazon Comprehend, Amazon Translate
Pretrained foundation models through Amazon Bedrock
Bedrock with retrieval-augmented generation
Fine-tuned models, such as models available through SageMaker JumpStart
Building and operating a custom model or full AI platform

Minimizing scope can reduce:

Governance responsibilities
Security obligations
Legal and privacy risk
Model resilience responsibilities
Operational complexity
Compliance burden

Exam tip

“Use the simplest managed service that satisfies the requirement” is often the correct governance strategy.

This does not eliminate responsibility, but it reduces the amount of infrastructure and model lifecycle responsibility owned by the customer.

21. Implementing an AI Governance Strategy

A practical governance process can follow these steps.

Step 1: Identify the scope of responsibility

Determine whether the organization is:

Using a third-party AI application
Calling a managed foundation model
Adding retrieval data
Fine-tuning a model
Training a custom model
Operating its own endpoint

The scope determines governance, security, privacy, and resilience responsibilities.

Step 2: Identify stakeholders and risks

Identify:

Business owners
Data owners
Data stewards
Security teams
Legal and privacy teams
Compliance teams
Developers
ML engineers
End users
People affected by decisions

Document:

Intended use
Out-of-scope use
Potential harms
Sensitive data
Required human oversight
Regulatory requirements

Step 3: Create AI governance policies

Policies should cover:

Acceptable AI use
Prohibited use
Data classification
Prompt handling
PII handling
Access requests
Model approval
Human review
Transparency
Explainability
Bias testing
Monitoring
Incident response
Retention and deletion
Vendor review
Change management

Step 4: Train employees

Training should be role-based.

Examples:

End users: acceptable use and prohibited data
Developers: secure prompting and output validation
Data stewards: classification and quality
ML teams: bias, explainability, testing, and drift
Security teams: threat modeling and incident response
Managers and owners: risk acceptance and escalation

Step 5: Implement technical controls

Possible controls include:

IAM least privilege
Lake Formation permissions
Encryption
S3 lifecycle rules
CloudTrail logging
AWS Config rules
Conformance packs
Amazon Inspector assessments
Bedrock Guardrails
SageMaker Clarify
Network isolation
Secrets management
Human approval workflows

Step 6: Define monitoring and escalation

Define:

What will be monitored
Which metrics matter
Thresholds for action
Who receives alerts
When human review is required
When a model must be disabled
How incidents are documented

Step 7: Review regularly

Governance should be a continuous process.

Review:

Model performance
Bias metrics
Data quality
Security findings
Guardrail events
Access permissions
Regulatory changes
User feedback
Incidents
Model and dataset changes

22. Governance Review Cadence and Strategies

The review cadence should be risk-based.

Examples

Continuous or real-time review

Appropriate for:

High-risk model outputs
Safety-sensitive applications
PII detection
Security events
Guardrail violations
Production endpoint health

Daily or weekly review

Appropriate for:

Operational metrics
Error rates
User complaints
Cost anomalies
Access anomalies
Model drift indicators

Monthly or quarterly review

Appropriate for:

Access recertification
Policy compliance
Bias testing
Data quality trends
Vendor and model reviews
Audit evidence
Risk register updates

Annual review

Appropriate for:

Governance framework updates
Formal policy review
Employee training refresh
Certification and compliance review
Business continuity and resilience testing

High-risk AI systems may require more frequent reviews than low-risk internal productivity tools.

23. Key Service Selection Guide

Requirement	Best-fit AWS service or feature
Obtain AWS audit reports and certifications	AWS Artifact
Collect customer compliance evidence	AWS Audit Manager
Record AWS API activity	AWS CloudTrail
Track resource configuration changes	AWS Config
Evaluate configurations against compliance rules	AWS Config rules
Automatically remediate configuration issues	AWS Config + Systems Manager Automation
Package compliance controls	AWS Config conformance packs
Find software and container vulnerabilities	Amazon Inspector
Review AWS best-practice recommendations	AWS Trusted Advisor
Filter harmful generative AI content	Amazon Bedrock Guardrails
Detect or redact PII in Bedrock interactions	Amazon Bedrock Guardrails
Analyze model bias and explainability	SageMaker Clarify
Catalog data metadata	AWS Glue Data Catalog
Profile and prepare data visually	AWS Glue DataBrew
Validate data quality	AWS Glue Data Quality
Track data origin and transformations	Data lineage capabilities
Control data lake access at row/column/cell level	AWS Lake Formation
Archive data for long-term retention	Amazon S3 Glacier classes
Automatically transition or delete objects	Amazon S3 Lifecycle rules

24. High-Value Exam Traps

Artifact versus Audit Manager

Artifact gives you AWS compliance reports.
Audit Manager collects evidence about your AWS environment.

Config versus CloudTrail

Config: resource configuration and compliance state.
CloudTrail: API calls and account activity.

Config versus Inspector

Config: configuration policy violations.
Inspector: software, application, and infrastructure vulnerabilities.

Trusted Advisor versus compliance certification

Trusted Advisor gives best-practice recommendations. It does not certify that an organization complies with a regulation.

Guardrails versus IAM

Guardrails control harmful content, denied topics, and PII behavior.
IAM controls who can access AWS resources and perform actions.

Clarify versus Guardrails

Clarify: bias, feature attribution, and explainability.
Guardrails: generative AI input/output safety controls.

Data owner versus data steward

Owner: policy and accountability.
Steward: daily operational data management.

Data catalog versus data lineage

Catalog: what data exists and what it means.
Lineage: where data came from and how it changed.

Retention versus deletion

Compliance may require retaining data, but privacy policies may require deleting it. A documented retention policy must reconcile both requirements.

Compliance reports versus compliance

Using an AWS-certified service does not automatically make the customer’s application compliant. Customer configuration and processes remain important.

AI standards versus laws

ISO/IEC 42001, ISO/IEC 23894, and NIST AI RMF provide standards or guidance.
The EU AI Act is a regulation.
A voluntary framework is not necessarily a legal requirement.

25. Final Exam Summary

For Task Statement 5.2, remember these core ideas:

AWS and the customer share security and compliance responsibilities.
AWS Artifact provides AWS compliance reports and certifications.
AWS Audit Manager collects customer evidence and creates audit assessments.
AWS Config evaluates resource configurations and supports automated remediation.
CloudTrail records AWS API activity and audit events.
Amazon Inspector identifies vulnerabilities in applications, containers, and infrastructure.
Trusted Advisor recommends AWS best-practice improvements.
Bedrock Guardrails helps block harmful topics, unsafe content, and PII.
SageMaker Clarify supports bias detection, feature attribution, and explainability.
Data governance manages data availability, usability, integrity, security, privacy, and quality.
Lake Formation provides fine-grained data lake access control.
Glue Data Catalog stores metadata; Glue DataBrew profiles data; Glue Data Quality validates data.
S3 lifecycle policies support cost-aware retention, archival, and deletion.
Governance must address data residency, logging, monitoring, observation, retention, and access.
The Generative AI Security Scoping Matrix helps identify how responsibility increases as the customer builds more of the AI solution.
A strong AI governance program includes policies, role-based training, risk assessment, monitoring, regular reviews, transparency, human oversight, and continuous improvement.