Should an AI audit scope include the training data?

Yes, when data quality, bias, or data governance are relevant risks. Training data directly influences system behavior, and many regulations require documentation of data provenance and quality measures.

How do I scope an audit for an AI system using a third-party API?

Include the organization's management of the third-party relationship (contracts, risk assessments, monitoring) within scope. The third-party system itself may be excluded, but the organization's due diligence should be evaluated.

Can the audit scope change after the audit begins?

Yes, but scope changes should be formally documented, justified, and approved. Uncontrolled scope changes undermine audit integrity and can lead to missed risks or wasted resources.

What is the difference between audit scope and audit criteria?

Scope defines what is being audited (the system, processes, and boundaries). Criteria define what standards the system is being evaluated against (regulations, standards, policies).

How does scope differ for a pre-deployment vs post-deployment audit?

Pre-deployment audits focus on design, development, and validation phases. Post-deployment audits focus on operational performance, monitoring, incident handling, and the effectiveness of controls in production.

Quick answer

Audit scope for AI systems should clearly identify the specific system version, lifecycle phases, data flows, third-party components, applicable regulations, and organizational boundaries, with documented justifications for any exclusions.

Updated June 2026 · MmowW AI Compliance

Defining Audit Scope for AI Systems: Boundaries, Inclusions, and Exclusions (2026)

Why Scope Matters

Scope definition is the single most important planning decision in an AI audit. Too broad, and the audit becomes superficial. Too narrow, and material risks may be missed. The scope must be precise enough to produce actionable findings while comprehensive enough to satisfy regulatory expectations.

Dimensions of AI Audit Scope

System Identification

Identify the AI system by name, version, and deployment context. A single AI model may be deployed in multiple contexts, each with different risk profiles. Specify which deployment(s) the audit covers.

Lifecycle Phases

AI systems have a lifecycle that includes design, data collection, training, validation, deployment, monitoring, and retirement. The audit scope should specify which phases are included.

Phase	Include When	Typical Audit Activities
Design	Pre-deployment audit	Requirements review, risk assessment evaluation
Data collection	Data quality is in scope	Consent verification, bias in data sources
Training	Model development is in scope	Training process documentation, validation approach
Deployment	Release process is in scope	Deployment procedures, rollback capability
Operation	Production systems	Performance monitoring, incident handling
Monitoring	Ongoing compliance	Drift detection, performance metrics

Regulatory Mapping

Map applicable regulations to the AI system. For EU-deployed systems, determine the risk classification under the EU AI Act and identify specific articles that apply. Layer sector-specific regulations (financial, medical, employment) as additional criteria.

Organizational Boundaries

Define which organizational units are included. AI systems often span multiple teams (development, operations, business units). Clarify which teams and processes fall within the audit boundary.

Third-Party Components

Modern AI systems frequently incorporate third-party components: pre-trained models, cloud services, data sources, and APIs. Define the extent to which third-party components are evaluated. At minimum, review how the organization manages third-party risks even if the third party itself is outside the audit boundary.

Scoping for EU AI Act Compliance

When auditing for EU AI Act compliance, the scope should align with the conformity assessment requirements for the system's risk classification.

Unacceptable risk: Not auditable (these systems are prohibited)
High risk: Full conformity assessment scope (Articles 8-15)
Limited risk: Transparency obligations (Article 50)
Minimal risk: Voluntary codes of conduct

Common Scoping Mistakes

Including too many AI systems in a single audit engagement
Excluding data quality when it is a primary risk factor
Omitting third-party model components from consideration
Failing to specify the system version (models change frequently)
Not aligning scope with regulatory requirements
Excluding post-deployment monitoring when the system is in production

Documenting the Scope

The scope statement should be a formal document, reviewed and approved by audit management and communicated to all stakeholders. It should include the following elements.

AI system name, version, and deployment context
Lifecycle phases included
Applicable regulations and standards
Organizational units and locations
Third-party components and their treatment
Time period covered
Exclusions with justification
Limitations and assumptions

Adjusting Scope During the Audit

Sometimes findings during fieldwork reveal that the scope needs adjustment. If the audit team discovers that a critical data source or third-party component was not included, the scope should be formally amended with documentation of the change and its rationale. Scope changes should be approved by audit management and communicated to stakeholders.

Check your AI compliance readiness — free.

Take the Readiness Check 3 minutes · 10 questions · no signup required

This article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.