What Is an AI Audit?
An AI audit is a systematic examination of an AI system’s development, deployment, and operation. Unlike traditional IT audits that focus primarily on security controls, AI audits evaluate a broader set of concerns:
- Technical soundness: Model performance, reliability, robustness
- Ethical alignment: Fairness, bias, transparency, accountability
- Regulatory compliance: Adherence to applicable laws and standards
- Governance effectiveness: Policies, procedures, oversight mechanisms
- Risk management: Identification, assessment, and mitigation of AI-specific risks
The goal is to provide independent assurance that AI systems operate as intended, comply with requirements, and don’t create unacceptable risks for the organization or affected individuals.
The Audit Gap
Policies describe intended governance. Operational records can add scoped evidence of which configured control path reported which outcome, while auditors still need separate evidence for coverage and effectiveness.
Types of AI Audits
AI audits come in several forms, each with different scopes, objectives, and evidence requirements:
| Audit Type | Focus | Who Conducts |
|---|---|---|
| Internal Audit | Policy compliance, control effectiveness, risk identification | Internal audit team |
| Technical Audit | Model performance, security, data handling, MLOps practices | Specialized AI auditors |
| Bias/Fairness Audit | Algorithmic bias, disparate impact, protected class analysis | Specialized firms, academic partners |
| Regulatory Audit | Compliance with specific regulations (EU AI Act, state laws) | External auditors, regulators |
| Certification Audit | ISO 42001, SOC 2 + AI controls, industry certifications | Accredited certification bodies |
| Due Diligence Audit | M&A, vendor selection, investment decisions | Consulting firms, specialized assessors |
Healthcare-Specific Considerations
Healthcare AI review criteria depend on the intended use, regulated role, data flow, and patient-risk boundary. Relevant questions may include:
- HIPAA scope: PHI handling, access controls, and audit controls where HIPAA applies to the organization and workflow
- FDA scope: Applicable device requirements when the product includes a regulated device software function
- Clinical validation: Evidence matched to intended use, user, setting, and patient-risk pathway
- Integration: EHR connectivity, data-flow boundaries, dependencies, and failure handling
Regulatory Landscape
Rules for AI are arriving from several directions at once, and the obligations overlap enough that work done for one counts toward another. Four of them shape most audit scopes today:
Colorado ADMT Law (SB 26-189)
SB 26-189 repealed and replaced the original Colorado AI Act before it took effect. It now sets pre-use notice, post-adverse-outcome disclosure, and documentation duties for covered automated decision-making technology (ADMT), with substantive compliance from January 1, 2027.
EU AI Act (High-Risk)
Specified healthcare uses can be high-risk, requiring conformity assessment, technical documentation, and ongoing monitoring. The AI Omnibus entered into force on July 27, 2026; relevant Annex III high-risk obligations apply from December 2, 2027. Annex I product-embedded systems follow on August 2, 2028.
California ADMT
Under the CPPA’s approved CCPA regulations, businesses using ADMT for significant decisions must comply with the applicable ADMT requirements beginning January 1, 2027. Scope and available consumer rights depend on the regulation’s definitions and exceptions.
ISO 42001
Voluntary international management-system standard for establishing, implementing, maintaining, and improving an AI management system. Certification and audit scope are set through the applicable conformity-assessment process.
These regulations share common themes: risk assessment, documentation, human oversight, and ongoing monitoring. A governance program built once around those four can answer several of the requirements at the same time, which is the argument for treating them as one program rather than four projects.
Evidence Requirements
Auditors may examine several complementary evidence forms. The GLACIS evidence layers below describe different properties; no layer alone establishes audit sufficiency or control effectiveness.
GLACIS Evidence Layers
Four complementary forms of AI governance evidence
Policy Documentation
Written policies, procedures, and guidelines. Documents intended practice but not operational events. Intent evidence.
Operational Records
Logs, dashboards, and reports describing activity. They may be incomplete or retroactively modified. Process evidence.
Execution Traces
Timestamped, per-inference records linking a specific output to the configured control states reported for that event. Scoped operational evidence.
Cryptographic Attestation
Signed records whose covered integrity and provenance properties can be checked independently. This verification does not establish control effectiveness, complete capture, safety, or compliance. Independently checkable evidence.
Evidence by Audit Domain
The table below sets out the artifacts an assessor usually asks for first in each domain. Gathering them in this shape saves a round of follow-up requests later.
| Domain | Required Evidence |
|---|---|
| Governance | AI policy, roles & responsibilities, board oversight records, ethics review minutes, training records |
| Risk Management | Risk assessments, impact analyses, risk registers, mitigation plans, residual risk acceptance |
| Development | Requirements specs, data lineage, training documentation, validation reports, model cards |
| Testing | Test plans, test results, bias testing, red team reports, performance benchmarks |
| Operations | Deployment records, monitoring dashboards, incident logs, change management, access logs |
| Continuous Monitoring | Drift detection, performance metrics, feedback loops, retraining triggers, audit trails |
Common Audit Findings
Five findings recur across healthcare AI audit and assurance guidance. Read them as a pre-audit checklist, since each one is cheaper to close before an assessor names it than after:
Missing Execution Evidence
Organizations may have policies and dashboards without event-level records connecting specific outputs to the control states reported at the time. Those records add operational context but do not, by themselves, establish that a control was effective.
Incomplete Model Documentation
Training data sources, preprocessing steps, and model architecture decisions are poorly documented or not linked to deployed versions.
Inadequate Bias Testing
Testing exists but doesn’t cover all protected classes, use cases, or edge conditions. Results aren’t tied to deployment decisions.
Weak Change Control
Model updates and configuration changes lack formal approval processes. Auditors can’t verify which version was running at a specific time.
Human Oversight Gaps
Policies require human review but no evidence exists that reviews actually occurred, or reviewers lack qualification documentation.
Preparation Sequence
Timing depends on scope and evidence readiness. Use this sequence after agreeing milestones with the relevant reviewer.
Define criteria, boundaries, and owners
- Define audit scope with auditor
- Identify all AI systems in scope
- Assign audit coordinators and evidence owners
- Conduct gap assessment against requirements
Inventory artifacts and identify gaps
- Gather existing documentation
- Identify and remediate gaps
- Implement missing controls
- Create evidence inventory
Run scope-appropriate testing and prepare evidence
- Complete technical documentation
- Run bias and performance tests
- Validate control effectiveness
- Prepare evidence packages
Check completeness before assessor access
- Internal readiness review
- Staff briefings and interview prep
- Organize evidence room/portal
- Address any last-minute gaps
Respond, trace, and preserve decisions
- Provide requested evidence promptly
- Facilitate interviews and walkthroughs
- Document auditor questions and requests
- Track and respond to preliminary findings
Remediation and follow-up
- Review draft findings
- Develop remediation plans
- Implement corrective actions
- Schedule follow-up verification
Building Audit-Ready Systems
Audit preparation is cheapest when it starts at design time. A system built to emit its own evidence leaves nothing to reconstruct under time pressure, and the reconstruction is where most of the cost and most of the argument sits.
Design Principles
- Evidence by design: Define which events the system must record automatically, why they are in scope, and which actions remain outside that coverage
- Tamper evidence: Under the stated signing and storage design, later modification of covered fields should be detectable. Tamper-evident records are not necessarily immutable
- Traceability: For events in scope, preserve the purpose-appropriate input references, model or configuration identifiers, and recorded control states
- Verifiability: Third parties should be able to validate evidence without full system access
Key Capabilities
Guardrail Execution Logs
For configured events, record which controls were invoked, their reported outcomes and order, timestamps, and the boundaries of coverage.
Model Version Tracking
Where the integration can support it, bind declared model, configuration, or prompt-template identifiers to the covered event. A recorded identifier does not prove which artifact actually executed without separate deployment evidence.
Selected Decision Context
For covered events, preserve purpose-appropriate context references and declared configuration. This does not reproduce model reasoning or guarantee identical replay.
Compliance Dashboards
Real-time views mapping operational evidence to specific regulatory requirements and control objectives.
The Payoff
Organized, review-ready records can reduce evidence-collection friction. Audit findings and compliance conclusions still depend on the applicable scope, criteria, testing, and reviewer.
Frequently Asked Questions
What is an AI audit?
An AI audit is a systematic examination of an AI system’s development, deployment, and operation to assess compliance with regulations, adherence to ethical principles, effectiveness of controls, and alignment with organizational policies.
Who conducts AI audits?
AI audits can be conducted by internal audit teams, external audit firms (Big 4 and specialized AI auditors), regulatory bodies, standards organizations (for certifications like ISO 42001), and third-party assessment organizations.
How long does an AI audit take?
There is no universal duration. Timing depends on the system boundary, criteria, evidence readiness, testing depth, reviewer process, and remediation needs. Agree the schedule with the relevant reviewer rather than planning around a generic estimate.
What happens if we fail an AI audit?
Consequences depend on the audit type. Internal audits lead to remediation plans. Certification audits may result in certification denial or conditions. Regulatory audits could trigger enforcement actions, fines, or operational restrictions.
How often should AI systems be audited?
Set cadence from applicable law, standards, contracts, risk, and the assurance objective. Material changes, incidents, expanded use, or changed dependencies may justify reassessment between scheduled reviews; continuous monitoring does not replace a required independent audit.