Operational Evidence for ISO/IEC 42001 Certification Audits
Who this is for: ISO/IEC 42001 certification auditors conducting Stage 1 and Stage 2 assessments. AIMS lead implementers preparing for certification. AI governance coordinators maintaining certified management systems between surveillance audits.
ISO/IEC 42001 is the first international standard for AI Management Systems (AIMS). It defines what an organization must manage: AI policy, risk assessment, model lifecycle, monitoring, fairness, transparency, and human oversight. TUV Rheinland achieved accreditation for ISO 42001 certification in January 2026 and has engaged clients across banking, food manufacturing, and medical devices.
Certification audits verify the management system exists. They verify documented procedures, policy statements, and organizational structures. What they cannot verify is whether the system operates continuously between audits. A management system on paper is not a management system in practice.
A management system without operational evidence is a policy binder on a shelf. The binder proves the organization planned. It does not prove the organization executed.
The gap between "documented" and "operational" is where nonconformities hide. An AI risk assessment documented in January may not reflect the risks present in July. A fairness threshold defined in a policy may be violated in production without anyone knowing. Model lifecycle procedures exist, but the model currently in production may not have followed them.
SWT3 closes this gap. Each witness anchor is an objective evidence event: timestamped, fingerprinted, immutable. When the auditor asks "show me evidence of continuous monitoring," the answer is not a screenshot of a dashboard. It is a ledger of cryptographic anchors spanning every day since the last audit.
Note: SWT3 does not replace the management system. It provides the operational evidence that proves the management system is working. Policy defines intent. Anchors prove execution.
The following table maps ISO/IEC 42001 clauses and Annex A controls to SWT3 procedures. Each row identifies the clause requirement, the procedure that produces objective evidence for that requirement, and the specific evidence produced by each witness anchor.
| ISO 42001 Clause | Requirement | SWT3 Procedure | Objective Evidence Produced |
|---|---|---|---|
| 5.2 AI Policy | Documented AI policy, communicated and maintained | AI-GOV.1 |
Governance framework attestation anchor with controls-defined/controls-active ratio |
| 6.1.2 AI Risk Assessment | AI-specific risk identification and evaluation | AI-DRIFT.1, AI-DRIFT.2 |
Drift measurements as operational risk metrics with consequence mapping |
| A.6.2.4 Model Lifecycle Management | Model integrity through development, deployment, retirement | AI-MDL.5, AI-MDL.6, AI-MDL.7 |
Weight hash verification, adapter attestation, quantization records |
| A.6.2.5 Continuous Monitoring | Ongoing measurement of AI system performance | AI-DRIFT.1 |
Drift metrics per evaluation window with detection method and threshold |
| A.6.2.6 AI Security | Adversarial resilience and threat mitigation | AI-SEC.1 |
Threat detection results with category counts and blocked status |
| A.8.3 Transparency | Stakeholder communication about AI capabilities and limitations | AI-TRANS.1, AI-EXPL.1 |
Disclosure events with type classification, explanation generation records |
| A.8.4 Fairness | Bias identification and mitigation processes | AI-FAIR.1 |
Disparity measurements across protected characteristics |
| A.8.5 Human Oversight | Mechanisms for human intervention and review | AI-HITL.1 |
Human review completion records with reviewer binding |
| 9.1 Performance Evaluation | Monitoring, measurement, analysis, and evaluation | AI-AUDIT.1 |
Audit log integrity verification with entry counts and format |
| 9.2 Internal Audit | Planned audit program with results documented | All procedures | Full ledger query scoped to audit period and system boundary |
Clauses 5 through 7 (Context, Leadership, Planning, Support) map to governance attestation through AI-GOV.1. These are the "does the system exist" clauses. They verify organizational commitment, resource allocation, and documented policy. The governance attestation anchor records how many controls are defined and how many are active, producing a ratio that quantifies management system coverage.
Clause 8 (Operation) maps to runtime procedures: AI-MDL.5/6/7 for model lifecycle, AI-DRIFT.1/2 for performance monitoring, AI-SEC.1 for security, AI-FAIR.1 for fairness, AI-TRANS.1 for transparency, and AI-HITL.1 for human oversight. These are the "is the system working" clauses. Each anchor is an objective evidence event proving that the operational procedure was executed on a specific date.
Clauses 9 and 10 (Performance Evaluation, Improvement) map to audit integrity through AI-AUDIT.1 and the full ledger query. These are the "is the system improving" clauses. Audit log integrity verification confirms evidence has not been tampered with. The ledger query scoped to the internal audit period confirms that all applicable procedures were covered.
The mapping is not one-to-one. A single clause may require evidence from multiple procedures, and a single procedure may satisfy elements of multiple clauses. Use this table as a starting point, not a substitute for professional judgment about the scope of the AIMS.
ISO 42001 certification follows the standard two-stage audit process. Stage 1 confirms the management system is designed and documented. Stage 2 confirms the management system is implemented and operating effectively. The evidence requirements are fundamentally different.
| Audit Stage | Focus | Key SWT3 Evidence | What It Proves |
|---|---|---|---|
| Stage 1 (Documentation) | Does the AIMS exist on paper? | AI-GOV.1 attestation, documented clearing levels, procedure definitions |
The management system is designed and documented |
| Stage 2 (Implementation) | Is the AIMS operating effectively? | AI-INF.1 volume, AI-DRIFT.1 history, AI-FAIR.1 measurements, AI-MDL.5 deployment records |
The management system is implemented and producing results |
| Surveillance | Is the AIMS maintained? | Anchor continuity (no gaps), drift trend stability, governance re-attestation | The management system continues to operate between certifications |
At Stage 2, request 90 days of AI-DRIFT.1 anchors. Continuous monitoring means continuous evidence, not a single report generated for the audit. If the organization produces anchors only during audit preparation, the gap between the last anchor and the audit date reveals the true monitoring cadence.
The most common Stage 2 nonconformity for AI management systems is "monitoring exists but evidence does not." Organizations install drift detection tools, configure fairness thresholds, and implement review workflows. They forget to record the results as objective evidence. The tool ran. The evidence was not captured.
The following checklist covers the ten most critical evidence items for an ISO 42001 certification audit. Each item identifies the clause reference, what to request, what to verify, and the most common nonconformity encountered at that checkpoint.
Request: AI-GOV.1 anchors for the assessment window.
Verify: Controls active is greater than or equal to controls defined. Framework version and last review date are populated.
Common nonconformity: Policy exists but was never attested, or attestation is more than 12 months old. The policy document was approved, but the management system never recorded that it was active.
Request: AI-DRIFT.1 and AI-DRIFT.2 anchors for the assessment window.
Verify: Drift monitoring is active and consequence categories are mapped to business impact levels.
Common nonconformity: Risk assessment is static. It was documented once and never updated with operational data. The risk register describes theoretical risks, not measured ones.
Request: AI-MDL.5, AI-MDL.6, and AI-MDL.7 anchors spanning the deployment window.
Verify: Weight hash matches across the deployment window. Adapter and quantization records exist for each model version.
Common nonconformity: Model was updated but the lifecycle procedure was not followed. No anchor exists at the time of deployment. The model changed; the evidence trail did not.
Request: AI-DRIFT.1 anchor timeline for the full assessment period.
Verify: Continuous cadence with no gaps exceeding 7 days. Metrics evaluated and detection method are consistent.
Common nonconformity: Monitoring tool is installed but only runs manually before audits. The anchor timeline shows clusters around audit preparation dates and silence between them.
Request: AI-SEC.1 anchors for the assessment window.
Verify: Threat categories checked is greater than zero. For PASS verdicts, threats detected equals zero.
Common nonconformity: Security scanning is not applied to AI-specific threat vectors. The organization performs traditional IT security scans but does not assess adversarial attacks, prompt injection, or model extraction.
Request: AI-TRANS.1 anchors for the assessment window.
Verify: Disclosure events match the system's declared communication plan. Disclosure types cover all required categories.
Common nonconformity: Privacy notice mentions AI use, but no disclosure events are recorded at the point of interaction. The notice exists. The delivery was never evidenced.
Request: AI-FAIR.1 anchors for the assessment window.
Verify: Protected attribute coverage matches the organization's declared scope. Disparity measurements are within declared thresholds.
Common nonconformity: Fairness tested at training time only, with no production monitoring. The model was fair on test data. Whether it remains fair on real-world populations is unknown.
Request: AI-HITL.1 anchors for the assessment window.
Verify: Human review occurred for high-risk decisions. Reviewer binding confirms a named individual completed the review.
Common nonconformity: Human oversight is documented in policy but not implemented in the system. The procedure describes who should review. No evidence exists that anyone did.
Request: AI-AUDIT.1 anchors for the assessment window.
Verify: Audit log integrity is confirmed. Entry count and format are recorded.
Common nonconformity: Logs exist but integrity is not verified. The logs could have been modified without detection. Existence of a log file is not evidence of log integrity.
Request: Full ledger query for the internal audit period, scoped to the declared system boundary.
Verify: Anchor coverage spans all applicable procedures. No procedures defined in the AIMS scope statement are missing from the evidence record.
Common nonconformity: Internal audit did not include AI-specific controls. The audit program covered information security (ISO 27001) but did not extend to AI management system requirements.
Between certifications, surveillance audits verify the AIMS continues to operate. The certification body typically conducts surveillance audits at 12-month intervals. SWT3 evidence serves this purpose without additional preparation.
For surveillance audits, compare the anchor timeline against the certification date. Evidence should begin immediately after certification and continue without interruption. Gaps indicate periods where the management system was not producing objective evidence. A 90-day gap in a 12-month surveillance period is not a minor observation. It is a potential major nonconformity against Clause 9.1.
The surveillance audit should also verify that corrective actions from the previous certification or surveillance audit have been implemented. If a nonconformity was raised at Stage 2 -- for example, missing fairness monitoring -- the corrective action evidence is a series of AI-FAIR.1 anchors beginning shortly after the nonconformity was recorded and continuing without interruption.
Install:
# Python
pip install swt3-ai
# TypeScript
npm install @tenova/swt3-aiRecord a governance attestation (implementation reference):
from swt3_ai import Witness
witness = Witness(api_key="your_key", tenant_id="your_tenant")
witness.witness_governance_framework(
controls_defined=45,
controls_active=43,
framework_version="ISO-42001-v1",
last_review_date="2026-09-01",
)This produces an AI-GOV.1 anchor that serves as objective evidence for Clause 5.2 (AI Policy) and contributes to the audit trail for Clauses 9.1 and 9.2.
Full SDK documentation: SDK Docs. Create a free account: Sign Up.