Which existing processes does ISO 42001 plug into — and how do development, procurement, security and legal share the work?
ISO/IEC 42001 governs how an organisation manages AI, not a single model. Its natural home is therefore the processes you already run, not a fresh pile of documents. We usually start from four of them:
- Product and system development: requirements, design, testing and release already have gates. We add checkpoints for the AI system impact assessment (clauses 6.1.4 and 8.4) and for verification and validation (A.6.2.4), without creating a new review board.
- Procurement and supplier management: bought-in models, SaaS and generative-AI APIs go through the supplier assessment you already have; we only add the AI-specific questions of A.10.3 (suppliers) and the allocation of responsibilities in A.10.2.
- Security and IT operations: if you hold ISO 27001, document control, internal audit and management review are shared as they are. What is new is operation and monitoring of AI systems (A.6.2.6) and event logging (A.6.2.8).
- Legal and customer contracts: responsibilities go into the contracts; when a product heads for the EU, the EU AI Act articles are mapped onto the same system rather than a second one.
That is our default approach: we start from the management system you already run and build the AI requirements into it instead of setting up a separate one, cutting duplicated effort now and maintenance cost later.
Which technologies make ISO 42001 real — AI inventory, data governance, MLOps and human oversight?
One rule governs technical implementation: every control should leave a record that the system produces on its own, not a document written the week before the audit. The table lists the seven items we implement most often; clause and control numbers follow the body and Annex A of ISO/IEC 42001:2023.
| Requirement (clause / control) | Implementing technology | Evidence left behind |
|---|---|---|
| Clauses 6.1.2, 8.2 AI risk assessment | Model risk-tiering criteria (purpose, degree of automation, who is affected), built into the existing risk assessment tool | AI risk assessment records and approvals |
| Clauses 6.1.4, 8.4; A.5.2 AI system impact assessment process | Impact assessment template based on ISO/IEC 42005:2025, tied to development gates | Impact assessment reports and re-assessment triggers |
| A.4.2 resource documentation; A.6.2.3 design and development documentation | AI inventory and model registry covering in-house, bought-in, SaaS and generative AI | Inventory version history, purpose and owner |
| A.7.4 data quality; A.7.5 data provenance | Data versioning, lineage tracking, datasheets | Data quality check records, provenance records |
| A.6.2.4 verification and validation | Automated tests in the MLOps pipeline: performance thresholds, bias and robustness tests; no deployment below threshold | Test reports, model cards, deployment approvals |
| A.6.2.6 operation and monitoring; A.6.2.8 event logs | Model drift monitoring, centralised inference logs | Monitoring records, log retention policy and the logs themselves |
| A.9.2 processes for responsible use; A.9.4 intended use | Human-oversight design: confidence thresholds trigger human review; a generative-AI gateway controls input data and reviews output | Intended-use statements, human review records |
If a product is going to the EU, the same evidence is then mapped to EU AI Act Article 9 (risk management), Article 10 (data and data governance), Article 12 (record-keeping) and Article 14 (human oversight). Note that a 42001 certificate is not AI Act compliance in itself; statutory steps such as conformity assessment still have to be completed separately.
We do not prescribe particular tools. Existing version control, CI/CD and logging platforms can usually carry these controls; what is typically missing is not tooling but clearly written thresholds, approvers and retention periods. So we first inventory which records your current tools already produce, then add the rules, and only then assess whether a new tool is really needed.
How does an AI risk travel through ISO 42001 controls to technical controls and monitoring metrics?
Risk management is where the chain starts. Take a customer-service team rolling out generative AI to answer customers:
- Risk identification: using the risk sources in Annex C of 42001, three risks surface — customer personal data typed into an external model, wrong answers misleading customers, and a supplier model update changing behaviour.
- Standard requirements: clause 6.1.2 risk assessment and 6.1.4 impact assessment, plus A.9.2 responsible use, A.10.3 suppliers and A.8.4 communication of incidents.
- Technical controls: the gateway masks personal data automatically; low-confidence or high-risk answers go to a person; a supplier model version change triggers re-assessment; AI incidents feed into the existing security incident process.
- Monitoring metrics: personal-data masking hits, share of answers sent to human review, whether re-assessment was completed after each version change, incident handling time.
The chain can be walked end to end because one team plans regulation, process and technology together: determining risk level and provider or deployer role under the AI Act; building the AI management system, impact assessment and human-oversight procedures under ISO 42001; then handling data governance, model testing, and bias and robustness validation.
Three documents — the AI system inventory, the AI risk and impact assessment, and the EU AI Act obligations matrix — let ISO 42001 and the AI Act share one set of evidence. AI risk is finally rolled into the enterprise risk management risk map, so the board sees it instead of it stopping at the IT department.
Which departments, systems, data and evidence need to be integrated vertically into one AI management system?
AI governance most often stalls because every department holds a piece and nobody sees the whole. R&D has the models, procurement the contracts, security the logs, legal the regulatory mapping — each in a different system. Vertical integration means the data and evidence for one AI system run from the department all the way to management review, and can be traced.
| Department | System | Data | Evidence |
|---|---|---|---|
| R&D / data science | MLOps pipeline, model registry, version control | Training and test data, model versions, test results | Verification and validation records, model cards |
| IT / security | Logging platform, identity and access management (IAM), existing ISMS tools | Inference logs, access records, incidents | Log retention records, incident handling records |
| Procurement | Supplier management system | Third-party model and API assessments, contract terms | Supplier assessment records, responsibility clauses |
| Legal / compliance | Contract management, regulatory mapping | AI Act role determination, obligations matrix | 42001 × AI Act mapping |
| Sales / product | Product requirements management, customer-service system | Intended use, user feedback, complaints | Instructions for use, intended-use statements |
| Internal audit / management | Internal audit and management review records | Audit findings, AI objective attainment | Internal audit reports, management review records (clause 9) |
Once connected, the AI inventory becomes the index: each AI system in it links to its risk assessment, impact assessment, tests, monitoring and review records. Auditors no longer collect files department by department, and management review can see straight away which systems have overdue assessments.
Success hinges on three handover points: when procurement brings in a new AI service, a row is added to the inventory at the same time; when R&D deploys a new version, the system determines whether the impact assessment must be redone; when security detects an AI-related incident, affected interested parties are informed under A.8.4. If these three points leave records, the system will not break between departments.
What does Ahura do in an ISO 42001 engagement, and how does the data flow?
The Ahura enterprise risk management platform is a Taiwan-native RegTech platform for enterprise risk management and compliance, developed by Winners Consulting and currently at proof-of-concept (POC) stage. Everything below refers to the POC version. In an ISO 42001 engagement, data flows like this:
- In: consultant deliverables (AI policy, AI system inventory, risk and impact assessments) and records produced by your systems (test reports, monitoring records, human review records) are brought into Ahura.
- Kept: Ahura is 100% on-premises; client documents and data never go to the cloud or leave the client environment. For AI governance files that describe training data and model internals, this matters a great deal.
- Signed: a full-process evidence chain records every sign-off and is compatible with EU eIDAS certificates.
- Retrieved: internal audit, certification bodies or regulators retrieve evidence from one place instead of a fresh collection exercise.
A group-style distributed architecture lets headquarters set one AI policy while each site complies locally; the interface supports Traditional Chinese, English and EU languages such as German, French, Dutch and Czech.
To be clear: an ISO 42001-specific module is planned under the modular extension design; for now the focus is the evidence chain, sign-off and on-premises retention. Gap analysis from uploaded documents, cross-standard comparison (for example, what ISO 27701 documents still lack for 27001) and regulatory change tracking are in development and will open as the POC progresses.
Ahura itself is building its governance on the ISO/IEC 42001:2023 framework (in progress) and bars China-origin models and cloud LLM APIs.
What verifiable governance, risk, efficiency and compliance gains does ISO 42001 integration produce?
We avoid abstract talk of "AI governance maturity" and stick to changes an auditor can check:
- Governance: every AI system in the inventory has a purpose, an owner and a supplier record; the AI policy and roles are approved by management (clause 5, A.2, A.3). The board sees AI risk because it sits on the enterprise risk map.
- Risk: AI risk and impact assessments have a method, an approval and re-assessment triggers — a model version change, a change of use or a new supplier sets off a re-assessment. High-risk use leaves human review records.
- Efficiency: organisations with ISO 27001 share document control, internal audit and management review on one backbone; one set of evidence answers 42001, the AI Act mapping and customer due diligence. Organisations with ISO 27001 usually reach ISO 42001 certification in 4–6 months.
- Ongoing compliance: evidence comes out of the MLOps pipeline, logging platform and review workflow as a matter of course, not from pre-audit catch-up; internal audit and management review run continuously under clause 9.
Metrics you can accept the work against include: share of AI systems covered by the inventory, share of impact assessments completed on time, number of releases that went live below a test threshold, and completeness of human review records.
VP- and director-level consultants lead the work end to end, never outsourced; our consultants hold ISO Lead Auditor qualifications. Winners Consulting provides consulting only and does not issue certifications; certification is performed by a certification body accredited under ISO/IEC 42006:2025. To see where your own AI use cases should start, contact us.
Technical integration for related services
Official sources
- ISO/IEC 42001:2023 — Information technology — Artificial intelligence — Management system (IEC Webstore, 2023-12-18)
- ISO/IEC 42005:2025 — AI system impact assessment (IEC Webstore, 2025-05-28)
- ISO/IEC 42006:2025 — Requirements for bodies providing audit and certification of AI management systems (IEC Webstore, 2025-07-07)
- Regulation (EU) 2024/1689 — Artificial Intelligence Act (EUR-Lex)
- DIN Media — ISO/IEC 42001:2023 table of contents (clauses 4–10, Annexes A–D)
