Scope and your role
Which AI systems are covered, and whether you act as developer, provider, user, or all three.
AI management systems
Govern AI before it governs your risk.
An AI management system that conforms to ISO/IEC 42001, holds up against the NIST AI RMF and the EU AI Act, and is tested as an adversary would.
- ISO/IEC 42001:2023
- NIST AI RMF
- EU AI Act
- Who it is for
- Organizations that build, buy or deploy AI
- The standard
- ISO/IEC 42001:2023, the certifiable AI management-system standard
- You receive
- A working AI management system and an audit-ready evidence pack
- The outcome
- Certification readiness, proven by internal audit and red teaming
The executive challenge
AI arrived faster than the controls around it.
Vendor models in underwriting, language models in service, assistants in engineering, and pilots nobody approved.
Few have an owner, a documented purpose or someone with authority to stop them. ISO/IEC 42001 gives every new use case a process that already knows how to evaluate it.
- Clauses
- 4–10
- Annex A
- 9
- Assessments
- 2
The management-system core, shared with ISO/IEC 27001
AI control objectives, A.2 to A.10
AI risk and AI system impact, distinct obligations
A policy that says “use AI responsibly” is not a control. A control is a named person who can stop a model.
Client outcomes
What changes when AI is governed
01
An owned AI inventory
Every system, whether built, bought or embedded, with an owner, a purpose and a risk tier.
02
Decisions you can defend
Risk and impact assessments that show why each use was approved.
03
Oversight that works
Human review designed into the workflow, with real authority to intervene.
04
Certification readiness
An evidence pack prepared for an accredited body’s Stage 1 and Stage 2 audits.
The standard, explained
What ISO/IEC 42001 actually asks
Clauses 4–10 define the management system; Annex A sets reference controls for AI. These are its distinctive demands.
AI risk assessment
A repeatable method for risks to the organization, individuals and society.
AI system impact assessment
Documented consequences for the people a system affects, revisited on change.
The AI life cycle
Design, verification, deployment and monitoring controls that were actually applied.
Data for AI
Provenance and quality of training and evaluation data, the most common weak point.
Third parties
Responsibilities allocated with the suppliers of the models you depend on.
Summarized from ISO/IEC 42001:2023; the published standard governs.
Interactive demonstration
An AI governance register, working.
Change a system’s risk answers and watch its tier and controls update. Open the red-team findings to see a hidden instruction reach a refund tool.
Fictional organization, synthetic data. Real engagements test real models under agreed rules of engagement.
Loading interactive demonstration…
Every asset, finding and score in this panel is synthetic and generated in your browser. Nothing is scanned, collected or sent.
Methodology
An AI management system that runs
- Discover
Find the AI
Built, bought, embedded and shadow use.
- Assess
Risk & impact
Assessments with use-case owners; gap against clauses 4–10.
- Remediate
Build
Policy, roles, life-cycle and supplier controls.
- Validate
Test & audit
Red teaming, internal audit and readiness check.
- Sustain
Keep current
New use cases, model changes and regulation.
Deliverables
What you receive
AI system inventory
Owner, purpose, data, supplier and risk tier for every system.
Statement of Applicability
Annex A controls selected and justified from risk treatment.
AI red-team report
Findings mapped to OWASP LLM Top 10 and MITRE ATLAS.
Internal audit & readiness
An independent view of readiness for certification.
Accreditation & scope
From inventory to certificate
What MarkTillson delivers
- Gap assessment, design and implementation support
- AI risk and impact assessments
- Internal audit and readiness assessment
- AI red teaming and LLM application testing
What completes certification
- ISO/IEC 42001 certification, awarded by an accredited certification body
- Legal opinion on EU AI Act classification, from your counsel
Framework mapping
One system, several obligations
42001 is the structure. The NIST AI RMF and EU AI Act shape what goes inside it.
| Control area | ISO/IEC 42001 | NIST AI RMF | EU AI Act | ISO/IEC 27001 |
|---|---|---|---|---|
| AI policy and accountability | Directly addressed | Directly addressed | Partially addressed | Partially addressed |
| Risk assessment and treatment | Directly addressed | Directly addressed | Directly addressed | Partially addressed |
| Impact assessment on people | Directly addressed | Directly addressed | Partially addressed | Not addressed– |
| Data governance and quality | Directly addressed | Directly addressed | Directly addressed | Partially addressed |
| Human oversight | Directly addressed | Directly addressed | Directly addressed | Not addressed– |
| Robustness and cybersecurity | Partially addressed | Directly addressed | Directly addressed | Directly addressed |
| Third-party management | Directly addressed | Directly addressed | Partially addressed | Directly addressed |
Directly addressed Partially addressed or different evidence expectations , Not a primary objective. EU AI Act obligations depend on system classification and your role. The Digital Omnibus on AI moved most stand-alone high-risk obligations to December 2027.
Industry considerations
Same standard, different stakes
Extend model risk management to language models, and classify credit and insurance-pricing uses carefully under the EU AI Act.
Over-reliance on generated summaries is a patient-safety risk. Impact assessments need a clinical-safety frame.
42001 increasingly answers enterprise procurement. The challenge: release gates that keep pace with weekly model changes.
Transparency, contestability and fundamental-rights impact assessment carry particular weight.
Risks we test
Where AI systems actually fail
Prompt injection
Hidden instructions in documents hijack a model that cannot tell data from commands.
Data disclosure
Retrieval that ignores entitlements; secrets in prompts.
Excessive agency
Agents with broad tools and no approval gate.
Drift
Performance shifts with no threshold, owner or trigger.
Upstream change
A supplier updates a hosted model overnight.
Vector stores
Poisoned embeddings and weak tenant isolation.
Engagement journey
Inventory to readiness
-
01
Orientation
Scope and the AI roles you play.
OutputScope statement
-
02
Inventory & gap
Discovery and clause-by-clause gap.
OutputInventory and gap report
-
03
Build
Methods, SoA and controls.
OutputManagement-system documentation
-
04
Test
Red teaming of priority systems.
OutputRed-team report
-
05
Audit
Internal audit and management review.
OutputReadiness report
FAQ
Questions we are asked
Who certifies us to ISO/IEC 42001?
An accredited certification body. We build, test and internally audit your management system, help you choose the body and support you through Stage 1 and Stage 2.
We hold ISO/IEC 27001. How much can we reuse?
Most of the management-system machinery: document control, internal audit, management review and corrective action. The AI-specific risk, impact, life-cycle and oversight content is new.
Does 42001 make us EU AI Act compliant?
Not by itself. It is an excellent vehicle for many AI Act obligations, but those depend on how a system is classified and your role. We map the gaps explicitly.
We mostly buy AI. Is the standard relevant?
Yes. For buyers the emphasis shifts to inventory, intended use, supplier assessment and monitoring.
Is AI red teaming safe for production?
Rules of engagement and stop conditions are agreed in advance. Most testing runs in staging; any production testing is tightly scoped and approved.
Related
Continue exploring
Speak with an AI governance specialist
Find out what AI you are really running.
A confidential conversation about your AI portfolio, your obligations and where the inventory should begin.