Who this guide is for
Compliance officers and MLROs at regulated firms who are being asked how the firm governs the AI inside its transaction monitoring, onboarding, pricing or client-facing tools. Founders preparing a licence application whose business model depends on models or automation. Boards and senior managers who carry ultimate responsibility for systems and controls and want to oversee AI the way they oversee credit or conduct risk. Outsourced compliance providers and technology leads evidencing AI governance for several entities under several regulators. The examples draw on the UK, the EU and the Gulf; the logic applies wherever a supervisor can ask "how do you govern this model?".
What is ISO 42001?
ISO/IEC 42001:2023 is the first international standard for AI management systems. It sets out the requirements for establishing, running, maintaining and improving an AIMS: the organisational framework through which a firm decides which AI it will use, assesses the risks and impacts of each system, controls development, procurement and operation, monitors performance in production, and reviews the whole arrangement at board level. It is certifiable, which means an accredited certification body can audit the AIMS and issue a certificate that a client, investor or regulator can rely on.
ISO/IEC 42001:2023 is the first international standard for AI management systems.
The standard follows the harmonised high-level structure used by every modern ISO management system standard: clauses 4 to 10 cover context, leadership, planning, support, operation, performance evaluation and improvement. If you have worked with ISO 27001 you will recognise the architecture immediately. What ISO 42001 adds is AI-specific content inside that structure: a risk assessment method tuned to algorithmic systems, an AI impact assessment on individuals and groups, controls for bias, transparency, human oversight and model change, and 38 reference controls in Annex A grouped under nine objectives that run from AI policy and internal organisation through impact assessment, lifecycle, data and third-party relationships.
How does an AIMS differ from ISO 27001 and data protection law?
The confusion usually starts with a firm that holds ISO 27001 and assumes its AI is therefore governed. It is not. ISO 27001 is defensive: it protects the confidentiality, integrity and availability of information and of the infrastructure the models run on, but says nothing about what those models decide. Data protection law, whether GDPR, UK GDPR or a local regime such as the ADGM Data Protection Regulations 2021, governs personal data: lawful basis, rights, automated decision-making, impact assessments. It protects the individual, not the wider risk of a model being wrong.
ISO 42001 covers the gap: a credit model that discriminates, a monitoring model whose alerts cannot be explained to the MLRO, a pricing model that drifts as the book changes, an automated process nobody can meaningfully override. The three regimes are complementary. An AI system taking automated decisions about customers using personal data needs all three: ISO 27001 securing the platform, data protection law governing the data, ISO 42001 governing the behaviour of the model.
Does it apply to us?
No rulebook we know of mandates ISO 42001, and most financial regulators have not yet written AI-specific rules. What they have written is a set of hooks that pull AI into the regulatory frame: a risk management framework proportionate to the business, outsourcing rules that apply to AI vendors like any other, IT and cyber risk guidance (the FSRA's IT Risk Management Guidance has a chapter on algorithm-driven solutions; the FCA, PRA and ICO have all raised their expectations on AI governance), and data protection law with its automated decision-making and impact assessment provisions. A supervisor who asks "how do you govern this model?" is asking under those rules, and ISO 42001 is the most widely recognised structured answer.
No rulebook we know of mandates ISO 42001, and most financial regulators have not yet written AI-specific rules.
Use the table below as a first screen.
| If this is true of your firm | Then |
|---|---|
| You build models or AI features into a product you sell or operate for others | The AIMS is core to your product assurance. Expect clients to ask for it in due diligence. |
| You use third-party AI in decisions about customers: onboarding, customer risk rating, alert triage, credit, pricing, suitability | You are accountable for those decisions regardless of who built the model. ISO 42001's third-party and oversight controls are the relevant part. |
| Any AI output is used in the EU, or you sell AI into the EU | The EU AI Act may apply. Classify each system now; the first hard deadline for the financial services use cases is 2 December 2027. |
| You are applying for a licence or permission | Expect systems and controls questions about any model in the business plan. A scoped AIMS is a clear answer. |
| Your AI is limited to drafting, summarising and internal productivity | Scope a light AIMS. The inventory and the acceptable-use rules still matter; the heavy controls can wait. |
What does ISO 42001 require?
What does Clause 4 ask you to understand about your context?
Clause 4 requires you to document the external and internal factors that shape your AI governance before you build anything. Externally that means the regulatory position above, the EU AI Act where it reaches you, data protection law, client expectations and the technology you depend on. Internally it means your AI strategy, culture, resources, existing management systems and reliance on third-party models and data.
Clause 4 requires you to document the external and internal factors that shape your AI governance before you build anything.
You must also list your interested parties and their requirements: customers, your regulator, your data protection authority, employees, investors, auditors, vendors and, for some products, the public. Finally you define the AIMS scope: which AI systems, business units, jurisdictions and lifecycle stages are inside. Scope should follow risk. Most firms start with the systems that touch customers or regulatory reporting and widen the scope as the AIMS matures.
What does leadership have to do under Clause 5?
Top management must own the AIMS visibly, not sign a policy and step back. The clause requires an AI policy stating acceptable and prohibited uses, human oversight requirements and escalation routes for high-risk decisions; a governance structure with named roles; and adequate budget, people and training. In a regulated firm the natural home is an AI governance committee reporting to the board or risk committee, drawing in the chief executive, the compliance officer, the MLRO where financial crime models are in scope, the data protection officer, technology, risk and the business owner of each system.
This is the clause where implementations are won or lost. A committee that meets once and never reconvenes produces a management system that exists only on paper, and a certification auditor will find that out in the first interview.
How does Clause 6 handle AI risk and impact assessment?
Clause 6 is the core of the standard and the place where most firms have the biggest gap. It requires a structured AI risk assessment method covering impact on safety, fundamental rights and economic harm; likelihood factors such as model complexity, data sensitivity, degree of automation and scale of deployment; the controls already in place and how effective they are; and the residual risk once controls are applied.
Alongside risk assessment sits AI impact assessment: the effect of each system on individuals and groups, including discrimination, transparency, user autonomy and the adequacy of safeguards. This aligns with the fundamental rights impact assessment that the EU AI Act requires of deployers of credit scoring and life and health insurance pricing systems before first use, and it complements the data protection impact assessment. Risk treatment then means choosing controls that bring each risk inside appetite, recording why any risk is accepted, and keeping residual risk visible to the governance committee.
What support does Clause 7 require?
Clause 7 covers the resources that keep the AIMS alive: budget and people, competence of those who operate it, awareness across the workforce, and controlled documentation. For AI the competence point bites hardest. The people who are supposed to oversee or override a model need to understand what it does, where it fails and when to intervene, and you need records that show they were trained.
What operational controls does Clause 8 expect?
Clause 8 is where the management system meets production. You need a lifecycle process for how AI use cases are proposed and approved, how models are developed and validated, how they are released, monitored and eventually retired. You need defined triggers for reassessment: new data sources, retraining, a wider user population, a change of deployment context, or a serious incident. And you need governance of third-party AI, meaning the models, APIs and vendors you rely on, through due diligence, contract terms and ongoing monitoring that are proportionate to how much of the decision you have outsourced to them.
How do you prove it works under Clause 9?
Clause 9 requires measurement, internal audit and management review. Measurement means defined indicators for each system: accuracy, bias measures, drift, incident counts, human override rates, complaint volumes. Internal audit verifies periodically that the documented processes operate as written, with findings, owners and deadlines. Management review means senior leadership examining AIMS performance, audit results, incident trends and resource adequacy at least annually, and minuting the decisions; auditors ask for those minutes early because they show whether AI governance reaches the top of the firm.
What does Clause 10 mean by improvement?
When something goes wrong, and with models something eventually will, Clause 10 requires a structured response: root cause, correction, verification that the correction worked. More broadly it requires you to use incidents, audit findings, regulatory change and new technology to keep improving the AIMS. This is the Plan-Do-Check-Act loop that turns a set of documents into a living capability.
The technical components firms underestimate
A credible AIMS also needs technical plumbing that compliance teams rarely own and technology teams rarely budget for.
An AI inventory that stays current. Not a spreadsheet built for the audit. Each entry needs identification (name, version, owner, status), technical description, data lineage, risk classification including any EU AI Act tier, governance status (last assessment, validation state, approver) and integration points. Governance metadata should travel with the model, not live in a system nobody updates.
Model documentation, versioning and data provenance. Model cards have become the standard form: architecture and training method, training data and where it came from, performance broken down by demographic group, known limitations, intended and prohibited uses, and oversight requirements. Every version must link to its data snapshot, the change that triggered it, its risk assessment and validation report, and a tested rollback path. AI risk is largely data risk, so datasets are versioned against the models they trained, with collection, licensing, quality checks and bias characteristics recorded.
Monitoring for drift. Models degrade as the world changes. Watch four dimensions: data drift (input distributions shift), concept drift (the input-output relationship changes), performance drift (accuracy or recall falls) and bias drift (disparate impact emerges after validation). Set thresholds that trigger investigation and a revalidation schedule that runs whether or not drift has been detected.
AI-specific incident management. AI fails differently from ordinary software: biased output instead of an error message, an unexplainable decision instead of a crash, gradual decline instead of an outage. Define incident categories for accuracy, fairness, safety, transparency and robustness failures, with severity levels, containment steps such as rollback or a switch to human review, root cause analysis and the notifications that may follow; the EU AI Act's Article 73 serious incident duties and your supervisor's incident rules may both bite.
Human oversight that is real. A person glancing at a screen before clicking approve is not oversight. Define when review happens, what the reviewer sees, what authority they have and what training they received, then build it: interfaces that expose the reasoning, overrides that stop a decision taking effect until confirmed, and logs of every intervention and its justification.
Which implementation approach fits?
There are three broad routes, and most firms blend them. A manual, document-led approach suits a small firm with one or two AI systems: cheap and fully under your control, but it does not scale, is hard to audit and goes stale within a year. A consultant-led approach suits a firm that needs certification on a deadline or lacks in-house experience; the risk is dependency, so insist on capability transfer. A platform-supported approach suits a firm with several AI systems and continuing obligations: a system of record for obligations, controls and evidence automates collection, tracks control status continuously and gives the single view that shared drives cannot, and it pays back as the number of systems under governance grows.
On cost, budget for four lines: internal time (for a mid-sized firm, typically one to two full-time equivalents spread across roles over 9 to 12 months), any external expertise for gap analysis, design and pre-assessment, tooling such as registries and monitoring, and certification body fees for the Stage 1 and Stage 2 audits plus annual surveillance. Weigh that against avoided enforcement, faster client due diligence, and a board that can answer the AI question.
A typical implementation timeline
The figures below are typical ranges drawn from implementation experience, not commitments. Your own timeline depends on the number and risk of your AI systems, the maturity of your existing governance and the resources you can dedicate.
| Firm profile | Typical time to certification-ready | Main variables |
|---|---|---|
| Small: 1 to 10 AI systems, straightforward use cases | 4 to 6 months | Dedicated part-time resource, clear scope |
| Mid-sized: 10 to 50 AI systems, several business units, third-party AI | 9 to 12 months | Vendor governance, regulated-industry scrutiny |
| Large: 50 or more AI systems | 12 to 18 months for the initial scope, then phased expansion | Pilot scope first, usually the high-risk systems |
| Any firm that already holds ISO 27001 | Reduce the range above by roughly 30 to 40 percent | Shared backbone: policy, risk, audit, management review |
After certification, expect annual surveillance audits and a full recertification audit typically every three years.
How it relates to other frameworks
| Framework | Relationship to ISO 42001 |
|---|---|
| ISO/IEC 27001:2022 | Same harmonised structure. Extend the ISMS rather than build a parallel AIMS: one governance committee, one risk register with AI risks flagged, one audit programme, one evidence repository. Firms with ISO 27001 typically reuse 40 to 50 percent of their governance infrastructure. |
| NIST AI Risk Management Framework | Voluntary guidance organised as Govern, Map, Measure and Manage. It tells you what to do in detail; ISO 42001 gives you the certifiable system for doing it consistently. The two nest together. |
| EU AI Act (Regulation (EU) 2024/1689, amended by Regulation (EU) 2026/1744) | The legal obligation. ISO 42001 maps to the Act's risk management (Article 9), data governance (Article 10), transparency and human oversight (Articles 13 and 14), quality management (Article 17) and post-market monitoring duties, but it is not a harmonised standard and gives no presumption of conformity. The first harmonised standard, EN 18286:2026 on quality management for high-risk providers, was listed in October 2026 and includes an informative mapping to ISO 42001. |
| GDPR, UK GDPR and local data protection law | Govern the personal data the model uses. The AI impact assessment complements, and can be combined with, the data protection impact assessment; automated decision-making and profiling provisions apply alongside. |
| Supervisory IT and cyber rules | Regulators' IT risk guidance and cyber rules (the FSRA's IT Risk Management Guidance and GEN 3.5 are examples) reach the systems AI runs on and increasingly the models themselves. An AIMS gives you the evidence to answer them. |
Two points on the EU AI Act deserve emphasis because earlier guides, including ours, got the dates wrong. First, the Digital Omnibus on AI (Regulation (EU) 2026/1744, in force 27 July 2026) moved the high-risk obligations for Annex III systems to 2 December 2027 and for Annex I systems to 2 August 2028; prohibitions and AI literacy (2 February 2025), general-purpose AI model duties (2 August 2025) and Article 50 transparency duties (2 August 2026) did not move. Second, AML transaction monitoring and fraud detection are not Annex III uses, because the credit scoring entry carves out fraud detection, so most compliance tooling is not high-risk, though Article 50 and general deployer duties can still apply. A firm with no EU establishment is caught only where it places AI on the EU market or its output is used in the EU, a limb whose reach for non-EU financial firms remains unsettled: say "may apply" and record your reasoning per system.
Special considerations for regulated firms
Four things distinguish a regulated firm's AIMS. Accountability is personal: senior managers and approved persons remain responsible for decisions a model informs, so oversight and override records are their protection as well as audit evidence. Outsourcing rules already apply to AI vendors, so the due diligence and contract terms your regulator expects for any critical provider extend to model providers. Financial crime models sit under the MLRO, who needs visibility of monitoring model performance and tuning decisions to survive a thematic review. And licence applications are where the question is asked first: a scoped AIMS, even a light one, beats a promise.
A readiness checklist
- Inventory every AI system you build, buy or use, including AI features embedded in vendor software, and record where each system's output is used and who owns it.
- Classify each system by business criticality, risk and regulatory category, including whether the EU AI Act may apply and under which limb.
- Run a gap analysis against clauses 4 to 10 and Annex A, separating documentation, implementation and effectiveness gaps, and prioritise by risk and regulatory exposure.
- Agree an AI risk and impact assessment method with a rating scale your teams can apply consistently, templates, and triggers for reassessment.
- Write the AI policy and the procedures for use-case approval, development and validation standards, data governance, human oversight, incident response, vendor onboarding and change control.
- Appoint the roles: a governance committee, AI owners in the business, model owners in technology, the data protection officer, risk and internal audit, and record them in a RACI.
- Embed the controls in the workflow, so that an unassessed model cannot reach production, retraining triggers a fresh assessment, and no system ships without a monitoring plan.
- Put AI-specific due diligence, contract clauses and periodic review in place for every third-party model and API.
- Run one full internal audit cycle and one management review, fix the findings and minute the decisions.
- Commission a pre-assessment, then book Stage 1 (documentation) and Stage 2 (implementation) with an accredited certification body.
Common mistakes
Treating ISO 42001 as a documentation exercise
The policies exist but engineering, product and business teams have not changed what they do. Controls are described in documents and enforced nowhere, and auditors see it immediately when they interview staff and sample evidence. Put the controls into pipelines, procurement and approval workflows so that deploying an unassessed model is harder than assessing it.
Scoping out the inconvenient systems
The flagship product goes into scope and the finance team's predictive spreadsheet, the marketing team's generative AI and the alert-triage vendor stay out. Scope follows convenience rather than risk. Do the firm-wide inventory early and honestly; phased expansion is fine, excluding high-risk systems because they are awkward is not.
Building a parallel management system
A firm with ISO 27001 builds a separate AIMS with its own risk register, committee and control library, leaving staff with two conflicting sets of instructions. Extend what you have: AI risks go into the enterprise risk register, the existing committee widens its remit and one audit programme covers both standards.
Taking vendor AI on trust
"We use a certified vendor" is not an answer a supervisor will accept, because you remain accountable for how AI is used in your firm. Ask vendors about training data, bias testing, logging and their own AI Act posture, write the requirements into contracts, review periodically and track their incidents.
Stopping at launch
Thorough validation before go-live, then nothing, until a complaint or a regulator says otherwise. Make a monitoring plan a precondition of deployment, alert on defined thresholds and revalidate on a fixed schedule regardless.
Scattering the evidence
Risk assessments in one system, model cards in another, incident reports in email, so the assessment behind a control takes a day to find. Standardise templates from the start and keep one evidence store cross-referenced to the clauses and your regulatory obligations.
Frequently asked questions
Yes. ISO 42001 is a standalone standard with its own certification. In practice a regulated firm handling customer data will also need ISO 27001 or an equivalent security management system, and the two share the same clause structure, so building them as one integrated management system is cheaper and avoids conflicting controls. If you already hold ISO 27001, extend it rather than start again.
No. Certification shows that your AI governance is systematic and audited. The AI Act requires system-specific conformity assessment for high-risk systems, and ISO 42001 is not a harmonised standard, so it gives no presumption of conformity. It does provide the management system that makes the Act's duties on risk management, data governance, oversight and monitoring far easier to meet and evidence.
Only where the firm places AI on the EU market or its AI output is used in the EU, for example decisions about EU-resident customers. Many non-EU firms will not be caught; those that are face real obligations, with the high-risk duties for credit scoring and life and health insurance pricing applying from 2 December 2027. Classify each system and record the reasoning.
Almost certainly not by name. Most financial regulators rely on general systems and controls rules, outsourcing rules, IT and cyber risk guidance and data protection law to reach AI, and some have started issuing AI-specific guidance. ISO 42001 is a recognised way to show a supervisor that AI risk is governed, not a rule in itself. Check your own rulebook and recent supervisory letters.
Yes, proportionately. You are accountable for decisions made with AI in your firm regardless of who built the model. ISO 42001 expects due diligence on vendors, contract terms covering governance and incident notification, and ongoing monitoring. The depth should match how much of the decision you have delegated to the tool.
An ethics policy states your principles. ISO 42001 requires you to operationalise them through defined roles, documented risk and impact assessments, lifecycle controls, monitoring in production, internal audit and management review. The difference is between aspiration and evidence, and an auditor or supervisor will ask for the evidence.
Typically 4 to 6 months for a firm with up to ten AI systems and simple use cases, 9 to 12 months for a mid-sized firm with third-party AI and several business units, and 12 to 18 months for a large estate taken in phases. A firm that already holds ISO 27001 can usually cut those ranges by 30 to 40 percent because the governance backbone exists.
Two stages. Stage 1 is a documentation review in which the auditor checks that the AIMS design meets the standard. Stage 2 is the implementation audit: interviews, observation of processes and sampling of evidence to confirm the controls operate as written. Nonconformities must be closed within set timeframes, and the certificate is then maintained through annual surveillance audits and recertification.
Key takeaways
ISO/IEC 42001:2023 is the certifiable standard for governing AI across its lifecycle, built on the same structure as ISO 27001. No regulator requires it by name, but systems and controls rules, outsourcing rules, IT risk guidance and data protection law all reach AI, and clients increasingly ask. The EU AI Act applies to a non-EU firm only where its AI touches the EU market or EU people; where it does, the financial services high-risk deadline is 2 December 2027. Certification is a governance credential, not legal compliance. What matters is the inventory, the risk and impact assessment, controls embedded in the workflow, monitoring after launch and real human oversight.
Sources
- ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system: https://www.iso.org/standard/42001 (the ISO site blocks automated fetching; edition confirmed by reference, 8 November 2026)
- ISO/IEC 27001:2022, Information security management systems, Requirements: https://www.iso.org/standard/27001 (checked by reference, 8 November 2026)
- Regulation (EU) 2024/1689 (the AI Act), EUR-Lex: https://eur-lex.europa.eu/eli/reg/2024/1689/oj (checked 8 November 2026)
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), OJ L 2026/1744, 24 July 2026, EUR-Lex: https://eur-lex.europa.eu/eli/reg/2026/1744/oj (checked 8 November 2026)
- European Commission, AI Act policy page: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai (checked 8 November 2026)
- Commission Implementing Decision (EU) 2026/2274 on harmonised standards (EN 18286:2026), 12 October 2026: https://eur-lex.europa.eu/eli/dec_impl/2026/2274/oj (checked 8 November 2026)
- NIST AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework (checked 8 November 2026)
- ADGM FSRA, Information Technology Risk Management Guidance, 20 November 2024 (example of supervisory guidance on algorithm-driven solutions): https://assets.adgm.com/download/assets/Information+Technology+Risk+Management+Guidance+20241120.pdf/014f7430a71811efa4b7bae6ac7fc29c (checked 8 November 2026)
- ADGM Data Protection Regulations 2021 (consolidated), as an example of local data protection law with automated decision-making provisions: https://assets.adgm.com/download/assets/ADGM+Data+Protection+Regulations+2021+Updated.pdf/146aa34858b011efb99a36e29b0f3a63 (checked 8 November 2026)