AI GovernanceDo the controls actually work?
AI Governance

AI Governance & Oversight

Governance and a management system are two different things. Governance directs and oversees, and a management system delivers against that direction. Most of what gets called AI governance is really AI management. That’s real work, and it covers most of what regulation requires. But it can’t decide what the AI is for. That decision belongs to the governing body. This page sets out both, in the order you would take them on.

Written against
ISO/IEC 38500:2024 · ISO/IEC 38507:2022 · ISO/IEC 42001:2023 · ISO/IEC 42005:2025 · ISO/IEC 42006:2025 · ISO/IEC 23894:2023 · NIST AI RMF 1.0 · IEEE 7000-2021 · Regulation (EU) 2024/1689
Last updated
Answered first

How does the board direct AI?

The governing body evaluates, directs and monitors: it decides what AI is for, which uses are prohibited, how far judgement may be delegated to machines, which decisions stay human, the risk appetite and the spending. Management then delivers against that direction in its own plan, do, check, act loop, which ISO/IEC 42001 specifies in clauses 4 to 10. Programmes usually fail where the two loops meet, so the direction has to be explicit.

What does an ISO/IEC 42001 management system deliver?

One auditable system that combines policy, risk assessment, controls, roles and continual improvement. Your controls come from your own risk treatment and are checked against the reference set in Annex A. An accredited certification body can certify the system. That helps with preparation and external verification; it does not prove that the AI produces good outcomes.

How does it fit the systems we already run?

It extends them: avoid a separate AI management system if an information security or quality management system already exists. Share the common machinery: document control, competence, internal audit, management review and corrective action. Keep the AI-specific elements distinct: the AI risk assessment, the impact assessment, the Statement of Applicability and the data provenance controls.

The principle underneath

Governance directs and controls. Management delivers.

Governance has a precise meaning. It isn’t another word for oversight, compliance or a committee. ISO/IEC 38500 defines the governance of IT as the system by which its current and future use is directed and controlled. The governing body does three things: it evaluates the options, it directs by setting objectives and policy, and it monitors performance against that direction.

Management is the other half, and a different job. It plans, builds, runs and monitors the activities that deliver what the governing body directed. Governance sets the what and the why; management owns the how. The two are easy to blur: a steering committee that reviews project status is doing management, whatever the meeting is called.

The same standard says plainly what governance aims at, and it makes a useful test: technology whose use is effective, efficient and acceptable. Effective means it does what it was directed to do. Efficient means it’s worth what it costs. Acceptable means you can defend it to the people it affects, to a regulator and to the market. ISO/IEC 38507 carries that framing into AI.

Governance — the governing body
Evaluate→Direct→Monitor↺

Evaluate the options and the pressures. Direct by setting objectives, policy and risk appetite. Monitor performance and conformance against that direction. None of the three verbs is project approval.

Management — the organization
Plan→Do→Check→Act↺

Plan against the direction received. Do: build, acquire, operate. Check through measurement, internal audit and management review. Act on what that showed. This is the loop ISO/IEC 42001 specifies, across clauses 4 to 10.

The two loops interlock, and programmes usually fail at the joins. Direct feeds Plan: management can’t align to a direction nobody set, so it substitutes its own. Check feeds Monitor: without reliable measurement, the governing body is directing blind, and usually knows it. And Monitor returns to Evaluate, so the direction gets revised instead of quietly going out of date.

Look at what a typical AI governance programme contains: inventories, policies, risk assessments, model documentation, monitoring, evidence, certification. Most of that is plan, do, check, act. The platforms are tools for it: workflow, registers, evidence capture. The committee is usually a review body that reports into the same layer.

The governance half is smaller and more specific, and no compliance programme forces you to do it. What will we use AI for, and what not? How much judgement are we prepared to hand to a system nobody can fully explain? How much opacity is acceptable when an outcome affects a person? What will we spend, and for what return? Who answers when it goes wrong? No management system can answer these for the organization.

When nobody answers them, management runs ahead of direction. Without direction, management optimises for the only goal anyone stated, which is usually evidence of compliance. So the inventory gets built, the policy gets written and the assessment gets filed. None of it serves an intent, because nobody set one. The documents pile up, and nobody works any differently.

That doesn’t make the management work optional. It’s most of the effort, most of what regulation requires, and most of what an auditor will examine. It’s the management half of the job, and doing it before the direction is clear is expensive.

The introduction to ISO/IEC 42001 makes two points we work to. First, an AI management system extends the management structures you already have, instead of running beside them. So the answer to “do we need a separate AI governance function?” is usually no. Second, an organization has to find an appropriate balance between governance mechanisms and innovation. Direction that only constrains is a failure of governance, not an excess of it.

How this page uses the word

On this page, “governance” means both halves, because that’s how the market uses the word. Where the difference matters, we name them. ISO/IEC 38507 covers the governing body’s work: direction, risk appetite, accountability and oversight of the AI portfolio. ISO/IEC 42001 covers the management system that delivers against it. The sections below deal with both.

Part one

Direction

Evaluate · Direct · Monitor

The half the governing body owns. What it decides, and who answers for it.

This half is made of decisions, not documents, so it helps to be concrete about them. There are three verbs, and a small number of decisions nobody else in the organization can take.

Evaluate · Direct · Monitor, in practice
Evaluate

The governing body needs a clear view before it decides anything: what the portfolio costs, what it returns, where the exposure sits and what has changed since last time. Most of that comes from the inventory and the indicators.

Direct

The decisions nobody else can take. What AI is for in this organization. Which uses are out of bounds regardless of return. How much judgement may be delegated to a machine, and where a human decision stays mandatory. Risk appetite, expressed as thresholds rather than adjectives. What will be spent, and against what expected return.

Monitor

Conformance against that direction, and performance against the objectives it set, reviewed on a fixed schedule. A board that only looks once something has broken is reacting, and by then the decision has been made for it.

Risk appetite is the decision everything downstream depends on, and ISO/IEC 38507 puts it with the governing body on purpose. How much bias, uncertainty, opacity and autonomy will the organization accept? The team building the system shouldn’t be making that decision on the organization’s behalf. Setting that boundary is where an organization says how much it is willing to be wrong about, and in which direction.

Governance often breaks over who is accountable. The controls exist and the policy names them, but when one doesn’t operate, nobody in particular has to answer for it.

The five questions

Human oversight with real authority

For every consequential AI system, five questions whose answers are names: Who owns it? Who monitors it? Who can challenge it? Who can stop it? Who answers for the outcome? If any answer is a function, a committee or a policy, the system is unowned.

Part two

Delivery

Plan · Do · Check · Act

The half the management system runs. What must exist, where it lives, and how it is tested.

04

What controls are, and where they come from

Permalink to “What controls are, and where they come from”

A control is a measure that maintains or modifies risk. Everything in this half of the page is a control, a way of deciding which controls you need, or a way of checking that they work. So it’s worth being exact about where controls come from.

ISO/IEC 42001 offers a reference set of controls in Annex A. You determine the controls your own risk treatment requires, then check against Annex A that nothing necessary was left out. The EU AI Act works differently. It imposes requirements rather than controls: Articles 9 to 15 state what a high-risk system must achieve, not how to achieve it. They cover risk management, data governance, documentation, logging, transparency, human oversight, accuracy, robustness and cybersecurity. Where they apply, map those requirements onto the controls you actually implement. Then every obligation has something concrete answering it, and one control can answer several obligations.

That mapping is the core of an AI management system built for a specific organization. You get one control list, derived from your own risk treatment. It is cross-referenced to the Annex A objectives and to the legal obligations that apply to the systems in scope, and the gaps are visible. Without it, you get a control set that satisfies an annex, a separate compliance workstream chasing the Act, and no single view of what is actually covered.

05

Responsible AI is a requirement, not a posture

Permalink to “Responsible AI is a requirement, not a posture”

A requirement to be fair doesn’t tell a delivery team what to build. Everyone accepts the obligation, but it arrives as a values statement with no way to act on it. IEEE 7000 closes that gap by treating ethical values as requirements: elicited from stakeholders and traced into the design, the way a performance or security requirement would be.

Specifying a control and enforcing it are different jobs, and governance is exposed in the gap between them. Written governance describes intent, while production enforces behaviour. In between sit a human-oversight policy with no point in the workflow where a human can intervene, an approval nobody recorded, and a retention rule the pipeline never implemented.

The gap widens with autonomy. Agentic systems plan, call tools and pass context to one another, so the decision path a control was meant to constrain is put together at runtime, not at design time. Controls written for request-and-response don’t bind a system that composes its own steps.

So enforcement has to be designed in: where a human can intervene, what gets logged, which data a system can reach and which decisions an agent may take unaided. That’s architecture more than paperwork, and it decides whether an audit is a formality or a problem.

What changes when you bought it rather than built it

The common mistake is to assume the vendor’s compliance covers yours. It does not. Their conformity assessment establishes that the system can be used lawfully. Nothing in it establishes that you are using it lawfully. A lot stays in your hands: which decisions the system may touch, what data you feed it, where a human can step in, and what you log.

07

One management system, several standards

Permalink to “One management system, several standards”

Don't build an AIMS island. If a company already has an ISMS or QMS, there is no reason to build another parallel inventory, risk methodology, supplier process, audit programme, management review, competence system and corrective-action process just because ISO/IEC 42001 appeared. Determine what can be shared, what needs extending, and what genuinely needs to remain AI-specific.

Three lenses on one business process
Three lenses on one business processRisk management is primarily concerned with…
ISO 9001Can our processes consistently deliver products and services that meet requirements and customer expectations?
ISO/IEC 27001What information do our business processes depend on, and what happens to those processes if its confidentiality, integrity or availability is compromised?
ISO/IEC 42001What risks and impacts arise from AI, and what AI opportunities should we pursue?

Formally, an ISMS preserves the confidentiality, integrity and availability of information through risk management. Why that matters is one level underneath: information supports the business processes where value lives. So we build the integration with information and business processes as the primary assets. Systems, applications, infrastructure and suppliers are supporting assets: they matter because of the processes and information that depend on them. ISO/IEC 27001 does not require a formal business-process inventory, so a system can be certified without one, and the link to the processes often stays implicit. We make it explicit. The question changes from "SAP is a critical asset" to "which business processes depend on SAP, and what happens to them if it is compromised?" An AI model is a supporting asset too: if it is unavailable, manipulated or corrupted, the process it serves is affected.

All three plan for risks and opportunities at the level of the management system. What each adds is its own: 9001 provides the process and performance backbone. 27001 provides the information-security and asset-protection layer, protecting the information the business processes depend on. 42001 provides the AI governance and impact layer: the risks, impacts and opportunities that arise because AI is used. ISO 31000 supplies the common risk logic underneath.

Take an order-to-delivery process. Through the 9001 lens, a risk is that incorrect production planning causes late customer deliveries. Through the 27001 lens, loss or corruption of production-planning information interrupts production. Now introduce AI-based production planning. 42001 adds that the optimizer could produce inappropriate schedules because of model, data or performance limitations. It also adds the other side: the optimizer could improve machine utilization, reduce lead times and improve delivery performance, the same delivery performance the 9001 lens measures. These aren’t three unrelated management systems. They’re three views of the same business process: one process, one risk architecture with distinct assessments inside it, rather than three independent risk registers.

An organization already running a quality management system does not need a second management system to govern AI. Building one produces two sets of documents, two audit programmes and two versions of the truth about who is responsible for what.

The join is structural. ISO 9001 and ISO/IEC 42001 are deliberately written to the same clause architecture. So context and interested parties, leadership and policy, document control, competence, internal audit, management review and the corrective-action loop are one set of machinery serving a second purpose. An AI management system built there inherits something that already works instead of standing one up alongside it. Corrective action is the most valuable of those inheritances, and where the two halves meet most often once the system is in daily use.

What holds for quality holds more widely. The same harmonised structure carries business continuity in ISO 22301, information security in ISO/IEC 27001 and compliance in ISO 37301 (obligations identified, documented and kept current). An organization holding any of them extends machinery it already trusts instead of building a parallel set for AI. The continuity join is the one AI has made acute, because ISO/IEC 42001 does not carry the discipline. Nothing in it asks for a business impact analysis, for strategies and plans to keep operating, or for an exercise programme. It requires the AI system to be operated, monitored and logged, and its third parties managed. It does not say what the business does when the system is gone. ISO 22301 keeps the organization running when an AI-dependent operation is disrupted. It says nothing about whether the model was right.

What must not be merged is the AI-specific work. The AI risk assessment, the assessment of impacts on individuals and society, the statement of applicability, and the controls over data and its provenance have no counterpart in quality management. Folding them into generic risk-based thinking dilutes them until they satisfy an auditor and govern nothing. Objectives and their planning belong on that side too. They look structurally shared, but this is precisely where the two standards diverge. Integrate the common structure. Keep the AI-specific half whole.

Two management systems only become one when information and actions flow between them in both directions. A nonconformity raised in operation reaches the AI risk assessment when its cause is the system’s behaviour rather than its execution. An impact finding that changes what the system may do re-enters as a change to requirements, controlled like any other.

Since October 2026 this has an official form: the Commission describes EN 18286, the harmonised standard for the Article 17 quality management system, as a layer on top of ISO 9001 or ISO/IEC 42001 rather than a replacement.

Certification is worth having, and it is not the point

Two things are true at once, and the case for certification only holds when both are stated.

Certification is worth having. An external audit checks whether the management system actually does what its documents say, and the answer is not always yes. Preparing for the audit is worth most of the fee on its own.

But no regulation requires it, and it does not make AI pay off. A certified organization can still run AI that returns nothing, because the certificate says how decisions are made — not whether they were good decisions, and not what came of them. It answers one question, about governance, and answers it well. Buy it for that, and expect nothing else from it.

08

Testing that the controls still work

Permalink to “Testing that the controls still work”

A control that was designed correctly but never tested is an assumption. Assurance is what turns it back into a control.

Assurance asks two questions. Design effectiveness asks whether this is the right control for the risk. For example, can a human review step prevent anything if it comes after the decision has already taken effect? Operating effectiveness asks whether the control is working now, at its current volume, with the people doing it today. A control can pass one test and fail the other, and audit findings get written in that gap.

Control indicators show whether each safeguard is operating as designed. They’re the earliest signal in the chain: controls degrade first, exposure rises next, and value is lost last. We sort them into three groups, and the last one checks whether the controls are keeping up with a portfolio that changes faster than the review cycle.

  • A · Are the controls in place?
  • B · Are they working?
  • C · Are they keeping up?
10

One management system, made auditable

Permalink to “One management system, made auditable”

ISO/IEC 42001 brings the guardrails, the board’s risk appetite and the ethical requirements together in one certifiable management system, giving AI the shape ISO/IEC 27001 gave information security.

  • 01

    ISO/IEC 42001 — the management system

    Policy, risk assessment, controls, roles and continual improvement across the AI lifecycle. A third party can certify it, which turns your claims about it into evidence.

  • 02

    Connected, not parallel

    The AI management system extends the risk and compliance practice already running. A second, separate governance function competing with the first is an expensive failure.

More questions

Who is accountable for each AI system?

Named people, not a function or a committee. Each consequential AI system needs people who own it, monitor it, can challenge it, can stop it, and answer for its outcomes.

How do we test that controls still work?

Test design effectiveness and operating effectiveness separately. Design effectiveness asks whether a control is the right one for the risk. Operating effectiveness asks whether it works in practice, at current volume and with current staff. Control indicators then show whether controls are in place, working, and keeping up with change.

How do the EU AI Act requirements relate to our controls?

The Act states requirements rather than controls, so we map its requirements onto the controls you implement, and one control can then answer several obligations. ISO/IEC 42001 Annex A gives a reference control set. The high-risk requirements in Articles 9 to 15 are the ones to map.

Get direction and delivery working together.

Tell us what governance exists today and where AI actually runs. We will respond with a read on how the direction and the management system line up, and a sensible first step — often an ISO/IEC 42001 gap analysis.