AI Law, Risk & Governance

Building AI Governance: Committees, Policies, Accountabilities

Marin Niehues23 December 202412 min read

Wood-panelled courtroom

AI governance is the decision system that defines who decides on AI applications, which rules they operate under and who answers for their results. This article describes a model with four building blocks, places the recognized frameworks and sets out a sequence for the build-up.

Why does a company need AI governance?

A company needs AI governance so that low-risk applications reach operations quickly and high-risk applications reach them with due care. This article provides an overview and does not constitute legal advice.

In many business units, governance is regarded as a brake. From a delivery perspective it is the opposite. It determines in advance which decisions a business unit takes itself and which require a review or executive management. Without that definition, every new application is decided by whoever speaks loudest or holds out longest. The result is pilots that wait months for approval, alongside tools that run in business units without any review.

Both depend on a model that classifies each application by its risk. The internal risk grid mirrors the risk classes of the AI Act: anything the regulation classifies as high-risk automatically goes into the highest internal class. Under Annex III this includes, for example, systems used to select candidates or to evaluate employees. Data protection and copyright feed in as review criteria, as does the AI literacy obligation under Article 4 of the AI Act. Interpretation in individual cases belongs to specialist lawyers who take part in the committee. An overview of the regulation is given in the article on the EU AI Act.

Is there a framework or template for AI governance?

There are three recognized reference points for AI governance: the standard ISO/IEC 42001, the AI Risk Management Framework of the US National Institute of Standards and Technology (NIST) and the risk logic of the AI Act. None of them is a template a company simply fills in. They provide the structure the company's own model follows.

ISO/IEC 42001 is the international standard for AI management systems and was published in 2023. It follows the structure of the ISO management system standards that companies know from ISO/IEC 27001. An annex contains 38 controls, from which the organization selects the applicable ones on a risk basis. The standard is certifiable, but there is no legal obligation to certify.

The NIST AI Risk Management Framework is voluntary and organizes the work into four functions (Govern, Map, Measure, Manage). It serves well as a grid for risk assessment and metrics. The AI Act classifies AI systems into risk levels from prohibited practices to minimal risk and attaches obligations for providers and deployers.

For most companies one combination is right. The internal model adopts the risk logic of the AI Act as its mandatory part and the structure of ISO/IEC 42001 as its outline. It can then be certified later without rework, should customers or tenders require evidence.

Who should sit on an AI committee?

An AI committee is a small group with decision authority, and before it needs members, it needs a mandate. The mandate answers three questions. Which decisions sit with the committee? Which applications may a business unit start without a submission? Within what period does the committee respond to a request? A committee that leaves these questions open becomes a collection point for concerns and lengthens every approval.

It includes the functions without which an approval would lack substance:

  • a representative of the business unit that uses AI
  • IT and information security
  • data protection and legal
  • HR, as soon as applications affect employees

The chair is a member of executive management or a person it appoints with decision rights. In companies with a Chief AI Officer, that role takes the chair. The committee meets at a fixed rhythm, works from a two-page decision paper and decides in the meeting. Postponements are exceptions with a stated reason. Voting rights stay with five to seven members; further specialists join as non-voting guests where needed.

Policies: Few Rules, Each with a Checkpoint

The most common error in policies is their length. A long document is read at launch and rarely opened afterwards. A short rulebook works better, in which every rule has a checkpoint: it is defined who verifies the rule, when, and what happens in case of deviation.

A robust basic set consists of three documents. The framework policy governs permitted use and data classes as well as the labeling of AI outputs. The risk grid assigns applications to classes. The approval process sets out the form and deadlines as well as the responsible parties.

Rules only work when the permitted route is faster than the detour. Anyone who receives an approved tool within a week has little reason left for a private account. The design of the approval process therefore determines compliance more strongly than any threat of sanctions.

Accountabilities and Metrics

Every AI application in operation needs a named owner of responsibility. A split into three roles, recorded in a register, has proved its worth:

  • The business owner in the using business unit is accountable for benefit and purpose.
  • The technical operator, usually IT or a platform team, is accountable for stable operation.
  • The review function, independent of the owner, assesses quality and risk at fixed intervals.

Without this assignment, nobody knows who decides in an incident. For class 1, the owner with a register entry is sufficient. From class 2 onwards, all three roles are named before launch.

The fourth element is steering by metrics. The executive team needs a few figures that it sees regularly:

  • number of applications per risk class
  • lead time from request to approval
  • share of applications with a named owner
  • number and handling time of incidents

Approval lead time is the most revealing figure. When it rises, business units work around the process and governance loses its effect. A mandate of five working days for class 2 and a fixed slot in the next executive meeting for class 3 has proved effective.

Building in the Right Sequence

The sequence decides acceptance. A three-step approach has proved its worth.

The first step is an inventory of all running applications, including those never reported. It shows the committee where the need for rules is greatest.

The second step creates a minimum model of mandate and risk grid, together with framework policy and register. It goes live within a few weeks and initially covers class 2 and class 3 applications. The first real cases test mandate and deadlines. Staff training starts in parallel, because the AI Act requires providers and deployers to ensure sufficient AI literacy among their staff.

In the third step, the model is adjusted after the first decisions. Rules that are too tight are loosened, gaps are closed and the metrics gain reference values. Only now does it pay to add detailed rules for individual business units or to align the structure with ISO/IEC 42001. How AI governance fits into a company's overall AI management is shown in the article AI management: what executive leadership must steer.

Niehues Consulting as Partner for the Build-Up

Governance is a delivery project that has to be built into the organization's decision paths. Niehues Consulting supports executive and business-unit leaders with three services. In AI management, requirements and a business case are produced together with the risk grid, as the basis for selecting applications and steering the AI portfolio. In operationalisation and implementation, processes and responsibilities are set up so that business units accept the rules. Delivery management and organization design define roles and decision paths so that committee and owners work reliably with the operator.

The approach follows the Niehues Delivery Model with the phases Understand, Build and Implement. Three of its review questions matter most for governance. In Understand: who decides, who delivers, who uses? In Build: which roles and decision paths does operation require? In Implement: who is accountable for operation from day one? Marin Niehues brings experience from projects at VC-funded scale-ups and DAX 40 groups in 20+ countries. The entry point is a free 30-minute initial consultation via the contact page.

Frequently asked questions

What is AI governance?

AI governance is the system of committees and policies together with accountabilities that steers the use of AI in a company. It regulates who approves applications by which criteria and who is responsible for them in operation. Good governance accelerates low-risk applications and secures risky ones.

What is the best AI governance framework?

The recognized frameworks are ISO/IEC 42001 for AI management systems and the NIST AI Risk Management Framework, with the risk logic of the AI Act as the binding layer in the EU. None of them is a fill-in template. Companies adopt the structure of the standard and populate it with their own mandate, risk grid and roles.

What is ISO/IEC 42001 and do we need certification?

ISO/IEC 42001 is the international standard for AI management systems, published in 2023. There is no legal obligation to certify. Certification pays off for providers of AI solutions and for companies whose customers require evidence.

Who should sit on an AI governance committee?

An AI committee consists of a small group with decision authority. Alongside the business unit as user, it includes IT with information security as well as data protection and legal, and HR where employees are affected. The chair is a member of executive management or an appointed person.

What roles does AI governance require?

AI governance requires three roles per application, namely the business owner and the technical operator as well as an independent review function. The owner is accountable for the benefit, the operator for stable operation. The review function assesses quality and risk at fixed intervals.

What is the difference between AI governance and data governance?

Data governance regulates ownership and quality of data, while AI governance regulates approval of and accountability for AI applications. The two interlock because every application uses data. The data classes from data governance therefore feed into the risk grid of AI governance as a criterion.

Do we need a Chief AI Officer?

A company needs one named person accountable for AI governance, and a dedicated executive role becomes necessary only with a large portfolio. With few applications, a member of executive management chairs the committee. Once the portfolio spans several business units, a dedicated role with budget and mandate pays off.

How many AI policies does a company need?

Three documents are enough to start: framework policy, risk grid and approval process. Further rules follow from the committee's first decisions. The shorter the rulebook, the more likely employees are to apply it.

When should a company set up AI governance?

The right time is the start of broad use, before the rollout and before applications multiply in the business units. If AI already runs without rules, the build-up begins with an inventory of existing applications. Legal questions in individual cases are for specialist lawyers.

Let’s talk about your initiative.

In 30 minutes we clarify where your organization stands and which step will pay off most now.

Request a consultation

30 minutes · free · no obligation