AI Governance

AI Governance Operating Model: Roles, Decisions and Accountability

  • AI Governance
  • Responsible AI
  • AI Risk Management
  • Data Governance

An organisation introduces an AI assistant to help staff search internal documents. IT configures it, procurement signs the contract, and employees start using it. A month later, the assistant reveals information to someone who should not have access.

Who owns the problem? Who can suspend the assistant? Who approved the access arrangements?

These questions belong in an AI governance operating model. Before deployment, an organisation needs a clear way to assign responsibility, make decisions and respond when something goes wrong.

What is an AI governance operating model?

An AI governance operating model describes how an organisation puts AI governance into everyday practice. It connects policies and principles to named roles, decision-making authority, working processes and evidence.

For each AI use case, it should explain who owns it, which decisions require approval, what checks support those decisions, and how concerns are escalated.

A systematic review by Batool and colleagues (2025) organises AI governance around four questions: who is accountable, what is governed, when governance happens and how it is implemented. Drawing on 28 articles, the review provides a useful starting point for thinking about organisational AI governance.[1]

The operating model below is a practical proposal informed by those questions. The review does not prescribe this exact structure or validate it as a universal model.

Start with decisions that need an owner

A useful first step is to identify the decisions your organisation actually makes about AI.

Can a team purchase an AI tool? Can staff upload confidential documents? Can an assistant send messages or change records without approval? Who decides whether a serious error requires suspension?

Each decision needs an accountable role, supporting expertise and a record of the outcome. The depth of review should reflect the potential consequences, the information involved and the system’s ability to take action.

A tool that drafts meeting agendas may need a simple approval process. An AI system influencing recruitment or accessing sensitive research information deserves closer scrutiny. Define the criteria for those different routes so teams know what is expected.

Define AI governance roles and responsibilities

The following roles are a starting structure. Smaller organisations may combine responsibilities, while larger organisations may distribute them across established teams.

Executive sponsor. Sets organisational expectations, provides resources and resolves issues beyond delegated authority. The sponsor should establish who may approve higher-risk uses and accept remaining risks within defined limits. Internal approval cannot authorise unlawful use.

Business or use-case owner. Owns the purpose and day-to-day use of the AI system. This person explains the intended benefit, identifies affected people, ensures appropriate use and follows up on performance or incidents.

Technical owner. Manages configuration, integrations, access controls, logging and technical changes. For purchased systems, the organisation still needs someone to coordinate these matters with the supplier.

Specialist reviewers. Privacy, security, legal and relevant subject experts assess issues within their competence. Their advice should inform approval. Assigning a review to the DPO, for example, should not automatically make that person the owner of every AI risk.

Approval authority. A designated manager or governance group decides whether the proposed use may proceed, under what conditions and when it needs further review. Record significant disagreements and how they were resolved.

Users and supervisors. Need clear instructions about permitted use, output checking, reporting concerns and escalation. Where human review is a control, give reviewers the time, information and authority to intervene.

For significant risks, arrange review by someone with sufficient independence from the delivery team to challenge its assumptions.

Make decision rights visible

A responsibility table helps translate these roles into action. The example below can be adapted to existing approval structures.

Decision Accountable role Supporting roles Evidence to retain
Propose an AI use case Business owner Users and technical owner Purpose, scope and expected benefit
Approve deployment Designated approval authority Business owner and specialist reviewers Assessment, test results and approval conditions
Grant access to information Relevant information owner Technical owner and security team Access decision and permission checks
Approve a material change Authority defined by the change process Business and technical owners Change assessment and updated testing
Suspend operation Named role with emergency authority Incident responders and business owner Reason, action taken and affected systems
Approve restart after an incident Designated restart authority Technical owner and relevant reviewers Remediation and verification evidence

“Everyone is responsible” provides little direction during an incident. Name the role that decides, the people who advise and the person who carries out the action.

Apply governance throughout the AI lifecycle

Before deployment, record the use case in an AI inventory. Identify its owner, supplier, intended users, information sources, limits and potential impacts. Complete the appropriate assessments and testing, then record approval conditions.

During operation, review evidence relevant to the use case. This might include incorrect outputs, inappropriate access, complaints, overrides and incidents. Decide who examines these signals and what would trigger action. Collect only the information needed, with suitable access and retention controls for logs.

When something changes, check whether the original approval still fits. A new model, expanded user group, additional data source or permission to act independently may require reassessment. Define material changes in advance.

At retirement, remove integrations and credentials, address retained information, update the inventory and preserve records needed to explain past decisions.

Set a review schedule appropriate to the risk, alongside event-triggered reviews. An annual review alone may leave important changes unexamined for too long.

Example: an AI assistant for internal documents

Consider an organisation introducing an assistant to answer questions from internal policies.

The operations manager owns the use case. IT manages the integration. Information owners decide which document collections may be connected. Security checks whether the assistant respects existing access restrictions, while privacy specialists assess any personal information involved.

The approval authority authorises a limited pilot. Its conditions include restricting access to approved documents, testing permission boundaries and requiring answers to link back to their sources. Users receive instructions for reporting incorrect or inappropriate responses.

During the pilot, the business owner reviews reported problems with the technical owner. A designated incident responder can disable the integration immediately if information is exposed. Restart requires evidence that the cause has been addressed and the relevant checks repeated.

If the team later proposes allowing the assistant to edit policies, that expanded capability triggers a new assessment. The original approval covered answering questions.

Put the model to work

Choose one existing AI use case and walk through its decisions. Identify its owner, approval authority, review requirements and suspension route. Then test a realistic scenario: an incorrect output, an access failure or an unexpected supplier change.

Use any gaps to improve the operating model. A practical measure of AI governance is whether the people involved can explain who decides, what evidence they need and how they act when the system causes concern.

For a starting point, see How to Start Your AI Governance Journey and What Is AI Governance?.

Reference

  1. Batool, A., Zowghi, D. & Bano, M. (2025). AI governance: a systematic literature review. AI and Ethics, 5, 3265–3279. https://doi.org/10.1007/s43681-024-00653-w