Every governance framework eventually confronts the same practical question: who, exactly, does what? Without an operating model – a defined structure specifying which functions are responsible for which AI governance activities – even a well-written AI policy becomes aspirational prose. This chapter maps the institutional architecture that makes governance real.
The Three Lines of Defense, Applied to AI
The three-lines-of-defense model – long a staple of financial services risk management and formally codified by the Institute of Internal Auditors (IIA) in its 2020 update to the Three Lines Model – applies directly to AI governance. The IIA’s updated model moves away from “defense” language toward a framework focused on governance, value creation, and risk management, while preserving the core structure of distributed accountability.
Applied to AI, the three lines operate as follows:
First Line – AI Product, Engineering, and Business Teams. These teams own and manage AI risk in day-to-day operations. They build models, select training data, deploy systems, monitor outputs, and are the first accountable party when a model behaves unexpectedly. In practice, this means AI developers must document model design decisions, conduct pre-deployment bias testing, maintain model cards (structured documentation of intended use, performance, and limitations), and flag emerging risks through defined escalation channels. EY’s AI governance discussion paper specifies that the first line owns model inventory management as a live operational responsibility, not a one-time compliance exercise.
Second Line – Risk, Compliance, Privacy, and Governance Functions. These functions provide oversight, expertise, challenge, and monitoring on AI-related risk matters. They review high-risk AI deployments before launch, set risk appetite boundaries, conduct periodic audits of model performance against approved parameters, and ensure regulatory requirements are translated into operational controls. They do not build models – they assess, challenge, and govern them. Jakub Schuett’s analysis published on the AI Governance Library notes that the second line in AI contexts includes legal counsel, compliance officers, red teams, and dedicated AI ethics or governance functions – and that the model’s key contribution is making organizational AI governance auditable.
Third Line – Internal Audit. Internal audit provides objective, independent assurance to the board and senior leadership on whether AI risk management and controls are actually working. In AI contexts, this means evaluating whether model inventories are complete and current, whether second-line review processes are functioning as designed, whether model documentation meets policy requirements, and whether incident response processes have been tested. PwC’s 2025 Responsible AI Survey recommends applying the three-lines model explicitly to AI governance and assigning each line clear performance indicators that can be reported to the board.
RACI for Key AI Decisions
Three AI governance decisions require explicit RACI (Responsible, Accountable, Consulted, Informed) assignment at every organization:
Model Approval. The first line (AI product owner, engineering lead) is Responsible for preparing documentation and initial testing. Risk/compliance is Accountable for final approval of high-risk deployments. Legal, privacy, and IT security are Consulted. The relevant business executive and board committee are Informed.
Incident Response. AI Ops or the technology team is Responsible for immediate containment. The AI product owner is Accountable for the response and remediation plan. Legal, communications, and privacy are Consulted depending on breach type. The CEO, board risk committee, and relevant regulators are Informed as required by policy and law.
Vendor and Tool Selection. Procurement and the AI product team are Responsible for the assessment. The second line (risk, compliance, privacy, legal) is Accountable for sign-off. IT security and data governance are Consulted. Finance and the relevant business unit executive are Informed. MSP Corp’s AI governance guide for IT teams notes that approval workflows must be explicit about what evidence is required at each stage – not just who approves, but what they are approving against.
Where Governance Sits in the Organization
The organizational home of AI governance varies by institution and reflects both maturity and organizational structure. Common models include:
Chief AI Officer (or Chief Responsible AI Officer). This emerging role – present at IBM, Google, and a growing number of major financial institutions – owns AI strategy and governance in an integrated function. Francesca Rossi at IBM holds the title of IBM Fellow and AI Ethics Global Leader, with a mandate spanning responsible AI policy, governance frameworks, and external engagement, including co-chairing OECD expert groups on AI futures and agentic AI.
Chief Responsible AI Officer / AI Policy Lead. Microsoft’s Natasha Crampton, previously Microsoft’s Chief Responsible AI Officer, led the company’s Office of Responsible AI and the Aether (AI and Ethics in Engineering and Research) Committee – the internal body responsible for providing guidance on AI governance decisions across Microsoft’s product portfolio. Salesforce has embedded AI ethics governance through Paula Goldman, Chief Ethical and Humane Use Officer, who oversees Salesforce’s Office of Ethical and Humane Use.
Governance via Existing Roles. Many organizations do not create a dedicated AI officer role and instead assign AI governance accountability to an existing CIO, CISO, CDO, or Chief Risk Officer. This is workable in the short term but risks AI governance being subordinated to narrower technical or risk objectives. The Board Developer analysis of CAIO roles (2025) recommends that boards define, in explicit terms, who owns each governance pillar – CAIO or equivalent for strategy, CIO or CDO for data and infrastructure, General Counsel for legal, and Internal Audit for assurance.
AI Ethics Committees: Composition and Charter
Several leading organizations have established formal AI ethics or responsible AI committees as the deliberative body for high-risk AI governance decisions. These are not advisory in the traditional sense – they have defined decision rights, review mandates, and escalation authority.
IBM’s AI Ethics Board, operating since 2018, sets policy, reviews specific use cases flagged for ethical risk, and coordinates with external standards bodies. Microsoft’s Aether (AI and Ethics in Engineering and Research) Committee – the governance body that preceded Crampton’s Office of Responsible AI – evaluated sensitive uses of AI, including the use of facial recognition, emotional state inference, and high-stakes AI applications. Google DeepMind maintains a responsible AI council as part of its frontier safety governance structure, distinct from the broader Google AI Principles framework.
For an organization establishing such a committee, a practical charter should include: membership spanning legal, privacy, risk, HR, product, and independent ethics expertise; defined quorum and voting or consensus thresholds; a register of use cases requiring mandatory review (high-risk AI systems as defined by OSFI E-23 and the EU AI Act, systems affecting hiring or credit, systems with autonomous action scope); and escalation pathways to the board risk committee for material decisions.
Centralize or Federate?
The choice between centralized and federated AI governance depends on organizational size, AI deployment volume, and the diversity of AI use cases across business units.
Centralized governance works best for organizations where AI deployment is concentrated in a small number of high-impact use cases, or where regulatory requirements (such as OSFI E-23 for federally regulated financial institutions) demand enterprise-wide consistency in model risk management.
Federated governance – where business units maintain primary AI governance accountability with central oversight – works better for large, diversified organizations with high volumes of AI use cases across distinct regulatory environments. The risk of federation is inconsistency and coverage gaps; the risk of full centralization is bottleneck and the diffusion of first-line ownership.
Most mature governance programs use a hybrid: a central AI governance function sets policy, standards, and review criteria; business units implement controls and maintain model inventories; the central function provides assurance and handles escalations. This is the model described in PwC’s responsible AI guidance as characteristic of organizations at the “embedded” governance maturity stage.
