AI Governance for Boards: The Four Questions Every Director Should Be Able to Answer

For most of the past decade, boards treated AI governance as a technology matter and delegated it entirely to the CTO or CIO. That position is no longer defensible. APRA's 2024 letter to regulated entities on AI risk management made explicit what prudential supervisors expect: boards and senior executives are accountable for how AI systems are governed, not just the outcomes they produce. The voluntary AI Safety Standard, released by the Australian federal government, reinforces that expectation for organisations outside the financial services sector.

Directors do not need to understand the mathematics of a language model. They do need to be able to ask the right questions and evaluate whether the answers they receive are credible. In our experience working with boards on AI governance, four questions separate directors who are genuinely across the issue from those who are not.

Question One: How does the organisation identify where AI is being used?

This sounds straightforward. In practice, most organisations cannot answer it with confidence. AI capabilities are embedded in software-as-a-service platforms, productivity tools, and vendor-supplied systems without always being visible to the IT or risk function. Shadow AI use by individual staff members compounds the problem. Without a reliable inventory of where AI is operating, governance of those systems is impossible.

Directors should press for a clear answer about how the AI inventory is maintained, who owns it, and how frequently it is reviewed. If the answer is "we're working on it" or "we rely on vendor disclosure", the board has identified a gap that warrants direct follow-up with management. An AI system that is not in the inventory is not being governed.

Question Two: Who is accountable for AI risk, and where does that accountability sit?

Accountability for AI risk is frequently distributed across the organisation in ways that leave gaps. The data team is accountable for model performance. The legal team manages AI-related contractual obligations. Security owns the controls around AI infrastructure. But when an AI system produces an adverse outcome, the question of who was accountable for preventing it often has no clear answer.

ISO 42001, the international standard for AI management systems, addresses this directly by requiring organisations to define roles, responsibilities, and authorities for AI governance. The APRA letter similarly expects that accountability structures be clearly defined and documented. Directors should be able to confirm that a named individual or function holds end-to-end accountability for each material AI system, not just accountability for individual components of it.

Question Three: How are AI systems tested before deployment and during operation?

Testing an AI system is not the same as testing conventional software. AI systems can perform well on standard test sets and still fail in production when inputs drift, when adversarial content is introduced, or when the system is used in ways that differ from its design intent. Pre-deployment testing needs to include adversarial testing, fairness and bias assessment, and security review specific to AI attack vectors such as prompt injection and data extraction.

Ongoing monitoring matters as much as pre-deployment testing. AI systems degrade over time as the data they were trained on diverges from current conditions. Directors should ask what monitoring is in place for each material AI system, what thresholds trigger a review, and when the last review occurred. If the answer covers pre-deployment testing but not ongoing monitoring, the governance picture is incomplete.

Question Four: What is the organisation's exposure if an AI system fails or is misused?

This is a risk quantification question, and it connects AI governance to the board's existing risk appetite framework. The potential consequences of AI system failure vary significantly depending on what the system does. An AI system used in credit decisioning carries different exposure than one used to draft internal communications. Directors need to understand the consequence profile of each material AI system and confirm that the controls in place are proportionate to that exposure.

Regulatory exposure is part of this picture. APRA entities that cannot demonstrate sound AI governance may face supervisory attention. Organisations that deploy AI in consumer-facing contexts carry reputational exposure if a system produces discriminatory, inaccurate, or harmful outputs. The board's risk committee should be satisfied that AI risks are captured in the enterprise risk register at an appropriate level of granularity, not just as a single line item.

If your board is working through AI governance for the first time, or if you want an independent assessment of where your governance stands against the APRA letter and the Voluntary AI Safety Standard, contact us at info@cyberlinx.com.au. We work with boards and senior leadership teams across a range of sectors on practical AI governance frameworks.