What a board should ask about AI risk before a rollout

An AI rollout usually starts as a procurement decision or a product build, and the paper that reaches the board describes the tool. The exposure sits in what the organisation answers for afterwards. A system that reads customer records, drafts communications or acts on a person's behalf produces actions nobody approved individually. That puts it on the same footing as any other material business risk the board oversees.

The Voluntary AI Safety Standard states that leaders cannot delegate or outsource accountability for the safe and responsible deployment and use of AI systems. The Guidance for AI Adoption turns that into a first action: assign a senior leader as the overall AI governance owner. That leader needs enough authority and understanding of AI capability and risk to oversee all AI use. It then asks for a specific person to be made accountable for every AI system the organisation uses. ISO/IEC 42001 clause 5.1 holds top management itself responsible for the AI policy and objectives, and for their fit with the organisation's strategic direction. It also holds top management responsible for integrating the management system into ordinary business processes and for the resources it needs. Clause 5.3 requires top management to assign responsibility and authority for two things: that the AI management system conforms to the standard, and that its performance is reported to top management. Put a name against each in the board papers.

The AI project underway has a name and a sponsor. The rest of an organisation's AI use often has neither, because staff adopt AI tools individually and vendors switch AI features on inside products the organisation already licenses. The Guidance for AI Adoption asks for an AI register: an up-to-date, organisation-wide inventory of each AI model and system. It is held in enough detail to inform stakeholders and to support a future conformance assessment. A register that lists only the sanctioned project misses the tools staff picked up on their own. It also misses the features that arrived in a product update, and any AI agent given standing access to internal systems.

A register entry that says AI risk names nothing anyone can act on. The third essential practice asks for a risk screening process that flags the systems and use cases carrying unacceptable risk or needing extra governance attention. It then asks for risk assessments and mitigation plans for each specific use case. The Voluntary AI Safety Standard sets the same requirement at system level, calling for a documented risk assessment for each AI system, including systems procured from third parties. Each assessment is written against documented use cases and their potential unintended use. A workable entry names the system, states the way it can fail in business terms, and names who answers for the consequence.

The Voluntary AI Safety Standard sets record keeping at the level a third party needs to assess compliance with the guardrails. That is a harder bar than an internal comfort document. Under ISO/IEC 42001 clause 5.3, reporting the AI management system's performance to top management is an assigned responsibility. Set the cadence and the detail so the board can see movement between quarters.

A quarterly pack built to that bar carries the current AI register against the previous one, so the board can see what was added and what was retired. It carries the gaps that remain against each practice a system has not met. It also carries every occasion on which a person intervened and what followed, and the incidents and near misses logged against AI systems, kept separate from general IT incidents. Where a gap has been accepted instead of closed, the acceptance needs a named owner, a compensating step and a date it comes back for review.

Two further items belong in the pack. The fifth essential practice calls for documented deployment authorisation and rationale from the accountable person, based on test results. The pack should show who authorised each system to go live and on what evidence. The sixth calls for alternative pathways, so that critical functions continue if an AI system malfunctions or is retired, which is a question to settle before the rollout.

The Corporations Act 2001 (Cth) contains no AI duty. It contains section 180(1), which requires a director or other officer to exercise their powers and discharge their duties with the care and diligence a reasonable person would exercise if they held that office, with those responsibilities, in that company's circumstances. The National AI Centre puts that duty in AI terms. A company director has a personal duty under section 180 to act with care and diligence. That duty includes ensuring adequate governance systems exist to manage the risks created by the organisation's AI systems.

The nearest Australian authority on what adequate means sits in cyber security. In Australian Securities and Investments Commission v RI Advice Group Pty Ltd [2022] FCA 496, the Federal Court declared that RI Advice contravened the Corporations Act, sections 912A(1)(a) and 912A(1)(h). The contravention was failing to have cyber security documentation and controls adequate to manage risk across the network operating under its licence. Rofe J found that it is not possible to reduce cyber security risk to zero. It can be materially reduced to an acceptable level through adequate documentation and controls. The judgment also states how adequacy is to be assessed. Assessing any particular set of cyber risk management systems requires the technical expertise of a relevantly skilled person. The standard is one for the Court to decide, but the Court's assessment would be informed by evidence from relevantly qualified experts in the field.

Those findings were made about a financial services licensee, and about cyber security rather than AI. What they call for is evidence of a particular kind. That evidence is a documented AI register, a risk assessment for each system, and oversight points examined by someone outside the team that built them.

Start with the register, because none of the other artefacts can be produced without it: every AI system in live use, named, owned and dated. Then confirm the intervention point for each one and test that it stops something. Then check that the two ISO/IEC 42001 roles, conformance and performance reporting, sit with named people instead of being implied by a job title. A gap analysis produced before the register exists will describe the wrong systems.

Where none of that exists yet, the work runs as an ISO 42001 readiness assessment. That assessment is a gap analysis against the standard, scoped to the AI systems actually in use. It produces a remediation plan sequenced against the date driving the work, and an evidence set prepared for a certification audit. It is led by a director who holds the Exemplar Global Certified Lead Auditor credential for ISO/IEC 27001:2022. That work runs from the engineering practice that builds and runs the firm's own production AI systems on Azure. Our AI cyber defence work covers the register, the intervention points and the assessments that sit behind the quarterly pack.