On 29 June, the AICD and the Human Technology Institute at UTS released Version 2 of A Director's Guide to AI Governance, updating the 2024 suite to reflect the arrival of agentic AI — systems that act inside a business, not just answer questions in it. There is a companion webinar on 30 July, and the full report sits behind the member wall, so the precise line-by-line changes are not yet public. But the headline is clear enough, and it is the right update. The reference document Australian directors will be measured against just moved.
Here is where most boards will get it wrong. A peak body updates its guide, and the reflex is to circulate it, note it in the minutes, and feel a little more covered than we did last week. That instinct is the trap. A guide is not governance. The AICD can hand you the best scaffolding in the country and your board can still be governing AI badly — because the document tells you what to think about, not whether your actual oversight holds up when an agent misfires at 2am on a Sunday. The update is a prompt to test your governance, not a compliance artefact to file.
And the timing has done us a favour by supplying a live worked example. In the very same fortnight the guide landed, the US Commerce Department switched off — then, on 1 July, switched back on — global access to Anthropic's frontier model, in exchange for the vendor agreeing to future "security protocols"; OpenAI gated its newest family to roughly twenty government-approved companies; and Malcolm Turnbull argued Australia should be demanding access to frontier models rather than deferring to US export controls. Read that as a director and the abstraction disappears. Model access is no longer a procurement detail. It is a government-negotiated variable that can be turned off — which means any board relying on a single frontier vendor is now carrying a continuity risk it never actually underwrote. The guide tells you to govern emerging AI risk. The sovereignty story is that risk, arriving on schedule.
- We have received the updated AICD guide. Rather than noting it — against each of its recommendations, can management show me where our current practice already complies, and where it does not? I want the gap list, not the summary.
- The update exists because AI can now act autonomously inside our business. For every agentic tool we have deployed, who is the named accountable person when the agent does something we did not intend — and is that the same person who signed off deploying it?
- If our primary AI vendor's model became unavailable to us next week — by export control, price, or policy — which of our processes stop working, and what is our fallback? Can anyone in this room answer that today?
- Do we have a written AI risk appetite that says yes or no to a specific deployment, or do we have a values statement about being "responsible" with AI? If I asked for the language that would decline a proposal, could you point to it?
- What leading indicators would tell us an AI deployment is drifting wrong before it becomes a customer complaint or a regulator's letter — and are any of them on a board dashboard?
- The 10 December automated-decision-making transparency deadline is dated and real. Is our policy and contract remediation underway now, or are we planning to discover the gap in November?
- The guide gets circulated and minuted, and no one is tasked with producing a gap analysis against it. Acknowledgement masquerading as oversight is the oldest failure in governance.
- Agentic AI being governed under the same procurement framework built for SaaS subscriptions. An agent that executes work is a fundamentally different risk object than a licence for a productivity suite, and the board-management accountability line blurs the moment software starts acting on its own.
- A single-frontier-vendor AI strategy with no articulated fallback — now demonstrably a continuity exposure, not a tidy simplification. Even ASX-listed software is deliberately going multi-model this same fortnight. Single-vendor is looking like the outlier.
- "We are waiting for the regulation to settle" offered as a governance posture. The Australian direction is visible and partly dated. Waiting is a decision, and it is the wrong one.
The upside is that this is cheap, unglamorous work with an outsized payoff. A written AI risk appetite, a register of every AI and agentic deployment across the business, a named accountable owner for each, a short set of leading indicators on the board dashboard, and an explicit multi-vendor position — none of that is exotic, and most of it is oversight the board should already be doing. Do it against the new guide as your checklist and you are not chasing a moving regulatory target; you are building the substance that meets whatever framework arrives, while your competitors are still forwarding the PDF around the audit committee.
The downside is the asymmetry, and it is stark. Treating the guide as a document to file rather than a diagnostic to run costs you almost nothing today and a great deal precisely when you can least absorb it — when an agent acts outside its intended scope, when a frontier model you had built a core process around becomes unavailable, when the December deadline lands and the contracts have not been touched. Those exposures do not announce themselves through audit findings; they arrive as headlines and regulator correspondence. There is no version of this where getting ahead of it is wasted effort, and none where being caught flat-footed is cheap.
Do not let this guide become the best-circulated, least-actioned document on your board's shared drive. Put it on the next agenda as a working session, not an information item, and task management with one deliverable: a gap analysis of current practice against every recommendation in Version 2, with owners and dates attached to the gaps. In the same session, treat the sovereignty story as your case study and pressure-test one thing concretely — if your primary model vanished next week, what breaks and what is the fallback — because that single question exposes whether your AI governance is real or decorative. Insist on a written risk appetite, a deployment register with named accountability including for agentic tools, leading indicators on the board dashboard, an explicit multi-model position, and a live plan for the 10 December automated-decision-making deadline. The AICD has given every Australian director the scaffolding. Whether it becomes governance is the board's job, not theirs — and a director who will not do that work is holding a guide, not governing.
Researched and drafted by Brad's agentic AI team. Edited and published by Brad Ferris.