Governing Intelligence · Part III. Putting it to work
12Leading governed AI: the executive agenda
Frameworks do not govern; leaders do. Everything in Sections 5 to 11 can be built, and still fail, if AI governance is treated as a technical program to be delegated rather than an accountability to be led. The Royal Commission into the Robodebt Scheme made that point in the hardest way available: the failures it described were failures of leadership and culture before they were failures of software. This section sets out the executive agenda for governed AI: what the accountable authority personally owns, the questions leaders should keep asking, the evidence they should insist on seeing, the external scrutiny they must be ready to answer, and the investment and workforce decisions that only they can make. Appendix J condenses it into a standing quarterly agenda that can be adopted as written.
12.1What the accountable authority personally owns
The Public Governance, Performance and Accountability Act 2013 already settles the question of who answers for AI. The accountable authority’s duty to govern the entity, promoting the proper use and management of public resources, an efficient, effective, economical and ethical operation, and the officials’ duties of care and diligence, apply to an algorithm exactly as they apply to a budget. AI does not create new duties; it gives old duties new reach and new speed. What is new is how easily accountability can appear to blur, across vendors, platforms, committees and models, and the first act of leadership is to refuse that blurring. Decisions made with AI are the agency’s decisions. Harm done at machine speed is the agency’s harm. The authority can delegate work without limit; it cannot delegate the answering.
| Only the accountable authority can | Why it cannot sit lower |
|---|---|
| Endorse the framework, the risk appetite and the boundaries | These express how much risk the agency will carry in the Crown’s name; that is an authority-level judgment, not a program setting. |
| Appoint and empower the accountable officials | The policy requires named officials; only the authority can give them standing, budget and the right to say no. |
| Set the tone on honesty and escalation | Whether people surface problems early is decided by how leaders receive bad news, and nowhere else. |
| Resource the foundation before the ambition | Data, security, provenance and capability are unglamorous; only executive sponsorship protects them from being deferred. |
| Answer publicly for the agency’s AI | To the minister, the Parliament, the auditors and the public; a delegate can brief, but the authority answers. |
12.2The questions leaders keep asking
A leader does not need to understand transformer architectures to govern AI well, any more than a chair needs to be an actuary to govern an insurer. What is needed is a short set of questions, asked on a rhythm, with consequences when the answers are weak. The table below gives one standing question per domain and the warning signs that the answer should trigger. International guidance for governing bodies, notably ISO/IEC 38507 on the governance implications of AI, converges on the same discipline: the board’s job is not to operate the system but to demand evidence that it is operating as intended.
| Domain | The standing question | Warning signs in the answer |
|---|---|---|
| Strategy | Which few AI uses matter most this year, and what did we decide not to do? | No refusals; a portfolio that only grows; ambition without a named trade-off. |
| Operating Model | Who is accountable for each use case, and would they say so unprompted? | Accountability described as shared; officials named on paper but absent from decisions. |
| Architecture | Would our AI decisions survive an audit tomorrow, from provenance alone? | Evidence that must be assembled by hand; logs that answer what but not why. |
| Design | What can each system do without asking a human, and how do we stop it? | No one can state the autonomy boundary; the kill switch has never been rehearsed. |
| Governance | What changed since last quarter, drift, overrides, incidents, and what did we change because of it? | Nothing reported; metrics all green; learning loop described but not evidenced. |
12.3Governing on evidence: the executive dashboard
Executive reporting on AI should be generated, not compiled. If the register, the gates and the runtime controls of Sections 6 to 10 are in place, then a one-page dashboard can be produced from them directly, and the numbers carry the authority of the systems that produced them. The figure below shows the pattern: coverage, assurance, exceptions and learning, each with a standing question attached so the conversation starts where the risk is.
Coverage
Use cases on the register, and impact assessments current
Is there AI we have not registered?
Assurance
Gate decisions: passed, passed with conditions, stopped
Are conditions being discharged on time?
Exceptions
Incidents, near misses, overrides and drift since last quarter
What did we change because of them?
Learning
Changes to policy, tiers, controls and training
What did the assurance loop teach us?
Two habits make the dashboard bite. First, review it with the same seriousness as financial performance, on the same cadence and in the same forum, because for a service-delivery agency AI behaviour now is operational performance. Second, treat comfortable numbers with suspicion. An override rate of zero usually means rubber-stamping, not perfection. No incidents reported usually means no one is looking, or no one feels safe saying so. The dashboard exists to provoke the second question, not to close the meeting early.
12.4Answering for AI: audit, the Ombudsman and the Parliament
A Commonwealth agency answers outward as well as up. The Australian National Audit Office now audits AI governance directly, and Section 3.8 set out what its two audits to date established: the criteria are public, they map almost line for line onto this framework’s outputs, and the difference between partly effective and largely effective is whether the machinery exists and demonstrably operates. The Commonwealth Ombudsman receives the complaints of people affected by automated and AI-assisted decisions, publishes joint better-practice expectations with the OAIC and the Attorney-General’s Department, and has shown, through live investigations of automated assessment tools, that it will examine specific systems, not just policies. Senate estimates will ask what the agency uses, what it costs and what went wrong. Freedom of information will reach AI records, and from December 2026 the Privacy Act requires automated decision-making to be disclosed in the privacy policy itself. None of this should require preparation, and that is the point: an agency running this framework is audit-ready as a by-product of operating, because the register is current, every gate decision is recorded, every in-scope decision carries provenance, and a named official can explain any decision in plain language.
12.5Investing with discipline
AI investment fails in two familiar ways: a hundred pilots that never harden into services, or a single flagship that outruns its assurance. The discipline against both is to tie money to the same gates that govern risk. Fund the secure foundation once, as infrastructure, because the landing zone, the provenance backbone and the policy-as-code runtime of Section 7 are shared by every use case that follows. Then fund use cases in tranches released at the readiness gates: discovery funding to G1, build funding to G3, scale funding only after G4 has proven behaviour against reality. Cost every use case whole-of-life, including monitoring, assurance, capability and eventual retirement, not build alone. Hold benefits to the same standard as any program under the PGPA Act: claimed in the business case, measured in operation, reported on the dashboard. Use the DTA’s 2026 guidance on taking AI from proof of concept to scale as the test for graduation, because the gap between a promising pilot and a governable service is exactly where public money disappears. And protect the clean exit in every contract, because the cheapest AI capability is rarely cheap once it cannot be left.
12.6Leading people through the change
Section 6.6 described the cultural shift from gatekeepers to participants; leaders are the ones who make it real. Three behaviours matter more than any communications plan. First, use the tools. A leader who works with the agency’s assistants weekly will govern them better than one who reads about them quarterly, and the workforce notices the difference. Second, make honesty safe and visible. The staff member who reports a near miss, an override or an ungoverned tool has done the agency a service; if the reaction they meet is blame, the framework’s feedback loops go quiet exactly where they matter most. Third, celebrate the gate that says no. A stopped deployment, publicly explained, teaches the organisation that the gates are real, and it buys credibility for every yes that follows. Capability uplift, the six emerging roles of Section 6.5 and baseline literacy for everyone, is then an investment decision, and it belongs on the executive agenda beside the technology it governs.
Facts checked against official sources on 7 October 2026. Confirm obligations against the source instruments, which can change. This framework is not legal advice.