If you sit on a board or an executive team, six authoritative Australian sources have now sent you the same message about AI: prove your governance from the system itself, not just through static verification. The era of relying solely on policy documents is ending. Regulators now demand dynamic controls within your enterprise.
Six beats, one message
Warning. In October 2024 ASIC reviewed 23 financial services and credit licensees. Report 798 was pointedly titled Beware the Gap. Nearly half had no policies addressing consumer fairness or algorithmic bias, and fewer still told customers when AI was being used.1
Demand. Through a targeted engagement in late 2025, APRA reviewed selected banks, insurers and superannuation trustees and found governance, risk management, assurance and operational resilience were not keeping pace with AI adoption.2 On 30 April 2026 it demanded a step-change in AI risk management, tying its expectations to the existing prudential standards for information security, operational risk, governance and data risk. Your current obligations already cover AI, and APRA expects you to act like it.3
Bedside. In August 2025 the Australian Commission on Safety and Quality in Health Care released its AI Clinical Use Guide: practical checkpoints for clinicians before, while and after AI is used in care. The standards body for hospital accreditations expects AI to be governed at the bedside, not just in the boardroom.4
House rules. The Digital Transformation Agency's updated policy for responsible AI use in government came into force in December 2025, with requirements phasing in progressively until December 2026. It requires a register of in-scope use cases, accountable owners, and AI impact assessments before deployment.5
Playbook. In June 2026 the AICD and UTS's Human Technology Institute released version two of A Director's Guide to AI Governance. Boards now have the manual. No excuses remain.6
Machinery. On 15 July 2026 the Australian Government stood up an Office of AI inside the Department of the Prime Minister and Cabinet to coordinate national AI standards. AI has stopped being merely a technology topic and become a whole-of-nation one.7
Almost everything written in response to that drumbeat is policy: frameworks, committees, registers, attestation templates. All necessary, but they are the paperwork of governance rather than the substance of it.
Throughout 2026 I have been governing an AI system I built myself. A retrieval pipeline over my working corpus, local and cloud models, and agents that draft briefings overnight. Everything I now advise was road-tested on my own systems first, mistakes included.
Governance requires data routing
Governance must extend further than a policy paragraph about the appropriate use of third-party models. A feature of my stack is that it can call fast local models for personalisation; sovereign, auditable cloud endpoints for tougher confidential client work; and cheap international APIs that are neither private nor sovereign but perfect for my own non-confidential deep reasoning. This flexibility makes it convenient, powerful, economic, future-proof. Unfortunately, AIs leak information. Despite multiple strong LLM prompts to respect secrets, in one intensive build I counted eight API security secrets exposed across hundreds of working turns. The AI had been specifically asked NOT to read the secrets file, but it read it anyway, and the exposure was only caught when I built a watch for this. So I wrote a test harness that investigated multiple frontier cloud and local open models. I found instances of the AI reading confidential items outside its scope across every model I tested.
What if confidential client data was found by your AI and then included in training data? Policy and prompt based controls are aspirational. Architecture boundaries are absolute.
I now have audit trails and software gates on all data flows. Within retrieval, deterministic checks make it impossible for sensitive content to reach the wrong endpoint. Secure partitions make data unreachable by AI. Now I ask: if this leaked, what would it cost? If I don't like the answer, that data stays isolated.
My lesson. Control where your data can go, as well as what is reachable.
Trust must be earned on a ladder of verification
Before any agent touches anything that matters it must earn its stripes on tasks where a mistake is non-consequential. On one task I blew a months token allowance, and ran through a further $1,200 usage charges, on a weekend build that was confident it would finish but ultimately didn't work. On another task I ran three days of continuous design loops, perhaps twenty redrafts requiring human direction, to reach a buildable design. On another occasion when I swapped in a more powerful frontier model, in ten minutes it refactored my system stripping out hardened controls. These were trust building on my own dollar before using AI on client work.
Agentic AI's are powerfully smart and fast. They do not possess common sense. They have no experience with consequences. Strong design reviews are a given. Verification must step in quickly.
I now enforce a strict design method on my Agentic AI coder. It must produce verifiable written designs, definitions of done, written test cases. I have refined check my AI design agent, AI test manager, and AI code reviewer all run before any design proceeds to human approval. After build this pattern is repeated on test and deploy. LLMs must pass a test harness now before use. I pilot everything on smaller tasks and some items get abandoned. My agentic designs regularly include deterministic software test suites that run daily alongside the operational AIs.
My lesson. AIs must earn trust by stepping slowly up a ladder of verification.
Governance requires intentional decisions on where to place humans
I write a brief, agents draft and cite, another agent checks sources. I then verify that citations resolve and claims holds before anything reaches another person. I've had less strongly directed AI's write briefs that were too strict, risky, or alarmist. Multiple models have invented plausible sources, or improvised for hours, only admitting the ruse when caught. My experience has been that these hallucinations occur more frequently when I skimp on prompting or don't ground in verified data.
Artificial intelligence does not behave like traditional software. With decades of leading technology teams I have developed a radar for how that works. For AI I've had to leverage this into new skills. Managing AI is less about programming logic and more about steering intent, much like leading a group of talented but independent thinkers who require goals, context, and continuous alignment with organisational values. When I positioned myself as an active leader within my AI processes I've gotten better results and better alignment with the ethical or risk outcomes I'm pursuing.
When I've let AI work from memory or without strong collaboration, it has mostly failed, often in hard to find ways. The moment unchecked AI output reaches a client to save time, that speed costs you magnified risk.
My lesson. Embedded human judgement must lead AI loops.
Routing. Verification. Human judgement.
Read the six beats again. APRA's data risk and information security expectations are about routing. ASIC's concerns about undisclosed AI and unassessed bias, and the Commission's bedside checkpoints, are about where the human leads. The DTA's accountable owner, register and impact assessment are a verification ladder.
The regulators are not asking you to admire your policy. They are demanding that you inspect your plumbing.
Three short questions for your next AI paper
Show me the route. For each class of data, which systems and models can it physically reach, and what enforces that? Architecture? Policy and AI prompting are not enough.
Show me the verification. What has this system been trusted with, what did it get wrong, and what earned each step up? How does verification step in when required?
Show me the human. Does human judgement still lead? Who is the person who owns outcomes before they become consequential? If humans are reduced to quality control, who will bear the cost of inevitable failures?
If those cannot be answered from your system itself, evidenced in dynamically checked business processes, governance has not happened yet whatever the documentation says.
The gap worth closing
There may be a gap between professions. Governance specialists might not have run the infrastructure, engineers might not have carried governance accountability, and leaders might not have gotten hands on with AI agents. The organisations that close this gap will put people at the intersection: close enough to the plumbing to know the risks, senior enough to be accountable for them.
AI Governance requires Engineering. Without this, what is produced is stationery.