AI Governance

Overview

What is AI governance?

AI governance is the organisational arrangement that makes someone accountable for each AI system in use, and that ensures the risks of that system are identified, treated, monitored and recorded. It covers who owns an AI use case, what the organisation permits and prohibits, how AI risk is assessed, how humans stay in control of consequential decisions, and what evidence exists to demonstrate all of that after the fact.

It is distinct from AI development and from data science. An organisation does not need to build models to need AI governance. Most Australian organisations are deployers: they buy or subscribe to AI capability built by someone else, embed it in a process, and inherit the risk without inheriting the visibility. That gap between where the risk lands and where the knowledge sits is the governance problem.

AI governance is also not a parallel framework. The organisations that manage it well fold it into the risk, security and resilience governance they already have, rather than standing up a separate regime that competes with it for attention.

The standards

Which standards apply to AI governance in Australia?

Three standards do most of the work, and they are designed to interlock.

The practical significance of this structure is that AI governance is not a greenfield build for most organisations. AS ISO/IEC 42001 follows the same harmonised management system structure as other ISO management system standards, so it slots alongside existing arrangements. ISO/IEC 23894 is explicitly written to be used in connection with ISO 31000, which means an organisation with a functioning risk framework already owns the foundation.^{1, 2}

Agilient does not certify organisations against AS ISO/IEC 42001 and is not a certification body. Certification is a separate activity performed by an accredited certification body. What an advisory firm can do is establish the governance, assess the gap, build the controls and evidence, and prepare an organisation for whatever assurance it chooses to pursue. See the security risk management pillar for how the underlying risk method works.

Government guidance

What does the Australian Government expect?

The National AI Centre published Guidance for AI Adoption on 21 October 2025. It sets out six essential practices for safe and responsible AI, and it comes in two versions: a foundations version for organisations early in their AI use, and an implementation version for organisations that build or customise AI, use it in complex ways, or manage higher-risk use cases.³

The sixth practice, maintaining human control, is the one that runs through all the others rather than sitting beside them.

The guidance is voluntary and creates no new legal duties. That is frequently misread as meaning it does not matter. It matters for two reasons. It is the clearest available statement of what the Australian Government considers reasonable practice, which makes it the benchmark an organisation is likely to be measured against if something goes wrong. And it aligns with AS ISO/IEC 42001, so adopting it is not wasted work if formal certification is later pursued.

The practices also map to Australia’s AI Ethics Principles: human, societal and environmental wellbeing; human-centred values; fairness; privacy protection and security; reliability and safety; transparency and explainability; contestability; and accountability.⁴

What changed

What happened to the Voluntary AI Safety Standard?

The Voluntary AI Safety Standard set out 10 guardrails for safe and responsible AI. It has been superseded. On 21 October 2025 the National AI Centre published Guidance for AI Adoption, which the Department of Industry, Science and Resources describes as updated and simplified guidance that evolves the Voluntary AI Safety Standard.^{3, 5}

This matters more than a renaming usually would. Many Australian organisations built their first AI policy against the 10 guardrails during 2024 and 2025, and a policy that still cites the superseded standard as its authority is now pointing at the wrong document. The underlying expectations have not reversed, and an organisation that genuinely adopted the guardrails is not starting again. But the mapping should be refreshed, and the six essential practices are the current reference point.

The direction of travel is worth noting when setting AI policy. The guardrails were updated in 2024 to align more closely with the Government’s proposed mandatory guardrails for high-risk AI. Organisations building AI governance now are better served designing to a standard, AS ISO/IEC 42001, than to guidance that is still moving.

Sector obligations

Who has obligations beyond the general guidance?

General guidance is the floor. Several sectors carry obligations that reach AI use directly, and these are where AI governance stops being optional.

• Australian Government entities. The Protective Security Policy Framework (PSPF) carries policy advice on the use of generative AI with OFFICIAL information, and more recent advice addressing frontier AI and access to generative AI capability. Entities also carry technology and information security requirements that AI deployments engage directly.
• Critical infrastructure. Responsible entities under the Security of Critical Infrastructure Act must manage material risks to their critical infrastructure assets under a critical infrastructure risk management program. Where AI is embedded in an operational process, it is part of that hazard picture, not separate from it.
• Regulated data. Privacy obligations attach to personal information regardless of whether a model is involved. AI use cases that ingest personal information, or generate inferences about people, engage those obligations directly.
• State government. Several jurisdictions run their own AI assurance requirements for agencies, which apply in addition to any national guidance.
• Health and hospitals. Clinical and administrative AI carry very different risk profiles, and both touch sensitive information.
• Transport and logistics. Automation decisions carry safety consequences, so human oversight and testing evidence matter more than elsewhere.
• Local government. Small teams carry the same obligations as large ones with a fraction of the capability, which makes scoping and sampling decisive.

Where an organisation already runs a security risk assessment cycle, the efficient path is to bring AI use cases into that cycle rather than to run a separate AI risk process alongside it.

When to act

When does an organisation need AI governance advice?

The request usually arrives from one of a small number of directions. Any one of them is a reasonable trigger; together they usually mean the gap is structural rather than incidental.

• The board has asked a question no one can answer. Typically some version of “where are we using AI, and who is accountable for it?” The inventory does not exist, and the answer is assembled by email.
• AI arrived without a decision being made. Capability appeared inside tools the organisation already licensed. No one procured “AI”, so no one governed it.
• A regulator, customer or insurer has asked. AI governance questions increasingly appear in tender responses, supplier assurance questionnaires and renewals.
• A use case has moved from experiment to consequence. An internal drafting aid is one risk profile. AI touching a hiring, eligibility, safety or customer decision is another entirely.
• Certification is being considered. An organisation intending to pursue AS ISO/IEC 42001 certification needs the management system built and the gap closed before a certification body arrives.
• Something went wrong. An incident, a near miss, or an output that should never have reached a customer.

Where the trigger is a live obligation rather than curiosity, the practical starting point is an AI governance maturity assessment, which establishes what is actually in place before anything is committed.

Approach

How to build AI governance that holds up

The failure mode is not a lack of policy. It is a policy no one can evidence. An AI governance arrangement is only worth what an organisation can demonstrate when a board, an auditor or a regulator asks. Agilient works the problem in five steps.

The third step carries most of the value and is the one most often skipped. Guidance for AI Adoption makes the same point: the same tool can create very different risks depending on how it is used, so each use case is reviewed on its own.³

How we help

How Agilient supports AI governance

Agilient is an independent security, risk and resilience consultancy. It advises on the governance of AI; it does not develop, sell or implement AI systems. That independence is the point: an organisation asking whether a control is adequate should not be asking the party that sold it.

How the work is scoped

The delivery model is deliberately lean, because AI governance engagements fail when they become a discovery exercise that consumes the organisation’s time and returns a document no one uses.

• Scoped and desktop-first. An evidence-based review of what already exists, before anyone is asked to sit in a workshop.
• One structured consultation per stream, with targeted interviews only where a gap genuinely requires one.
• High-impact use cases are sampled rather than attempting whole-of-organisation coverage.
• Scope, exclusions and dependencies are stated up front, so the engagement does not drift.
• Deliverables are written to be tabled, not translated: audit-ready and board-ready as produced.

AI failure is a resilience event as much as a compliance one. Where an AI system sits in a process that matters, the response belongs in the arrangements that already exist for disruption: see business continuity and business resilience programs. Read more about the full service line on the AI governance consulting page.