Governing AI agents: a practical framework for ANZ boards

Australian and New Zealand boards don't need a new AI law to govern AI agents. Directors' duties, the Privacy Act and, for banks, insurers and super funds, APRA's prudential framework already apply. In practice, governing agents comes down to five things the board should be able to see: a register of every agent, an autonomy tier for each one, a named human owner, limits on what each agent can access, and monitoring with a tested way to switch the agent off.

Board members around a conference table

This guide sets out what regulators and director bodies in Australia and New Zealand now expect, a tiered framework for approving agents, and the questions a board should ask management each quarter. It is general information, not legal advice.

Why do AI agents need different governance from other AI?

A chatbot answers questions. An agent takes actions: it plans a task across several steps, calls other systems, and changes records, sends messages or approves requests. The AICD and the Human Technology Institute at UTS describe agentic AI as generative AI combined with other tools so it can plan and act across multiple steps towards a goal, with varying degrees of autonomy. Version 2 of their A Director's Guide to AI Governance (June 2026) is the most current board-level reference in Australia.

Adoption has already started. In HTI's Corporate Leaders Survey, cited in the guide, 32.5% of organisations reported at least one use of agentic AI and another 12.9% planned to adopt it. IT and security, marketing and sales, and back-office work were the most common areas.

The guide names three ways agents raise the risk compared with other AI:

  • Unintended actions. An agent can do something its owner never authorised because its instructions were unclear, it misread its goal, or it made an error partway through a task.
  • Orchestration. Agents that work across several systems, or with other agents, widen the attack surface. An attacker can target the agent or any system it can reach.
  • Scale and continuity. An agent running around the clock can repeat an error thousands of times before anyone notices.

Liability doesn't move to the software. The guide's position is that agents are IT systems, not employees or contractors acting for the organisation. It points to the Air Canada case, where a Canadian tribunal rejected the airline's argument that it wasn't responsible for what its website chatbot told a customer.

What do Australian and New Zealand regulators already expect?

Australia has chosen not to legislate for AI separately. The AICD guide records that the December 2025 National AI Plan confirmed Australia won't introduce standalone AI legislation. Existing laws will apply instead, with targeted reforms where gaps emerge. The National AI Centre also reminds directors that their duty of care and diligence under section 180 of the Corporations Act includes making sure adequate governance exists for AI risks (Essential AI practices). The sources that matter for agents are below.

Source Date Binding? What it asks for
APRA letter to industry on AI 30 Apr 2026 Supervisory expectations for APRA-regulated entities Board AI literacy, an inventory of AI tools and use cases, human involvement in high-risk decisions, identity controls for agents, AI supply-chain mapping, continuous assurance
Privacy Act, APP 1.7 to 1.9 In force 10 Dec 2026 Yes, for APP entities Privacy policies must disclose the kinds of decisions that computer programs make or substantially support, and the personal information they use
Guidance for AI Adoption (National AI Centre) Oct 2025; implementation guidance published 5 May 2026 Voluntary Six practices: decide who is accountable, understand impacts, measure and manage risks, share essential information, test and monitor, maintain human control
AICD and HTI, A Director's Guide to AI Governance (v2) Jun 2026 Voluntary Board risk appetite for AI, an AI register, and agent controls covering permitted tasks, access limits and logging
AS ISO/IEC 42001:2023 Adopted 16 Feb 2024 Voluntary, certifiable An AI management system: policy, roles, risk assessment, controls and continual improvement
Institute of Directors (NZ): A Director's Guide to AI Board Governance Jul 2024 Voluntary Nine principles, including categorising and addressing risks, building board capability, and overseeing AI use and data governance

The APRA letter is the most specific about agents. It found that identity and access management "has not yet adjusted to nonhuman actors such as AI agents". It expects controls over agentic and autonomous workflows, and second-line and audit teams that can independently assess agentic workflows. It also says point-in-time, sample-based assurance doesn't suit probabilistic systems, and expects continuous monitoring scaled to how critical each use case is. The letter is addressed to APRA-regulated entities. Its expectations are still a sound benchmark for any board, because they describe what a regulator will look for after something goes wrong.

The Privacy Act change matters because many agents sit inside decisions about people: approving refunds, triaging complaints, screening applicants. From 10 December 2026, those uses need to be disclosed, and a person approving the agent's recommendation doesn't remove the duty. Our guide to automated decisions and the Privacy Act covers what the disclosure needs to say.

What does a practical agent governance framework look like?

The hard question for a board is not "do we allow agents?" but "who can approve which agent?". We sort agents into four autonomy tiers and attach the approver and minimum controls to the tier. This is our own working framework, built from the AICD controls, the NAIC practices and the APRA expectations above. Adjust the thresholds to your own risk appetite.

Tier What the agent does Example Who approves it Minimum controls
0. Assist Drafts or summarises; a person sends or acts Drafting a reply to a service ticket Business owner Register entry; acceptable-use policy; read-only access
1. Act with approval Prepares an action; a person approves each one Raising a purchase order for sign-off Business owner and risk Its own identity; scoped write access; action log
2. Act within limits Acts alone inside set limits; people review samples and exceptions Closing duplicate tickets; issuing refunds under a dollar limit Executive committee Hard limits in the target system; daily exception report; tested kill switch
3. Significant decisions Acts alone, or is the key input, on decisions that significantly affect people, money or operations Declining credit; suspending an account; changing production systems Board or board risk committee All of the above, plus an impact assessment, Privacy Act disclosure check, independent testing and continuous monitoring

Two rules keep the tiers honest. First, tier the agent by the most consequential action it is technically able to take, not the one it was built for. An agent with write access to the CRM is a tier 2 agent even if it was designed to draft emails. Second, re-tier the agent whenever its tools, data or vendor model change. The NAIC implementation guidance also asks organisations to define in advance when an AI system should be shut down and who decides. Settle that for every tier 2 and 3 agent before it goes live.

What should go in the agent register?

Practice 4 of the NAIC guidance asks for an organisation-wide AI register. Its fields include accountable people, purpose, capabilities and limitations, datasets, test results, risk assessments and review dates. The National AI Centre publishes a template. For agents, add six fields the standard template doesn't emphasise:

  1. Autonomy tier, using the table above, with the date it was last reviewed.
  2. Systems and tools the agent can call, with the permission level in each (read, write, approve, pay).
  3. Identity it runs as. This should be the agent's own service identity, never a shared or personal login.
  4. Hard limits, such as dollar thresholds, record counts per day, and actions it may never take.
  5. Model and vendor chain, including the underlying model provider. APRA calls these fourth-party dependencies.
  6. Decisions about people: whether the agent makes or substantially supports one, and whether the privacy policy covers it.

Agents built into vendor platforms belong in the register too. The AICD guide warns that AI is increasingly embedded in third-party products and used on a shadow basis without oversight. Atlassian Rovo, Salesforce Agentforce and Microsoft Copilot Studio all let business users create agents. If your licences include them, ask management how many agents exist today and who created them. The answer usually shows how complete the register is.

How do we control what an agent can access?

Agents fail in the same places as any other system: access, patching, authentication and incident response. The difference is speed. In our cyber security review for an Australian asset manager, the critical findings were inadequate access controls, outdated software, weak authentication without MFA on critical systems, and an incomplete incident response plan. Each of those gaps is more dangerous when an agent can act on it continuously. Four controls deal with most of the exposure:

  • Least privilege, per agent. Give each agent the narrowest access its tier allows, enforced in the target system rather than in the agent's instructions. A prompt that says "never issue refunds over $500" is not a control. A refund limit in the finance system is.
  • Separate identities. Every agent gets its own credentials, so its actions can be traced, limited and revoked without touching anyone else's access.
  • Logged actions. The AICD guide lists logging and auditability of agent actions as a core control. Logs should capture what the agent did and why, in a form your audit team can review.
  • Integration security. APRA names prompt injection, insecure integrations and manipulation of autonomous agents as attack paths. Every connector an agent uses is an integration, and should be designed and tested like one.

Most of this work is integration and identity design, not AI. It is where our agentic AI systems integration practice spends most of its time. For a wider view of how agents connect CRM, service, marketing and ERP platforms, see our guide to integrating the connected enterprise.

What should the board ask management each quarter?

The AICD guide doesn't expect directors to become technical experts. It does expect them to ask the right questions and get assurance. The APRA letter adds that many boards still lack the literacy to challenge AI risk effectively. These eight questions fit on one page of a board pack:

  1. How many agents are in production, by tier, and how many were added since last quarter?
  2. Which agents moved up a tier, and who approved the change?
  3. How many agent actions were reversed, overridden or escalated, and what caused them?
  4. Were there any incidents involving an agent, including near misses?
  5. Which vendors or underlying models changed, and were the affected agents retested?
  6. Which agents make or support decisions about people, and does our privacy policy reflect them?
  7. What did agents cost this quarter, including usage-based model and platform charges, against the value they were approved to deliver?
  8. When did we last test the kill switch on a tier 2 or 3 agent, and how long did it take?

Question 7 is often missed. The AICD guide notes growing concern among Australian business leaders in 2026 about the rising cost of agentic AI, at a time when the return on AI investment is still uncertain. Agents that run continuously consume tokens and platform credits continuously. Treat cost per completed task as a governance metric, not just a finance one. Our 40 agentic AI use cases by department include a scorecard for judging value before an agent is approved.

Do we need ISO 42001 certification?

Usually not as a first step. AS ISO/IEC 42001:2023 is a certifiable management system standard, adopted in Australia in February 2024. It gives you a structure of policy, roles, risk assessment, controls and continual improvement that maps well onto the six NAIC practices. Use it as the skeleton for your AI governance framework from day one. Certification is worth pursuing when customers, government tenders or regulators start asking for evidence, or when you already hold ISO/IEC 27001 and can extend the same audit cycle.

Frequently asked questions

Are directors personally exposed if an agent causes harm?

Directors' existing duties apply, including the duty of care and diligence under section 180 of the Corporations Act. The AICD guide treats board oversight of AI as part of those existing duties. A board that can show a register, tiered approvals and regular reporting is in a far stronger position than one that can't.

Does keeping a human in the loop solve the governance problem?

Only if the human has the time, information and authority to say no. A person who approves hundreds of agent recommendations a day without reviewing them is not real oversight. That arrangement also doesn't remove the Privacy Act disclosure duty.

Is New Zealand any different?

The approach is similar. New Zealand also relies on existing law rather than a standalone AI act. The Institute of Directors' nine principles cover the same ground as the AICD guide: strategy, risk, board capability, and oversight of AI use and data. A framework built on the tiers and register above works on both sides of the Tasman.

Where to start

Start with the register. Ask management for a list of every agent in production or pilot, including those built inside vendor platforms. Assign each one a tier, then compare its actual access against what that tier allows. The gaps usually show where the first controls are needed. Our agentic AI team can help you build the register, tier your agents and design the identity and integration controls behind them. Contact us to talk it through.

Let's work together.

Turn insight into action with a team that delivers.