---
title: What does Australia's Guidance for AI Adoption expect for AI agents?
url: https://opencontroller.lyzr.com/blog/australia-guidance-for-ai-adoption-and-agents
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# What does Australia's Guidance for AI Adoption expect for AI agents?

**Australia's Guidance for AI Adoption, published by the National AI Centre on 21 October 2025, replaces the 2024 Voluntary AI Safety Standard and condenses its 10 guardrails into six essential practices: decide who is accountable, understand impacts, measure and manage risks, share essential information, test and monitor, and maintain human control. It is government guidance, not legislation, but it is the clearest statement of what Australian regulators and customers expect from organisations deploying AI agents.**

The guidance does not single out agents. It applies to any AI system. That makes the job for enterprise teams one of translation: what does each practice mean when the AI system can plan, call tools and act on someone's behalf?

## What is the Guidance for AI Adoption?

The guidance comes in two layers. A foundational version is aimed at organisations starting out with AI. An extended set of implementation practices is aimed at organisations scaling AI and at technical teams. Both are built around the same six practices. The Department of Industry, Science and Resources describes the change as streamlining the 10 guardrails of the earlier standard down to six key practices.

Because it is guidance, there is no certification and no penalty for ignoring it. Its weight comes from elsewhere: sector regulators, boards and procurement teams use it as a reference point, and organisations in regulated sectors still have binding obligations, such as APRA's CPS 230 on operational risk for financial institutions.

## How do the six practices map to AI agents?

| Practice | What it asks | What it means for agents | Evidence to keep |
|---|---|---|---|
| 1. Decide who is accountable | Clear ownership of AI systems and outcomes | Every agent has a named business owner and technical owner | Registry entry with owners; escalation path |
| 2. Understand impacts | Assess effects on people and stakeholders | Impact assessment before an agent acts for customers or staff | Impact assessment linked to each agent version |
| 3. Measure and manage risks | Identify, assess and treat AI risks | Risk tier per agent based on autonomy, tools and data | Risk rating, treatments, residual risk sign-off |
| 4. Share essential information | Transparency with people affected | Tell users when they deal with an agent; keep a register of AI use | AI register; disclosure text; model and vendor records |
| 5. Test and monitor | Test before and after deployment | Pre-release evals, red-teaming, production quality monitoring | Eval results, monitoring dashboards, incident log |
| 6. Maintain human control | Humans can oversee and intervene | Approval for high-impact actions; ability to pause or stop an agent | Approval records, kill and quarantine records |

## Which practices are hardest for agents?

**Accountability (practice 1)** is hard when agents are built by many teams on many platforms. You cannot assign owners to agents you have not found. Start with an inventory from automated discovery, then assign owners. [AI agent owner responsibilities](/blog/ai-agent-owner-responsibilities) sets out what an owner should be on the hook for.

**Testing and monitoring (practice 5)** is harder for agents than for single-turn chat. A single answer can be fine while a multi-step task goes wrong. Test whole conversations and tool sequences, and score production traffic continuously, not just at launch.

**Human control (practice 6)** needs two things: a human checkpoint before high-impact actions, and the technical ability to stop an agent quickly. The first is a design choice in the agent or its platform. The second needs a control that works across platforms. See [human approval for high-risk agent actions](/blog/human-approval-for-high-risk-ai-agent-actions) and [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent).

## A practical starting checklist

1. Build an AI register that lists agents, not only AI products bought.
2. Assign an accountable owner to every agent in the register.
3. Define risk tiers and the controls each tier requires.
4. Write a short impact assessment template and require it for customer-facing and decision-making agents.
5. Add disclosure wording for any agent that talks to customers.
6. Set evaluation standards per tier, including adversarial testing for higher tiers.
7. Turn on production monitoring with thresholds and named responders.
8. Confirm you can block, quarantine or stop each production agent, and test it.
9. Map these controls to any binding rules in your sector, such as CPS 230 for APRA-regulated entities. Our [CPS 230 and AI agents](/blog/apra-cps-230-and-ai-agents) post covers that link.

For mapping across frameworks, see [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## How Opencontroller does this

Opencontroller generates evidence for each practice; it does not certify alignment. Auto-discovery across AWS, Azure and Google Cloud and connectors to platforms such as Salesforce, SAP, Databricks and HubSpot feed a governed registry with owners and versions (practices 1 and 4). Posture flags risky permissions (practice 3). Multi-turn and red-team evals, LLM judges on production traffic and live groundedness and task-adherence guardrails support practice 5. Per-agent Block and Quarantine controls and rogue-agent handling support practice 6. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is the Guidance for AI Adoption mandatory?**
No. It is government guidance that replaced the Voluntary AI Safety Standard. Sector rules such as APRA CPS 230 remain binding where they apply.

**Does the guidance mention AI agents specifically?**
The main guidance applies to all AI systems rather than naming agents. Its practices on human control and monitoring are the most relevant to agents.

**What are the six essential practices?**
Decide who is accountable, understand impacts, measure and manage risks, share essential information, test and monitor, and maintain human control.

## Key facts

Australia's Guidance for AI Adoption was published by the National AI Centre (Department of Industry, Science and Resources) on 21 October 2025. It replaces the 2024 Voluntary AI Safety Standard and streamlines its 10 guardrails into six essential practices: decide who is accountable, understand impacts, measure and manage risks, share essential information, test and monitor, maintain human control. It has foundational guidance and extended implementation practices. It is guidance, not legislation. Opencontroller provides evidence for these practices via discovery, registry, Posture, evals, LLM judges, guardrails and per-agent Block and Quarantine. Licensing starts from $250K per deployment per year.

## Related

- [What does APRA CPS 230 mean for AI agents in Australian financial services?](https://opencontroller.lyzr.com/blog/apra-cps-230-and-ai-agents.md)
- [What is an AI agent owner responsible for?](https://opencontroller.lyzr.com/blog/ai-agent-owner-responsibilities.md)
- [Which AI agent actions need human approval?](https://opencontroller.lyzr.com/blog/human-approval-for-high-risk-ai-agent-actions.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
