AI governance in Australia: a practical starting point

This article is general information, not legal advice. It reflects published guidance and law as at 14 October 2026. For decisions about your own obligations, speak with a privacy or technology lawyer.

Most businesses don't need an AI committee or a 40-page framework to start. They need to know which rules already apply, make a handful of clear decisions, and make those decisions easy for staff to follow. This is a practical starting point for an owner, COO or general manager.

What is AI governance?

AI governance is the set of decisions, owners and checks that control how your organisation uses AI: which tools are approved, what information can go into them, who reviews the output, what you tell customers, and what happens when something goes wrong. Good governance makes safe use the easy default, so staff use AI openly and well instead of quietly and inconsistently.

What the current guidance and law actually say

The Australian Government's National AI Plan says its approach builds on existing, largely technology-neutral laws, with regulators keeping responsibility for AI-related harms in their areas and targeted laws added where needed. It is also establishing an AI Safety Institute to monitor, test and share information on AI risks. For most businesses, the practical starting point is the Privacy Act and the guidance below.

The Privacy Act and the OAIC's AI guidance

The Privacy Act 1988 generally does not cover small businesses with annual turnover of $3 million or less, but some are covered regardless of turnover, including health service providers and businesses that trade in personal information. If you are covered, the OAIC's guidance on using commercially available AI products makes six points worth acting on:

• Privacy obligations apply to personal information you put into an AI system and to output that contains personal information.
• Personal information that AI generates or infers about someone is a collection under APP 3.
• Using personal information in AI for a secondary purpose needs consent, or must be within reasonable expectations and related to the original purpose (APP 6).
• You must take reasonable steps to ensure accuracy (APP 10), which in practice means human oversight and treating outputs as probabilistic, not as fact.
• As best practice, don't enter personal information, and particularly sensitive information, into publicly available generative AI tools.
• Update privacy policies and notices to reflect AI use, and make sure public-facing AI tools such as chatbots are clearly identified.

Training or fine-tuning on your own data

A separate OAIC guide covers developing and training generative AI models. It applies to businesses that fine-tune a commercially available model or use personal information they already hold to train one. Its key messages: reusing customer data for AI training is a secondary purpose that will often be hard to justify on reasonable expectations alone, so seek consent or offer a meaningful opt-out. And data being publicly available does not, by itself, make it fair to use.

Automated decisions: new privacy policy obligations from 10 December 2026

The Privacy and Other Legislation Amendment Act 2024 received Royal Assent on 10 December 2024, and its automated decision-making provisions commence on 10 December 2026. From then, they apply where you arrange for a computer program to make a decision, or do something substantially and directly related to making it, that could reasonably be expected to significantly affect an individual's rights or interests, and personal information is used to run the program. Your privacy policy must then describe:

• the kinds of personal information used in those programs;
• the kinds of decisions made solely by the program; and
• the kinds of decisions where the program does something substantially and directly related to making the decision.

The OAIC released a fact sheet, a flowchart and updated APP 1 guidelines on 30 September 2026. Processes worth checking include automated credit or payment-term decisions, application screening and eligibility rules. Map them now, while there is still time to describe them properly.

The Guidance for AI Adoption

Published by the National AI Centre on 21 October 2025, the Guidance for AI Adoption condenses the earlier Voluntary AI Safety Standard's 10 guardrails into six essential practices. There is a foundational version for organisations starting out and a more detailed version for those scaling up. The six practices are:

1. Decide who is accountable.
2. Understand impacts and plan accordingly.
3. Measure and manage risks.
4. Share essential information.
5. Test and monitor.
6. Maintain human control.

The foundations version asks you to name a senior leader as the overall AI governance owner, create an AI policy and keep an AI register. Everything below maps to those practices.

A one-page AI use policy outline

One page that people read beats twenty they don't. Cover these ten headings:

1. Purpose and scope: who it applies to, including contractors.
2. Owner: the named person accountable for AI use, and who staff ask.
3. Approved tools: a pointer to the register, and a rule that work data stays out of personal accounts.
4. Data rules: which data classes can go into which tools.
5. Human review: what must be checked, and by whom, before AI output is used.
6. Disclosure: when you tell customers AI was involved, and how chatbots identify themselves.
7. Prohibited uses: for example, decisions about individuals without human review, or uploading client files to unapproved tools.
8. Confidentiality and intellectual property: existing client confidentiality terms still apply when AI is involved.
9. Incidents: what to report, to whom and how fast.
10. Training and review: required training, and the date the policy will next be reviewed.

Classify data before anyone types it into a tool

Staff make dozens of small decisions a day about what to paste where. Four classes make those decisions quick.

Approve tools deliberately

The approved-tools register

The register is a simple table, owned by one person, that answers "can I use this, and for what?" For each tool, record:

• tool name, plan and business owner;
• approved use cases and the highest data class allowed;
• settings confirmed, such as the provider's use of your data for training restricted and retention limited;
• where data is stored and processed, including whether it leaves Australia;
• who has access, and how access is removed when people leave;
• date approved and next review date.

Include AI features switched on inside software you already use, such as your CRM, email platform or practice system. They are easy to miss and often process more customer data than anything else.

Questions to ask an AI vendor

These build on the due diligence questions in the OAIC's guidance:

• What data has the product been trained and tested on, and is it relevant to how we will use it?
• What are its known limitations, and what safeguards address them?
• What are the data flows? Can you or third parties access what we enter or generate?
• Are our inputs used to train your models, and can we switch that off?
• How long are inputs and outputs retained, and can we limit that?
• Where are the servers, and could personal information be disclosed outside Australia?
• What known security risks exist, and how are they managed?
• What evidence of testing can you share, and how will you tell us about material changes or incidents?

Human oversight rules that work in practice

• A named person is responsible for every AI-assisted output that leaves the business.
• Facts, figures, names and commitments are checked against a source before use.
• No decision that significantly affects an individual is made by AI alone without a human review step and a way for the person to query it.
• Records show which content was AI-generated, so it is treated as an assessment rather than fact.
• Anyone can pause a tool that produces harmful or wrong output, and critical work has a manual fallback.

Train staff and plan for incidents

Training

Train by role, not by tool. A short session should cover the policy, the data classes with real examples from your own business, how to check output, and how to report a problem. In the OAIC's fictional example of an employee uploading a customer's hardship application to a public AI tool, the company's response included restricting uploads, tightening its policy and scheduling regular refresher training. Build that rhythm in from the start.

When something goes wrong

AI incidents include wrong advice sent to a customer, a chatbot making commitments you can't keep, and personal information pasted into an unapproved tool. Treat the last one as a possible data breach. The OAIC's response steps are contain, assess, notify where required, and review. If you are covered by the Privacy Act and suspect an eligible data breach, meaning one likely to result in serious harm, you must take all reasonable steps to complete your assessment within 30 calendar days. Log every AI incident, even minor ones. The patterns show where the policy or training needs work.

Where to start this month

1. Name the owner.
2. Publish the one-page policy and the four data classes.
3. Start the register with the tools staff already use, including personal accounts you want to retire.
4. List any automated decisions about individuals before 10 December 2026.

Our AI governance & implementation work puts the policy, register, data classes, vendor reviews and training in place. It usually runs alongside AI strategy & roadmaps or the rollout of AI assistants & knowledge systems. The professional firm knowledge assistant worked example shows these controls in context, and AI assistants for your team covers where to start with assistants. For a quick view of your starting point, try the AI readiness assessment. We check against published guidance, but we are technologists and marketers, not lawyers.