A board AI governance policy template for Australia, clause by clause

[Organisation name] board AI governance policy

Approved by the board on [date] · Minute reference: [ ] · Executive owner: [name, role] · Next review: [date, no more than 12 months away]

1. Purpose and scope

This policy sets how the board of [organisation name] governs the organisation’s use of artificial intelligence: what the board decides itself, what it delegates and to whom, the risk it is willing to take, and what it is told and when.

It applies to every AI system the organisation builds, buys or uses, including AI features inside software already in use, and AI used by contractors and service providers on our behalf. “AI system” has the meaning given in our [organisation AI policy / staff AI use policy].

This policy sits above our [organisation AI policy, approved by the board / CEO] and our staff AI use policy, approved by [executive owner / CEO]. Where they conflict, this policy prevails.

Commentary. The business of a company is “managed by or under the direction of the directors” (Corporations Act, s 198A(1), a replaceable rule your constitution may vary).⁸ This policy is the direction for AI. Borrow the definition of an AI system from the federal template rather than writing your own. It includes generative AI and chatbots and excludes spreadsheet formulas and “if–then” macros, which keeps the policy off ordinary automation.¹

2. What the board decides, and what it delegates

The board reserves to itself:

• (a) approving this policy and the AI risk appetite in clause 3;
• (b) approving every Tier 3 AI use (clause 6) before it goes live;
• (c) approving any AI use likely to have significant effects on staff, such as removing or substantially changing roles (clause 10);
• (d) approving AI commitments over $[amount] a year, and any AI service that a critical operation depends on;
• (e) [listed entities: deciding, or delegating to the disclosure committee, whether an AI matter must be disclosed to ASX];
• (f) approving the AI roadmap, and any change to what AI is for in the organisation or to the measures the roadmap sets.

Everything else in this policy is delegated to the executive owner, who may sub-delegate approval of Tier 1 uses to process owners. The executive owner exercises the delegation in line with this policy and any direction of the board. This delegation is recorded in the minutes of the meeting that approved this policy.

Commentary. Unless the constitution says otherwise, directors “may delegate any of their powers” to a committee, a director, an employee or any other person. The delegate must follow the board’s directions, and the delegation must be recorded in the minute book (s 198D and its note).⁸ Delegating does not end the board’s responsibility. Under s 190 a director answers for the delegate’s exercise of the power unless the director reasonably believed, after proper inquiry where needed, that the delegate was reliable and competent, and the federal guidance says the same of AI in plainer words.⁹¹⁰ Clauses 12 and 13 are how the board keeps that belief reasonable. The dollar figure in (d) is yours to set. Tie it to the delegations of authority you already have.

3. Our risk appetite for AI

The tiers are the approval tiers in clause 6: Tier 1 is approved by the process owner, Tier 2 by the executive owner on a written risk assessment, and Tier 3 by the board.

We will not accept any AI use that we could not explain to the person it affects, or that breaks the law.

Commentary. The AICD guide says the board should discuss its risk appetite for AI and communicate it to management, and that a board may set different appetites for different technologies, tools and groups of people, such as generative against agentic systems or staff-facing against customer-facing uses.⁶ The table does that. ASIC found that the more mature of the 23 financial services and credit licensees it reviewed “included AI explicitly in their risk appetite statement”, and that less mature ones did not.¹¹ If you have a risk appetite statement, put this table in it. The board, not management, chooses the words in the middle column.

4. Who is accountable

• (a) [Name, role] is the executive owner of AI, accountable to the board for this policy, the organisation and staff AI policies, the register and the reports in clause 12.
• (b) Every AI use has a named process owner: the manager who owns the work the AI does. When work moves into an AI system, accountability for that work stays with the process owner. It never sits with the system or with the vendor.
• (c) No AI use goes live until its process owner is named on the register.
• (d) [If you have one: the management AI committee’s membership, mandate and quorum are set out in [document]. It meets at least [every two months] and reports through the executive owner.]

Commentary. The federal template gives each AI system an “accountable person”, and Foundations asks for a senior leader to own AI governance overall.¹⁴ The AICD guide says that in smaller organisations the CEO may take this role, and that a dedicated AI committee may not be practical in a small organisation.⁶ Clause 4(b) goes one step further than the federal “system owner”. It ties accountability to the work, not the tool, so that replacing a vendor or adding a second tool does not leave a gap. ASIC asks licensees the same two questions: who is accountable at model level and overall, and whether they get the reporting to do the job.¹¹

5. The AI register

The executive owner keeps a register of every AI use, using the fields in the National AI Centre’s AI register template and these six more:

1. the process owner;
2. the risk tier, who approved it and when;
3. whether personal information is used, and the country where it is stored or processed;
4. whether the use makes, or substantially shapes, decisions that could significantly affect a person;
5. the vendor contract and its renewal date;
6. the date the entry was last reviewed.

The register is reviewed at least [every six months], and its changes go to the board in every report under clause 12.

Commentary. The federal register template has 12 fields, from system name and owner to screening outcome and the people affected.² None records who approved the use, where the data goes, which decisions it touches or when it was last checked, which are the things a board asks. ASIC found that some licensees “required extra time to collate use cases” for its review and suspected that “a lack of an AI inventory” contributed.¹¹ In one case study a licensee found models missing from its own register. The Australian Prudential Regulation Authority (APRA) lists “an inventory of AI tooling and AI use cases” among its minimum expectations for regulated entities.¹² Use the Excel version of the federal register. In the Word version we read, the descriptions in the field table sit one row out of step from “Known limitations” downward.²

6. Approving new AI uses by risk tier

Before money or time is committed, every proposed AI use is screened with [the National AI Centre’s AI screening questions / our screening questions].

• Tier 1: every screening answer is “no”. The process owner approves it and adds it to the register.
• Tier 2: any answer is “yes”, and none of the Tier 3 tests applies. The executive owner approves it on a written risk assessment, and it is reported to the board at the next meeting.
• Tier 3: the use (a) makes, or substantially shapes, a decision that could significantly affect a person’s rights or interests; (b) acts without a person’s approval toward customers or in a core system; (c) uses sensitive or health information; (d) supports a critical operation; or (e) costs more than $[amount] a year. The board approves it before it goes live, on a paper covering purpose, fit with clause 3, privacy, the vendor, human oversight, consultation, the manual fallback and how its value will be measured.
• Prohibited: the uses listed in clause 3 are not approved at any tier.

An existing use is screened again, and may change tier, when what it does, the data it uses or its vendor changes.

Commentary. The federal screening tool asks seven yes-or-no questions and says plainly that “there are no scores or automatic outcomes”. A “yes” means you should consult “senior leadership or a governance committee”.³ The federal “Create an AI policy” page asks you to check before publishing that your policy says “who approves higher-risk use cases”.¹³ The tiers are that answer. The AICD checklist for small and not-for-profit boards asks the board to “establish a proportionate process for approving the use of AI systems in line with their potential impact”.⁷ Tier 3’s first test uses the words of the Privacy Act’s new automated-decision rule (clause 8), so a use that triggers the privacy-policy duty also reaches the board.¹⁴

7. Human oversight

For every Tier 2 and Tier 3 use, the register names the person who can pause, override or switch it off, and says how. Where a critical operation relies on AI, a manual way of doing the work is kept and tested [once a year]. Decisions about a person’s employment, pay or [credit / eligibility / services] are made by a person. Where work carries a professional sign-off (an engineer’s, an accountant’s, a lawyer’s, a registered health practitioner’s), the person who signs remains accountable for AI-assisted work and meets their profession’s standards for it; the register notes which uses touch such work. [Delete if it does not apply.] AI used in clinical care sits under [our clinical governance framework] and the therapeutic goods rules for software, not under this policy.

Commentary. This is the federal template’s “human oversight and control” statement and the sixth practice in the Guidance for AI Adoption, “Maintain human control”, turned into something the register can show.¹⁴ APRA expects regulated entities that use AI for critical operations to have “credible fallback processes”.¹² The rule that a person approves anything an AI tool sends, pays or changes is in Article 1’s staff policy, section 6. What each Australian professional body says about its members’ AI-assisted work, and why clinical AI needs its own regime, are in how to write an AI policy. A board also pays for AI in reading time. Every paper it receives is now cheap to lengthen, and the staff policy’s rule that anything sent is as short as the decision it supports (Article 1, section 14) is one the board can ask management to apply to its own papers.

8. Privacy and data

[Choose one.] The Privacy Act covers us because [our annual turnover is over $3 million / we are a health service provider / other reason]. or The Privacy Act does not cover us, and we follow this clause anyway.

• (a) No personal information goes into an AI tool that is not on the register.
• (b) Before personal information goes to an AI vendor that stores or processes it outside Australia, the executive owner records the country and the steps taken to make sure the vendor handles it as the Australian Privacy Principles (APPs) require.
• (c) AI tools that hold personal information sit inside our information security controls.
• (d) A privacy impact assessment is done before any Tier 3 use involving personal information.
• (e) From 10 December 2026, our privacy policy describes the kinds of personal information used, and the kinds of decisions made or substantially shaped, by computer programs where a decision could significantly affect a person. The executive owner confirms to the board [before 10 December 2026 and at each annual review] that it is current.

Commentary. Whether the Act covers you is a turnover test, not a headcount, with some businesses covered whatever they earn. The test, and the automated-decision rule behind (e), are set out in AI regulation in Australia in 2026.¹⁵ (b) is APP 8.1 and s 16C, which make an overseas recipient’s breach the entity’s own. (c) is APP 11.1, the security principle. (e) is APP 1.7 to 1.9, commencing 10 December 2026.¹⁴¹⁶ (d) follows the recommendation of the Office of the Australian Information Commissioner (OAIC) of a privacy impact assessment as part of privacy by design.¹⁷

9. Vendors and service providers

Before an AI service is contracted, the process owner checks the vendor’s own terms on whether it trains on our data, where it stores it, how long it keeps it and what our administrator can control. For Tier 3 uses and any AI service a critical operation depends on, the contract also covers our right to information and audit, notice of material changes to the model or service, incident notification, changes to data handling, and how we would leave. The executive owner tells the board when a vendor changes terms in a way that would change a register entry.

Commentary. The four vendor checks are the ones in Article 1’s staff policy, section 2. The contract list follows what APRA found missing at large banks, insurers and super funds: contracts “often lagged practice”, with “limited evidence of specific provisions addressing audit rights, model updates and deviations, incident notification or changes to data handling”, and few entities had “tested exit and substitution strategies for critical AI providers”.¹² The AICD guide says the board should have visibility over the risks of external AI vendors and their data protection settings.⁶ How to choose and connect AI software, and the questions to ask a vendor, are in choosing AI software and connecting it to your systems.

10. Staff, volunteers and consultation

• (a) The staff AI use policy tells staff and volunteers which tools they may use and how.
• (b) Any proposal for an AI use likely to have significant effects on staff (ending roles, changing what a role does, retraining, changing hours) comes with a consultation plan under [our award / enterprise agreement, clause number]. Where the board approves the use, consultation starts as soon as practicable after that decision.
• (c) [Optional: what staff get from the organisation’s use of AI, for example roles that fit in their hours, training time or role redesign, and how the board will know whether that is happening.]
• (d) [Choose one.] Nobody is required to use AI for their own work. or Some roles must use AI in prescribed and proven ways. For each such role the way is documented [in the role description and the systems map], and nothing outside it is required. Either way, the staff AI use policy says how a person who does not want to use AI is supported, and how a concern is raised.

Commentary. The consultation duty sits in your award or enterprise agreement. Once an employer makes a definite decision on a major change in technology likely to have significant effects on employees, it must give notice, discuss the change and give the information in writing, starting as soon as practicable after the decision (Clerks award, clause 38, for example).¹⁸ A board approval under clause 2(c) or 6 can be that definite decision, which is why the plan comes with the board paper. The clause text, and the person-or-AI question for a single role, are in should your next hire be a person or AI. (c) and (d) are the board’s half of two sections in the staff policy, what staff get (Article 1, section 16) and staff who are not sure about AI (section 13). The reasoning, and why a promise about jobs is spoken by the owner rather than written into a policy, is in how to write an AI policy.

11. Incidents and escalation

Staff report AI incidents under the staff AI use policy. The executive owner tells the chair within [24 hours], and the board at or before its next meeting, when:

• personal information may have been involved in a data breach;
• AI output has caused harm, or has reached a customer or the public when it should not have;
• a regulator contacts us about our use of AI;
• a vendor has a security incident affecting our data;
• a Tier 3 use fails, is switched off or is found to be making errors;
• [listed entities: the matter may be market-sensitive, in which case it goes to the disclosure officer at once].

Commentary. The federal template sends incidents to the system owner or compliance monitor, and it does not say when the board hears.¹ The AICD lists, among its governance red flags, that “AI risks are not reported to the board or are not escalated in a timely way”.⁶ For an organisation the Privacy Act covers, a suspected eligible data breach starts a clock. The OAIC expects an assessment within 30 days of becoming aware of it.¹⁹

12. What the board is told, and how often

The executive owner reports to the board [every quarter / at every meeting], in plain language, on:

• changes to the register since the last report;
• Tier 2 approvals made, and Tier 3 proposals coming;
• incidents and near misses;
• [three to five measures agreed with the board, e.g. value delivered against each Tier 3 business case, errors caught by human review, staff policy training completion];
• progress against the AI roadmap’s measures, with its [month-6 and month-12] reviews brought to the board as decisions;
• vendor term changes and changes in the law.

Once a year the executive owner reports in full against the risk appetite in clause 3, and proposes any change to it.

Commentary. ASIC rated periodic board reporting on AI risk as better practice and ad-hoc or absent reporting as poorer practice.¹¹ The AICD guide says the board should work with management to set the measures and “cadence of reporting” it expects, and that board reporting should avoid unnecessary technical language.⁶ The roadmap line, with clause 2(f), is where strategy meets this policy. The plan itself lives in How to build an AI roadmap, and the policy only says that the board approves it and hears how it is going. The AICD guide’s operating model puts strategy first of its four parts.⁶ APRA expects boards of regulated entities to oversee an AI strategy “supported by effective monitoring and reporting (including for third party dependencies), with clearly defined triggers”.¹²

13. Board competence

Each director completes [an AI governance briefing or course] within [six] months of appointment. The board holds [one session a year] on AI with the executive owner or an outside adviser. AI governance is in the board skills matrix. The board may take independent advice on AI at the organisation’s expense, up to $[amount] a year without further approval.

Commentary. The business judgment rule in s 180(2) turns partly on whether the director informed themselves to the extent they reasonably believed appropriate, and reliance under s 189 needs the director’s own assessment. Both are explained in your duty of care over AI.²⁰ APRA found that “many Boards are still developing the technical literacy required to provide effective challenge on AI related risks”, and expects boards to “maintain sufficient understanding and literacy with respect to AI”.¹² What each director’s own exposure is lies in a director’s personal exposure on AI.

14. Review

The board reviews this policy by the date at the top, and sooner if there is a significant AI incident, a change in the law, a new Tier 3 use or a material change by a vendor. The executive owner proposes the changes; the board approves them.

Commentary. The federal template’s review clause sets an annual review with similar triggers, and leaves approval to the “policy approvers”.¹ Here the approver is named. It is the board. The first review date to diarise is 10 December 2026, when APP 1.7 to 1.9 commence.¹⁶