← Back to Insights

Buying frontier AI: the assurance Boards should demand now

Australia’s proposed frontier AI rules sharpen a question every Board should already be asking: can we explain why the AI we buy is fit for the authority, data and decisions we give it?

Stronger duties for model developers would be welcome. They would not replace the buyer’s governance.

The organisation deploying AI still needs evidence about its own use, clear accountability and a workable response when something goes wrong.

Australia’s AI adoption guidance makes that responsibility explicit. [1, 5]

What is changing, and what is not

On 10 November 2026, the Australian Government announced plans for mandatory safety and security protections applying to frontier AI companies wishing to operate in Australia. Its six proposed obligations cover security controls throughout development and deployment; management of serious model risks; independent model evaluation; rapid incident notification to government; published safety and security frameworks; and responsiveness to updated standards and rules. [1]

The Government proposes a targeted AI safety regulator, working with the Australian AI Safety Institute for technical expertise and oversight. These are announced protections, not a basis for claiming that suppliers are already certified under a new regime. The release says legislative detail remains subject to further work and consultation; it does not specify commencement dates, numerical scope thresholds or penalty amounts. [1]

Existing law is a separate matter. Directors’ duties, privacy, consumer, anti-discrimination and sector-specific requirements may already apply to an organisation’s use of AI. Which obligations apply depends on the organisation and the activity, not simply on whether its supplier calls a model “frontier”. [2]

For financial services, the warning is already concrete. In April 2026, APRA reported overreliance on vendor presentations, opaque supply chains and limited evidence of tested exit arrangements. APRA’s expectations apply to its regulated entities. Other Boards can use the observations as a practical challenge to management, not as a claim that APRA regulates every sector. [4]

The privacy deadline that cannot wait

From 10 December 2026, APP entities must include additional privacy-policy information where they arrange for a computer program to use personal information to make, or do something substantially and directly related to making, decisions reasonably expected to significantly affect an individual’s rights or interests. [3]

The required information covers the kinds of personal information used and the kinds of relevant decisions made solely by the program or supported in that substantial and direct way. Human review does not automatically take a process outside scope. This is not a blanket disclosure rule for every use of AI, nor is it confined to frontier models. Management should identify relevant decision processes now and obtain advice on coverage and any applicable exceptions. [3]

Six questions that turn procurement into governance

The following are WeGovernAi’s recommended buying controls, drawing on official guidance. They are not a claim that every control is a universal legal requirement. Scale the evidence and safeguards to the consequences of failure. [5, 6]

1. What are we approving, and who owns it?

Require an AI register that records purpose, accountable executive, affected people, permitted data and the authority given to the system. Distinguish assistance from action. Ask management to set acceptance criteria and unacceptable uses before selecting a supplier, supported by an impact and risk assessment. The Board should set direction and risk appetite; management should own implementation and demonstrate that controls work. [4, 5]

2. What evidence supports this use, beyond the demonstration?

Identify the application vendor, underlying model developer and material hosting and subcontractor dependencies. Request documentation, known limitations, evaluation methods and relevant independent assurance. Then test the complete service against the organisation’s own tasks and acceptance criteria. A strong model benchmark or a supplier’s safety framework is not a substitute for testing the system the organisation will actually operate. Record gaps and the residual-risk decision. [4, 5]

3. What happens to our data, and what can we enforce?

Map inputs, retrieved information, outputs and logs, including who can access them and where they are handled. Assess privacy, confidentiality, retention, deletion and training uses. Seek contract terms covering those matters, material model changes, assurance access, incident cooperation and exit assistance. “Not used for training” does not settle all data risks. OAIC recommends, as best practice, not entering personal information, especially sensitive information, into publicly available generative AI tools. Enterprise products still require due diligence. [4, 5, 6]

4. Would we hear about an incident in time to act?

Agree customer-notification triggers, timeframes and a monitored contact route. Connect supplier reporting to internal cyber, privacy, legal and operational response. The proposed frontier duties refer to notification to the Australian Government; buyers should not assume that this guarantees notice to them or fulfils their own reporting obligations. Rehearse suspension, investigation, evidence preservation and escalation before a critical deployment. [1, 5, 6]

5. Can people intervene, and can the business continue?

Give human overseers the knowledge, time and authority to challenge or stop unsafe operation. For agents, restrict permissions, log actions and require approval for high-impact changes. Test how access can be revoked. For critical processes, demonstrate a credible fallback and assess whether another provider is genuinely substitutable. Several applications can share the same underlying model dependency, so supplier counts alone do not establish resilience. [4, 5]

6. What will the Board see after launch?

Require reporting on material AI uses, performance, incidents, complaints, near misses, control failures and accepted risks. Retest material changes and review assurance independently of vendor marketing. A useful Board report should show what has changed, where limits were breached and what management did, rather than simply list tools deployed. Keep decisions, testing and approvals traceable throughout the service lifecycle. [4, 5]

Use frameworks to organise evidence, not replace it

Australia’s October 2025 Guidance for AI Adoption evolves and simplifies the 2024 Voluntary AI Safety Standard into six essential practices. The older ten guardrails remain a reference, but should not be presented as the only current Australian guidance or as enacted law. [7]

Map one set of operational evidence to relevant frameworks: ISO/IEC 42001:2023 for an AI management system; NIST’s AI RMF for Govern, Map, Measure and Manage; and the OECD AI Principles for human rights, transparency, safety and accountability. These serve different purposes. A management-system certificate should not be treated as proof that a particular application is safe for the buyer’s intended use. [8, 9]

A practical 90-day agenda

As a management plan, not a regulatory timetable, WeGovernAi recommends:

  • First 30 days: identify consequential AI uses and hidden dependencies, name accountable owners and prioritise evidence gaps. Address the 10 December privacy-policy deadline before it arrives, not at the end of this period. [3, 5]

  • Days 31-60: assess critical suppliers, negotiate priority contract protections and test incident contacts and intervention controls. [4, 5]

  • Days 61-90: exercise a critical-process fallback, obtain independent challenge of key assurance gaps and bring residual-risk decisions to the appropriate governance body. [4, 5]

The Board’s test is simple: can management show why the service is acceptable for its intended use, who owns the risk and how the organisation would contain a failure?

If the answer depends entirely on the supplier’s confidence, the governance work is unfinished. Good procurement gives an organisation room to innovate without giving away control.

General governance information, not legal advice. Sources checked on 15 November 2026.

Sources and endnotes

[1] Prime Minister of Australia. New Australian AI safety and security protections
10 November 2026. Six proposed duties, proposed regulator and legislative status.
https://www.pm.gov.au/media/new-australian-ai-safety-and-security-protections

[2] Department of Industry, Science and Resources. Guidance for AI Adoption: The legal landscape
Live guidance. Existing Australian laws that may apply to AI and supply-chain relationships.
https://www.industry.gov.au/publications/guidance-for-ai-adoption/legal-landscape

[3] Office of the Australian Information Commissioner. APP guidelines, Chapter 1
Updated 30 September 2026. Paragraphs 1.35-1.52: 10 December commencement, scope, disclosures and human review.
https://www.oaic.gov.au/privacy/australian-privacy-principles/australian-privacy-principles-guidelines/chapter-1-app-1-open-and-transparent-management-of-personal-information

[4] Australian Prudential Regulation Authority. APRA Letter to Industry on Artificial Intelligence (AI)
30 April 2026. Board oversight, supplier concentration, assurance and tested exit arrangements. APRA-regulated entities only.
https://www.apra.gov.au/news-and-publications/apra-letter-industry-artificial-intelligence-ai

[5] National AI Centre. Guidance for AI adoption: implementation guidance
Live guidance. Practices 1-6, including accountability, impact assessment, supply chains, incidents, testing and human control. Buying controls adapt this guidance.
https://www.ai.gov.au/staying-safe-and-responsible/essential-ai-practices/guidance-ai-adoption-implementation-guidance

[6] Office of the Australian Information Commissioner. Guidance on privacy and the use of commercially available AI products
Published 21 October 2024; page also displays an update of 17 January 2025. Product due diligence, privacy by design, ongoing review and the public-tool recommendation.
https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products

[7] Department of Industry, Science and Resources. Voluntary AI Safety Standard
5 September 2024; live page checked. Ten guardrails and notice that the 21 October 2025 guidance evolves the standard into six practices.
https://www.industry.gov.au/publications/voluntary-ai-safety-standard

[8] National Institute of Standards and Technology. AI RMF Core
AI RMF 1.0. Govern, Map, Measure and Manage; voluntary and context-based risk management.
https://airc.nist.gov/airmf-resources/airmf/5-sec-core/

[9] Organisation for Economic Co-operation and Development. AI principles
Adopted 2019; updated 2024. Human rights, transparency, safety, accountability and lifecycle risk management.
https://www.oecd.org/en/topics/ai-principles.html