ASIC and APRA have asked financial entities to move from AI awareness to action. This 30-day plan helps boards and risk teams turn that request into evidence they could show a regulator.
In June 2023, the Federal Court ordered Insurance Australia Limited (IAL) to pay a $40 million penalty. More than 600,000 of its customers had not received the full loyalty discounts they were promised. ASIC said IAL used a pricing algorithm that limited the discounts renewing customers could receive, so their premiums did not fall below a set price point.
The court penalised IAL for failing to deliver what it had promised customers, not for using an algorithm. That distinction is the key to how regulators are likely to approach AI, because they will judge what your systems did to customers and markets against duties that already apply.
On 27 August 2026, ASIC and APRA asked financial market entities to move from awareness of frontier-AI risks to tested action. The regulators said frontier AI can increase the speed, scale and sophistication of cyber threats, and that it compresses the time available to respond. Their joint statement followed nine industry roundtables in June and July with more than 600 attendees.
For boards and executives, the call raises a direct question. If a regulator asked tomorrow, could you show where AI is used, who owns it, what could go wrong, how your controls work and what happens when they fail? For many entities, the honest answer today may be “partly”, and 30 focused days can change that.
The call adds no new AI rules, so your existing obligations are the test
The August statement is a clear supervisory signal and not a new AI statute or prudential standard. Your existing financial services, consumer protection, governance, operational risk and information security obligations still apply when a process uses AI.
APRA made the same point in April 2026. It said it was “not proposing to introduce additional requirements at this stage” and described its AI letter as guidance. APRA also said it will take stronger supervisory action where an entity’s risk management falls short. The absence of new rules does not reduce the pressure, because the existing rules already reach AI.
Regulators will ask how you used AI, not whether you used it
A matter that involves AI can take several forms. AI might be the alleged cause of harm, a tool used in misconduct, or one part of a wider control failure. Using AI does not by itself prove a breach. A regulator is more likely to ask how you used the system, what you represented or decided, which risks you could foresee, and whether your controls met your duties.
ASIC applies a technology-neutral approach, so AI does not replace existing duties. For an AFS licensee, the baseline can include its obligations under section 912A of the Corporations Act. It can also include directors’ and officers’ duties and the prohibition on misleading or deceptive conduct. The exact obligations depend on the facts, the licence and the activity.
These duties land on people as well as organisations. ASIC Chair Joe Longo has said directors’ duties are principle-based and apply broadly. He said directors “must pay special attention” to those duties as companies deploy AI. For a director or accountable executive, weak AI governance can become a personal question, and not only a question for the entity. The Financial Accountability Regime adds to this, as we set out in our earlier article on APRA’s expectations.
APRA-regulated entities also need to confirm which prudential standards apply to their sector. CPS 230 has applied since 1 July 2025. Its transition for pre-existing contracts with material service providers ended on 1 July 2026 at the latest. CPS 230 requires entities to manage operational risk, maintain critical operations through severe disruption and manage service provider risk. CPS 234 covers information security, including incidents and information assets managed by third parties. Other standards, such as CPS 220, have a different scope, and superannuation entities may have corresponding SPS standards. Do not assume every standard applies to every entity.
ASIC and APRA have already told you which gaps they expect to find
ASIC’s 2024 review shows why a policy document alone may not protect you. ASIC examined 624 AI use cases across 23 licensees in banking, credit, insurance and financial advice. Nearly half of those licensees had no policies on consumer fairness or bias, and even fewer had policies on telling consumers when AI was used. ASIC said existing consumer protection provisions, director duties and licensee obligations already require appropriate governance.
APRA’s April 2026 letter found similar problems, including unclear ownership, missing AI inventories, point-in-time assurance and limited visibility of AI suppliers. We explain those gaps and APRA’s expectations of boards in APRA is watching your AI. The plan below turns them into work you can start this week.
Australia’s AI enforcement story is already running through existing law
Some commentators ask when Australia will see its first AI enforcement case. The more useful fact is that regulators already pursue algorithm-driven harm with the laws they have.
The 2023 IAL penalty is one example. A second IAG proceeding, filed in August 2023, is still before the court. ASIC alleges that customers of brands including SGIO, SGIC and RACV did not receive the full benefit of promised loyalty discounts on home insurance. In January 2024, Joe Longo said this was an action “in which AI-related issues arise”, where ASIC believes a demand model was part of the pricing process. As at 1 October 2026, ASIC’s case page reports that the liability hearing ended on 1 May 2026 and judgment is reserved. These are allegations, not findings, and the case is not a ruling that using AI breached the law.
For your readiness work, the lesson is that a regulator can investigate an AI outcome with familiar questions. Were customers treated as you told them they would be? Were decisions fair and properly controlled? Did you understand the limits of the system? Could you explain the outcome, find the affected people and fix the harm? A technical explanation of the model cannot replace evidence of accountable governance.
A 30-day plan turns awareness into evidence you can show
Thirty days will not complete your remediation, but it can give you an honest baseline, clear owners, and a dated plan, which is what a board or regulator may reasonably expect to see first. Each week below produces an output you can test.
Days 1 to 7: build an AI inventory you can use
List each material AI system and use case, including pilots, AI features embedded in vendor products, and agentic tools that can take actions. For each one, record:
- the business owner and accountable executive
- the purpose and the customers or market activity it affects
- the model, version, data categories and human decision points
- the provider and any material subcontractors
Flag uses that could affect eligibility, pricing, advice, claims, complaints, trading, fraud controls, critical operations or sensitive information. Where you do not know something, record the gap and do not treat the unknown as low risk. APRA’s letter expects an inventory and clear ownership across the AI lifecycle.
Output by day 7: an inventory with owners, and a shortlist of higher-risk uses.
Days 8 to 14: map each higher-risk use to the duties that apply
Ask your legal, compliance, and risk teams to map each higher-risk use to the obligations that apply to your entity, and not to a generic AI checklist. For conduct that ASIC regulates, examine licensing, advice, disclosure, fair customer outcomes, marketing claims and market integrity rules where relevant. For APRA-regulated entities, identify the CPS or SPS requirements for operational risk, continuity, service providers, security and risk management.
Record the legal basis, control owner, evidence source and any gap for each use. Keep binding law and standards separate from regulator letters, speeches, reports and consultation papers, so that your board can see which is which.
Include privacy law in this map. From 10 December 2026, entities that use computer programs to make decisions that significantly affect individuals must explain those decisions in their privacy policy. Our article on Australia’s AI disclosure deadline explains what that requires.
Output by day 14: a duty-to-control map that shows gaps and owners.
Days 15 to 21: test how things fail, not only how accurate the model is
Run scenarios for the failures that could hurt customers or operations:
- misleading or fabricated outputs
- drift or degraded performance, and biased outcomes
- unauthorised actions by an agentic tool
- compromised credentials and prompt injection
- a supplier outage, a material model change or the loss of a critical AI provider
Test the full path for each scenario: detection, escalation authority, human override, customer protection, manual fallback, recovery and communications. For each material provider, check what your contract and technical arrangements let you see, audit, control, replace or exit. The ASIC and APRA roundtable paper stresses tested response and recovery plans, strong cyber fundamentals, dependency mapping and board decisions agreed “before a crisis hits”. For the ongoing testing and monitoring capabilities that should follow this first round, see the QA framework in APRA is watching your AI.
Output by day 21: test results, with each failed or untested control listed.
Days 22 to 30: put tested evidence in front of the board
Give the board or risk committee the inventory, the most important use cases, the largest control gaps, supplier concentrations, test results, incidents and near misses. Give each remediation item an owner and a date. Ask the board to record decisions on risk appetite, permitted use, human review, escalation and the conditions for suspending a system.
A good board pack lets directors identify the difference between a control that exists on paper and a control that has been tested and works.
Output by day 30: recorded board decisions and a dated remediation plan.
Build the evidence file before you need it
For each material use case, keep a proportionate record that would let an independent reviewer reconstruct what happened. The record could include:
- approvals and risk assessments
- model, prompt, configuration and version history
- relevant inputs and outputs
- testing and monitoring results
- human review and override records
- complaints and customer impacts
- change tickets
- supplier due diligence and contracts
- incident timelines, escalation decisions and remediation evidence
Set retention, access and privacy controls with your legal and records teams. Do not collect sensitive data without a clear purpose in the name of preparedness. The aim is to keep relevant, lawful evidence and a clear decision trail, not to create a new unmanaged store of data.
If an incident occurs, start your existing incident, legal hold, and customer remediation processes. Preserve relevant records before routine deletion or system changes remove the trail. Do not assume one reporting deadline applies to every event. CPS 234 requires notice to APRA within 72 hours of becoming aware of a material information security incident. CPS 230 requires notice within 72 hours for certain operational risk incidents, and within 24 hours for a disruption to a critical operation outside tolerance. Confirm the trigger and deadline for your entity and its facts, and coordinate any ASIC, APRA and other reporting obligations.
Ready means you can answer five questions without forming a new committee
An entity that is ready for AI enforcement can answer five questions with evidence it already holds:
- Where was AI involved?
- Who owned the decision?
- Which rule and control applied?
- What evidence shows the control worked?
- What did the entity do when the control did not work?
These are the same questions that sat behind the 2023 IAL penalty, where a pricing system produced a result that did not match what customers were promised. Readiness does not require you to predict the first AI case or to treat every model as a risk to be avoided. It asks that, if someone challenges a customer, market, or operational outcome, your organisation can explain its choices, show its controls, and act quickly to contain and fix the harm. The next 30 days are enough to find out how close you are.
Frequently asked questions
Did ASIC and APRA create new AI rules in August 2026?
No. Their 27 August statement asked financial market entities to act on frontier-AI risks, and their roundtable paper set out recommendations for preparedness. APRA’s April 2026 AI letter was guidance, and APRA said it was not proposing additional requirements at that stage. Existing laws and the prudential standards that apply to you remain the binding baseline.
Is using AI automatically a breach of financial services law?
No. Using AI does not by itself establish a breach. The relevant question is whether the entity’s conduct, decisions, representations, governance and controls met the obligations that apply to it. ASIC has said existing obligations apply when licensees use AI.
Has ASIC brought Australia’s first AI enforcement case?
Not in the sense of a court finding that AI use broke the law. ASIC has said AI-related issues arise in its current case against IAG subsidiaries about loyalty discounts. As at 1 October 2026, ASIC reports that judgment is reserved in that case, so its allegations have not been decided.
What if a key AI provider will not disclose model details?
Treat limited transparency as a supplier risk and an assurance risk. Identify what you can still test, monitor, contract for, restrict or validate independently. Assess concentration, portability and exit options. If you cannot understand or control a material risk, escalate it and reconsider the use case or the conditions of its use. APRA’s AI letter addresses supply chain visibility, assurance and dependence on critical providers.
This article is general information and not legal advice. It reflects publicly available material as at 1 October 2026. Your entity’s obligations and notification duties depend on its facts and on the current instruments, so seek advice before you act.