Building on the first two articles in our series – AI disputes: the invisible defendant? and AI disputes: AI in the boardroom – this article examines what Australian financial services licensees and Australian credit licensees (together, Licensees) should be doing now to manage AI’s risks while capturing its opportunities.

Key takeaways

  • Existing obligations apply: Licensees' general obligations apply to their licensed services however they are delivered, including through AI, and Licensees remain responsible for the results.
  • Governance is lagging adoption: ASIC and APRA have warned that some Licensees are adopting AI faster than they are updating their risk and governance frameworks – a gap likely to widen as adoption accelerates.
  • Adequacy of controls and systems: The law does not require perfection, but it does require adequate, risk-appropriate controls that work in practice.
  • Human oversight remains critical: Licensees should understand their AI tools' limitations and maintain meaningful human oversight, particularly of decisions that impact consumers.
  • Decisions must be explainable: If ASIC asks how a customer outcome was reached, Licensees should be able to explain the process and produce a clear evidentiary trail.
  • Regulatory scrutiny is increasing: ASIC and APRA have called for “decisive action” on frontier AI risk, and ASIC has announced a supervisory review of banks' AI use.
  • Enforcement is next: ASIC has named “consumer harm arising from poorly implemented technologies, including AI and automated systems” as one of its enforcement priorities for 2027, and will be looking for test cases.

Artificial intelligence has the potential to transform the provision of financial services and credit in Australia. It provides opportunities for more efficient, accessible and tailored products and services. However, AI can also amplify existing risks to consumers and introduce new ones.

ASIC, Report 798, Beware the gap: Governance arrangements in the face of AI innovation, October 2024

Why AI risk matters now

Artificial intelligence (AI) is becoming central in Australian financial services to assess risk, detect fraud, personalise and price products, communicate with customers and handle complaints.

For Licensees, AI is not simply an IT issue; it is a governance, operational risk and conduct issue. While there is no bespoke AI legislation, existing obligations already apply to this new class of risk, and APRA and ASIC have warned that not all Licensees are ready for it. As Joe Longo, then ASIC Chair, put it in July 2025, “we’ve got to kick the tyres a bit and see just how far our existing technology-neutral framework will flex”.

The practical questions include when an AI inaccuracy becomes misleading conduct and when an automated decision reveals a control failure. Agentic AI raises further novel questions – for example, if a customer gives their AI agent their bank account details to action a task, what is the Licensee’s responsibility if something goes wrong?

How AI is regulated

Australia has no standalone AI legislation. Instead, AI is regulated through existing technology-neutral laws and sector-specific frameworks.

Three recent developments show a national AI-specific framework taking shape:

  1. Office of AI and national AI Standards

    On 15 July 2026, the Government established the Office of AI and announced mandatory national AI Standards, expected to be legislated in early 2027, which will regulate large data centres and frontier AI developers operating in Australia, including on training and copyright and, as more recently announced, safety and security.

  2. Joint Select Committee on AI

    The committee is examining AI's opportunities and risks, including whether existing laws are adequate, and names financial services as a sector of comparative advantage. The committee gave early advice on 12 October 2026. Its final report is due by 30 November 2026 and any gaps it identifies could shape reform affecting Licensees.

  3. AI safety standards and regulatorOn 10 November 2026, the Government announced mandatory safety and security obligations for frontier AI companies, enforced by a proposed new regulator. Whether they will extend to deployers such as Licensees is not yet clear, but Licensees can expect them to filter through supplier contracts, incident notification arrangements and the model information available to them.

Two nearer-term changes are more pressing for Licensees. From 10 December 2026, Licensees bound by the Australian Privacy Principles must include certain information in their privacy policies where they use personal information in the operation of computer programs, including AI, to do a thing that is substantially and directly related to making decisions that could significantly affect individuals. Further, the Scams Prevention Framework will impose new obligations on banks, mostly from 31 March 2027, adding a further layer of accountability as AI-driven fraud and deepfake scams outpace traditional controls.

Regulatory views

Regulators have also become more specific about their expectations (see our previous article AI governance: existing obligations, regulatory expectations and increased scrutiny). On 30 April 2026, APRA called for a “step-change” in AI risk management, warning that governance and resilience practices “are not keeping pace with the scale, speed, and complexity of AI adoption”. ASIC followed on 8 May 2026 with an open letter urging all Licensees to strengthen their cyber resilience urgently, and joint APRA-ASIC roundtables in mid-2026 stressed that “the question is no longer whether entities understand frontier AI risk”, but whether they can keep operating and recover when incidents unfold at greater speed.

ASIC's attention is also moving from governance to consumer outcomes. On 30 September 2026, ASIC announced a supervisory review of banks' new and proposed AI use cases and their impact on customers, particularly in customer-facing interactions, decision-making and lending. On 18 November 2026, it named “consumer harm arising from poorly implemented technologies, including AI and automated systems” as one of its nine enforcement priorities for 2027. ASIC will target poorly designed or inadequately supervised systems that produce misleading information, make inappropriate decisions or prevent customers from obtaining effective support. Licensees should expect ASIC to be looking for an AI test case in due course.

For APRA-regulated Licensees, AI risk is also a prudential issue, to be managed through existing prudential standards, including CPS 220 (SPS 220 for superannuation trustees), CPS 230 and CPS 234. AUSTRAC’s regime is equally technology-neutral: an entity using AI for transaction monitoring or customer due diligence remains responsible for the outcome.

AI and financial services obligations

Artificial intelligence, automated decision-making, and chatbots may improve the delivery of financial services, but do not alter a firm’s legal obligations.

Sarah Court, ASIC Chair, ASIC Annual Forum 2026, November 2026

Section 912A(1) of the Corporations Act 2001 (Cth) (Corporations Act) and s 47(1) of the National Consumer Credit Protection Act 2009 (Cth) (NCCP Act) impose materially similar general obligations on AFS and credit licensees respectively. Central among them is the obligation to provide financial services, or engage in credit activities, efficiently, honestly and fairly, which applies to all Licensees, including banks and insurers. Licensees must also maintain adequate risk management systems and have available adequate resources, including technological and human resources. For APRA-regulated Licensees, similar matters are governed by the prudential standards discussed above. These obligations are principles-based and technology-neutral: they apply whether services are delivered by people or by AI.

Courts have already found that technology risk failures can breach these obligations. In both ASIC v RI Advice Group Pty Ltd [2022] FCA 496 and ASIC v FIIG Securities Limited [2026] FCA 92, failing to adequately manage cyber security risks was held to breach the obligations in s 912A(1)(a) and (h) to provide financial services efficiently and fairly and to have adequate risk management systems. RI Advice resulted in declarations by consent. FIIG, which also admitted inadequate resources, was ordered to pay a $2.5 million penalty – the first civil penalty for cyber security failures under the general AFS licensee obligations. In RI Advice, the Court observed that cyber security risk cannot be reduced to zero, but it can be “materially” reduced to an “acceptable level” through adequate documentation and controls.

The same reasoning is likely to apply to AI: inadequate governance of AI can itself amount to a breach of the general obligations.

When AI gets it wrong

Generative AI raises a more immediate risk because it communicates directly with customers – answering questions, drafting correspondence and, increasingly, recommending products. Machine-generated words are not exempt from consumer protection and financial services laws. As discussed in our first article, in Moffatt v Air Canada 2024 BCCRT 149, Air Canada was held liable for negligent misrepresentation for inaccurate information provided by its chatbot; the tribunal rejected the “remarkable” argument that the chatbot was a separate legal entity. Australian courts are likely to take the same approach and have already attributed algorithm-driven representations to the business deploying them (see Trivago N.V. v Australian Competition and Consumer Commission [2020] FCAFC 185).

AI outputs are often non-deterministic, and models can “hallucinate” – confidently asserting things for which they have no basis. The test for misleading or deceptive conduct is objective: the question is whether the output, as the business’s own conduct, was misleading or likely to mislead, regardless of intention. A lack of adequate human oversight will not be an excuse and may in fact compound the exposure, supporting claims for breach of the general conduct obligations.

Algorithmic pricing

Not all AI disputes will involve generative AI. Conventional algorithmic systems already create significant exposure, as the Robodebt scheme demonstrated for the Government. In financial services, algorithmic pricing has already attracted ASIC enforcement action and class actions.

In 2023, the Federal Court imposed a $40 million penalty on Insurance Australia Limited (IAL) for false or misleading representations to more than 600,000 customers. IAL's pricing algorithm limited how far renewing customers' premiums could fall, so promised loyalty and no-claims bonus discounts were not passed on in full. The Court also found that IAL breached the efficiently, honestly and fairly obligation.

AI advice: information, general advice or something more?

Customer-facing AI can give rise to advice risk even where it is designed only to provide information.

The Corporations Act broadly defines 'financial product advice' as a recommendation or opinion intended, or reasonably regarded as intended, to influence a product decision. ASIC has confirmed that digital advice, i.e. advice provided through algorithms and technology without direct human intervention, can constitute general or personal advice and is subject to the same regulatory framework as human-delivered advice – including the best interests duty where personal advice is given to retail clients.

The legal test therefore does not turn on whether a human generated the communication. A chatbot designed to provide information may start comparing products or suggesting one is more suitable. Its legal character will depend on what the system says across the full range of its outputs, and design intent is unlikely to be decisive.

Consumers are increasingly turning to AI for financial guidance: ASIC's Moneysmart research found that 18% of Gen Z Australians were using AI platforms for financial information, and 64% said they trusted them for money advice. Harmful financial advice is also an ASIC enforcement priority for 2027.

Personalisation, vulnerability and automated decisions

Cutting-edge technology can't leave your customers bleeding. 

Joe Longo, then ASIC Chair, ABA Banking Conference, July 2025

ASIC has identified harms that AI can create or amplify, including bias, discrimination, misinformation, exploitation of vulnerability and erosion of trust. Those harms become more acute when AI makes or influences decisions about individual customers, and increasingly sophisticated personalisation multiplies them.

An AI system might flag customers experiencing financial vulnerability so extra support can be offered. The same capability could identify customers less likely to switch, more willing to pay, or less likely to complain - differentiating customers in ways neither the customer nor the institution fully understands. The legal question is not only whether a prediction is accurate, but whether the way it is used is consistent with the Licensee’s representations and conduct obligations.

Liability may arise where the use of customer insights moves from better understanding customers to exploiting them, engaging the prohibitions on misleading and unconscionable conduct, and anti-discrimination, privacy and licensing obligations. The inquiry may extend to the inputs and proxies used, how the model was trained and tested, and what the business knew about patterns emerging from its operation.

For credit licensees, this scrutiny has a statutory footing. The NCCP Act's responsible lending obligations require inquiry into and verification of the consumer’s circumstances, and an assessment of whether the credit is unsuitable. Automated tools are not inherently non-compliant, but the obligations concern process as well as outcomes: a model that cannot show what it took into account leaves the Licensee unable to show its process was reasonable.

In April 2026, Money3 was ordered to pay $1.55 million after its analysts applied arbitrary minimum expense figures without checking them against bank statement data it already held. That failure was human, but automation can make the same shortcut faster, larger and harder to detect.

ASIC has also said that access to human review is “particularly important where customers are making a complaint, responding to a scam, pursuing an insurance claim, or seeking financial assistance”. Licensees should check not only what their automated systems decide, but whether customers can reach a person when they need to.

The new privacy disclosure requirement for automated decisions noted above is also likely to capture many credit decisions – for example, where a computer program approves or rejects loan applications.

Third-party AI: outsourcing the technology, not the regulatory exposure

Most Licensees will buy AI embedded in existing platforms or build on third-party models. ASIC found that many relied heavily on third parties for their AI models, but not all had appropriate governance arrangements to manage the associated risks. A Licensee remains responsible for its licensed services regardless of who built the underlying system.

If an AI system causes widespread customer harm, disputes with the developer or supplier may also follow. These are likely to turn on what the supplier said about the system’s capabilities, what it knew about how the Licensee would use it, who was responsible for testing and operating it, whether changes were disclosed, what information rights the Licensee had, and how any indemnities and liability caps apply.

Where many Licensees rely on the same provider or model, a single failure can spread across the sector – a risk APRA and ASIC have expressly flagged.

The evidentiary problem

AI also changes the evidentiary exercise if something goes wrong.

Human decisions are typically reconstructed from documents, emails and witness evidence. AI-assisted decisions require a different record: which model and version were running, what data and instructions applied, what output was generated, whether a human reviewed it and what monitoring followed. In an ASIC investigation, Licensees should expect compulsory notices to target exactly these matters, including model documentation, validation reports, system prompts, audit logs and board reporting.

Licensees should therefore approach AI governance on the assumption that a dispute will arise: governance records will be key evidence of what the business knew, which risks it identified and whether its controls worked. Licensees that can use those records to show their controls were adequate, risk-appropriate and effective in practice will be far better placed.

Courts are also beginning to direct parties to consider AI in running litigation itself.

Steps Licensees can take now

While AI risk cannot be eliminated, Licensees can take practical steps to mitigate exposure:

  1. Map AI use cases to your obligations, risks and controls

    Identify where AI is used across the business, including tools embedded in third-party products, and match each use case to the obligations it engages: general obligations, misleading conduct, advice, responsible lending and privacy. For each, record the key risks and the controls that address them, and identify any gaps.

  2. Designate clear responsibility for AI risk

    Assign individuals clear duties, with reporting lines to a board committee responsible for AI governance and risk oversight. For entities subject to the Financial Accountability Regime (FAR), consider how AI risk maps to accountable persons' responsibilities.

  3. Keep a human in the loop for material decisions

    AI may inform credit, claims or investment decisions, but human review should be required before material decisions reach a customer. For credit, make sure models support, not replace, the inquiry and verification the NCCP Act requires.

  4. Test customer-facing AI against the advice boundary

    Test what chatbots and digital assistants say, not just what they were designed to do, and put guardrails in place to stop factual tools drifting into recommendations or personal advice.

  5. Build the evidentiary record now

    For significant decisions, particularly those affecting customers, keep a record of the model and version used, the data and instructions it relied on, what it produced and any human review. Keep records on the basis that ASIC or a court may one day ask to see them.

  6. Govern third-party AI through contract

    Before engaging a supplier, obtain independent validation and secure service-level agreements, notification of model changes, audit and information rights, and appropriate indemnities. ASIC has identified applying the same expectations to third-party models as to in-house models as better practice.

  7. Train staff and monitor continuously

    Make sure staff using AI understand its limits and embed systems and processes to monitor outputs and trends against fairness and accuracy benchmarks as models evolve.

  8. Test against customer outcomes
    Assess AI-enabled products, communications and decisions by reference to customer outcomes, including for customers experiencing vulnerability. Make sure complaints and internal dispute resolution processes can identify and escalate AI-related issues.

Looking ahead

Australia’s AI regulatory framework is still developing, but existing licensing, conduct and consumer protection obligations already reach much of what Licensees do with AI. As AI moves from assisting employees to communicating with customers and making decisions, the legal questions will move beyond governance to conduct itself – the character of the service, the representations made, the adequacy of controls and the evidence available to explain what happened. With AI now among ASIC’s enforcement priorities for 2027, the first test cases are likely to follow.

Our next instalments will examine AI washing and legal professional privilege, with further topics to come. If there is a particular topic you would like us to cover in future articles, please let us know.