ASIC's Focus on Banks' Use of AI: What Records to Keep for Customer-Facing AI

Regulatory corporate plans are often concise, and that brevity leaves considerable room for interpretation.

In the text of its 2026–27 Corporate Plan, published on August 26, 2026, the Australian Securities and Investments Commission (ASIC) devotes essentially one item to banks’ use of AI.

To draw practical implementation lessons from that single line, it is necessary to read it alongside the regulator’s earlier, more detailed findings. This article explains that reading and the design questions it raises.

What the Corporate Plan Actually Says About Banks’ Use of AI

The Corporate Plan sets out five strategic priorities. Under the first of them, “Improve consumer and small business outcomes,” one of the focus areas is the impact of AI use.

The only activity specifically addressing banks’ use of AI reads as follows:

Use of AI — We will monitor increased use of AI by banks to identify harms or risks to consumers.

That is very nearly everything the Corporate Plan says about banks and AI. Specifics such as review criteria, logging requirements, testing across customer groups, or how accountable persons should be assigned do not appear in the document.

A note at the foot of the page states that expected timeframes exceed one year unless otherwise specified, suggesting that ASIC expects this work to continue beyond a single annual cycle.

There is one other passage. In the section setting out its operating environment, ASIC states that AI is transforming that environment, and that weakening information quality, reduced accountability and AI technologies can increase exposure to misleading and unreliable financial information outside traditional firm channels.

What that passage names, though, is information circulating outside traditional firm channels. Because it is not confined to financial institutions’ own customer-facing AI, reading it directly as a requirement on banks’ customer-facing AI goes further than the text supports.

Under the fifth strategic priority, which covers market integrity, there is a separate item on assessing and addressing AI-driven manipulation, deepfakes and misinformation. That is a different question from customer-facing AI.

What the Same-Day Media Release Added

On the same day as the Corporate Plan, ASIC published a media release titled “ASIC sets plan to be easier to deal with, harder to avoid” (26-200MR).

It states that ASIC will lift its focus on artificial intelligence, and names three things: how banks use AI in customer-facing services, how AI may affect consumers and investors, and how AI-driven manipulation, deepfakes and misinformation could affect market integrity.

The qualifier “customer-facing services,” which does not appear in the Corporate Plan itself, comes from this media release.

The release also quotes ASIC Chair Sarah Court: “AI can improve services, productivity and decision making, but its use must not weaken accountability or consumer and investor protections.”

That is the full extent of what the two primary sources explicitly state.

What the Corporate Plan Does Not Say

Some third-party commentary summarized the announcement in three points: that banks should demonstrate that their AI services do not produce misleading information, obscure accountability, or disadvantage particular customer groups.

That is a reasonable interpretation, but that set of three appears in neither the Corporate Plan nor the same-day media release in that form.

For that reason, this article keeps what is stated explicitly in ASIC’s primary materials separate from the analysis MIF draws from them.

A second point that can easily be misunderstood is the status of the document. The Corporate Plan is a planning document setting out supervisory focus and intended activities; it did not in itself create new AI-specific obligations.

ASIC itself stated in its 2024 Report 798 that the regulatory framework for financial services and credit is technology neutral, and that licensees need to consider their existing regulatory obligations before deploying AI.

The Corporate Plan should therefore be read not as announcing new regulation, but as identifying where ASIC intends to concentrate its attention under existing obligations.

What Report 798 Adds

To identify practical implementation implications, the Corporate Plan’s single line must be read alongside ASIC’s earlier, more detailed primary materials.

On October 29, 2024, ASIC published Report 798, “Beware the gap: Governance arrangements in the face of AI innovation.” It is a review of 624 AI use cases collected from 23 AFS and credit licensees.

The report sets out the risks AI can pose to consumers: unfair or unintended discrimination due to biased training data or algorithm design; incorrect information provided to consumers about products or services; manipulation of consumer sentiment or exploitation of behavioral biases; breaches of data privacy and security; and an erosion of consumer trust and confidence due to a lack of explainability, transparency and contestability.

ASIC’s interest is not confined to the accuracy of AI outputs.

The review also identified concrete gaps. Approximately half of the 23 licensees had specifically updated their risk management policies or procedures to address AI risks.

Only 12 of the licensees had AI policy documents, guidance or checklists referencing fairness or related concepts such as the risk of discrimination or bias; only 10 had documented requirements or principles about disclosure to consumers interacting with or affected by AI; and no licensee appeared to have implemented specific contestability arrangements for AI.

The report also states that weaknesses were observed in how licensees provided meaningful human oversight, and in how they monitored for and responded to unexpected model outputs.

The report gives examples of licensees weighing business risk more heavily than the impact on consumers.

In one example, licensees identified the risk of an incorrect model output but noted that the consumer could contest it or that staff could override it — without considering the harm if the output caused a consumer to abandon the transaction altogether, potentially without knowing they could contest it.

In another, licensees using overseas-developed models for identity verification identified the business risk of failing to identify fraud, but did not identify the potential for some groups to be disproportionately impacted if those models had not been adequately trained on data representative of their Australian customers.

Responses to unexpected outputs also differed. The report contrasts a better-practice example—in which a licensee conducted root-cause analysis and investigated consumer impact—with a poorer practice, in which another licensee merely adjusted model thresholds to bring outputs within its business risk appetite, without investigating the root cause or assessing the impact on customers.

The report closes with a set of questions for licensees. Under accountability, it asks who is accountable for AI use and outcomes, at model level and overall, whether those people get the reporting they need to do their job, and how consumer outcomes from AI are being measured.

If the one line in the Corporate Plan is a statement of what will be watched, Report 798 is a casebook of what was actually found wanting.

Distinguishing Explainability, Responsibility, and Traceability

From here on, this is our own reading of ASIC’s materials.

Discussions of customer-facing AI often collapse three distinct questions into the single idea of being “able to explain.”

Explainability refers to the extent to which a person can understand and explain which inputs or factors influenced a particular output, and how the model and the business rules led to the result.

Allocation of responsibility is a matter of organizational design: which decisions sit within whose remit, and who holds the authority and the duty to approve, reject or remediate.

Traceability means being able to establish from evidence which model, configuration, data, and business rules were used; who approved, reviewed, or corrected the output; and what was ultimately presented to the customer.

The three overlap, but none can substitute for the others.

ASIC has not set out a priority order between explainability and traceability. If anything, Report 798 lists the absence of explainability as a risk that erodes consumer trust in its own right.

In one of the report’s case studies, a scoring model used as an input to credit decisions was described in an internal report as a “black box with no ability to explain the variables in the scorecard or the impact they are having on an applicant’s score,” and was eventually replaced with a simpler model to ensure that the scoring outcomes and the model were explainable.

The case shows that the explainability of the model itself matters, not merely the records surrounding it.

MIF nonetheless takes the following position on implementation order:

First, ensure that the inputs, outputs, model version, referenced information, tool calls, and human review steps can be reconstructed after the fact. Then design an explanation method appropriate to the use case.

Without those records, there is no basis for testing whether an explanation is sound.

This is not ASIC’s official view. It is MIF’s design position, arrived at in light of the same materials.

Traceability in this sense does not mean that the system’s internal reasoning can be reproduced.

Generative AI operates probabilistically, so the same input does not necessarily return the same output.

Logs can help reconstruct the processing path and the conditions under which the system was used, but not the model’s true internal causal process.

What Should Be Verifiable After the Fact

What follows is not a list of ASIC requirements.

The table below is an illustrative “evidence package” for customer-facing AI, developed by MIF with reference to ASIC’s materials.

It should be read not as a prescribed configuration, but as a starting point to be adapted to the organization’s size, applicable regulatory requirements, and risk appetite.

Design target	What should be verifiable after the fact	Example implementation and operating controls
Basis for the answer	Which version of the relevant product information, contract terms, or fee schedule was used	Reference history, version control of product information, response logs
Conditions of use	Which model, system instructions, policies, and tool configuration were used to handle the request	Model inventory, configuration management, change history
Boundaries of the decision	What the AI was permitted to handle on its own, and what conditions required escalation to a human	Business rules, escalation criteria
Human involvement	Who approved, corrected, overrode, or rejected the output	Approval history, records of the staff involved, complaint handling records
Customer outcomes	Whether outcomes differ disproportionately across customer groups or usage contexts	Periodic monitoring, testing of consumer impact
Means of remediation	Whether affected customers can be identified and notified, and whether the error can be corrected and appropriate redress provided	Incident response, customer contact, redress procedures
Records governance	What is retained, for how long, who may view it, and when it is deleted	Logging infrastructure, access control, retention and deletion policy

Assessing customer outcomes requires particular care.

Looking at differences across customer groups does not mean sensitive attributes may be collected freely.

Organizations should first establish what data may lawfully be used and what privacy and anti-discrimination constraints apply, and then test whether outcomes differ disproportionately across customer groups or usage contexts.

The more useful measures are not aggregate averages alone, but concrete customer outcomes: rates of incorrect guidance, escalation, refusal, completion, and complaints, together with the time required for correction or redress.

The right-hand column mixes things that can be enforced in product or system settings with things left to internal operating rules.

Keeping logs and setting escalation conditions can be enforced by the system itself, but making accountability actually function comes down in the end to role definitions and day-to-day operations.

The table should not be treated as a one-off exercise. Distinguishing controls that are technically enforced from those that depend on operating practice reduces the risk of discovering gaps only after deployment.

More Recordkeeping Creates New Risks

Emphasizing traceability can encourage the assumption that more records are always better. Report 798, however, also identifies breaches of data privacy and security as consumer risks.

Retaining detailed customer inputs, referenced documents, and staff actions expands what can be reviewed after the fact, but it also increases the volume of sensitive data held and the potential impact of a breach.

Conversation data retained for complaint handling can also magnify the impact of a breach.

Recordkeeping design should therefore define, at the same time, what is retained and for what purpose, how long it is kept, who may access it, what is masked, and how and when it is deleted.

“Records governance” sits at the end of the table above because that row is a precondition for the others.

Keeping everything “just in case” is not sound traceability design; it may simply accumulate a different category of risk.

ASIC Also Subjects Its Own AI Use to Approval and Oversight

The Corporate Plan also describes ASIC’s own use of AI.

It states that ASIC will continue building its data and AI literacy and will enable AI adoption across the organization, including generative AI tools, and that it will use AI responsibly in accordance with its AI Policy and AI Assurance Framework.

Among its responses to enterprise risk, ASIC identifies an AI governance and assurance framework—including an AI Board—to provide independent approval and oversight of AI use cases.

ASIC’s use of an independent approval and oversight mechanism alongside its expansion of AI is one indication that it does not treat AI governance as a matter of model performance alone.

However, ASIC’s internal arrangements should not be treated as requirements for the firms it regulates.

Conclusion

In its 2026–27 Corporate Plan, ASIC states that it will monitor the growing use of AI by banks to identify harms or risks to consumers. The same-day media release makes clear that this focus includes AI used in customer-facing services, and reiterates that AI must not weaken accountability or consumer and investor protections.

Read alongside Report 798, the resulting picture is broader than model accuracy alone. Explainability, fairness, responsibility allocation, human oversight, ongoing monitoring, documentation, contestability, and remediation all need to be considered together.

MIF translates the implementation side of this into a practical priority: preserving enough evidence to reconstruct how an output was produced and handled. That is MIF’s design position, not an ordering established by ASIC.

The practical lesson when reading a short planning document is to identify both what it says and what it leaves unstated, then consult the regulator’s earlier, more detailed findings. Before customer-facing AI goes live, an organization should know who is responsible for explaining an adverse outcome, which records will support that explanation, and how the error will be corrected.

Sources and Verification

This article was prepared using information reviewed as of September 2026. It draws primarily on the PDF text of the Corporate Plan 2026–27 published by the Australian Securities and Investments Commission (ASIC) on August 26, 2026, and on the media release “ASIC sets plan to be easier to deal with, harder to avoid” (26-200MR) published the same day.

What the Corporate Plan itself states is a commitment to monitor increased use of AI by banks to identify harms or risks to consumers.

Statements about customer-facing services, accountability, and consumer and investor protections draw on the media release published the same day.

Statements about incorrect information, unfair treatment arising from bias, explainability, transparency and contestability, meaningful human oversight, and ongoing monitoring and documentation draw on Report 798, published by ASIC on October 29, 2024.

The definition of traceability used in this article, the proposed evidence set, and the distinctions drawn among traceability, explainability, and responsibility allocation reflect MIF’s analysis.

ASIC’s supervisory approach and any resulting enforcement outcomes may evolve over time. Please refer to ASIC’s official announcements for the latest information.

Key References

• Australian Securities and Investments Commission, ASIC Corporate Plan 2026–27, August 26, 2026.
• Australian Securities and Investments Commission, media release ASIC sets plan to be easier to deal with, harder to avoid (26-200MR), August 26, 2026.
• Australian Securities and Investments Commission, Report 798: Beware the gap: Governance arrangements in the face of AI innovation, October 29, 2024.