LLMs touch every part of the AML/CTF program. Governance has to follow.
Context for general readers: AML/CTF is the regime that requires banks, casinos, money transfer services, and a growing list of other businesses to identify their customers, monitor for unusual activity, and report suspicious matters to AUSTRAC, the Australian financial intelligence agency. The framework runs on the AML/CTF Act 2006 and the supporting AML/CTF Rules. The Tranche 2 reforms passed in late 2024 expanded coverage to lawyers, accountants, and real estate agents, with phased commencement through 2026 and 2027. The compliance burden is substantial, and the temptation to use AI tools to manage it is high.
This article covers the four highest-volume uses of large language models inside AML/CTF programs in Australian reporting entities, the supervisory considerations attached to each, and the governance posture that practitioners should be adopting now.
Where LLMs are operating in AML/CTF programs
1. Suspicious matter narrative drafting
When an analyst triages a transaction monitoring alert and decides to escalate, the suspicious matter report (SMR) submitted to AUSTRAC includes a narrative section. Large language models are now widely used to draft these narratives from structured alert data.
The supervisory consideration: the SMR is a regulatory submission. The Reporting Entity (not the AI tool) is responsible for its accuracy and completeness. Where the LLM output is treated as the final narrative without substantive review, several risks emerge. Hallucinated detail is the most obvious; less obvious is the systematic omission of contextual information that the structured alert data does not capture but that the analyst would have included by hand.
The governance posture: human-in-the-loop is not optional. The analyst must read the LLM output critically, verify it against source data, and add or correct narrative content. Documentation should evidence this review, not just the existence of the LLM workflow.
2. KYC document summarisation
Customer onboarding under enhanced customer due diligence requirements often involves reviewing complex source-of-wealth and source-of-funds documentation. Multi-page tax returns, corporate ownership structures, trust deeds. LLMs can summarise these efficiently.
The supervisory consideration: the AML/CTF Rules require the Reporting Entity to be reasonably satisfied of the information underlying customer due diligence. A summarisation that omits or misrepresents key information undermines this. The analyst's certification of customer due diligence completion must be defensible against the source documents, not against the LLM summary.
The governance posture: LLM summaries should be a starting point for analyst review, not a substitute for source-document examination on high-risk customers. The materiality threshold matters; for low-risk standard customers, LLM-assisted review may be appropriate. For politically exposed persons, sanctioned-jurisdiction nexus, or unusually complex ownership structures, the LLM is an aid, not an answer.
3. Transaction monitoring alert triage
Transaction monitoring systems generate large volumes of alerts. The vast majority are false positives. LLMs are increasingly used to triage these alerts, providing a recommended disposition (close, escalate, request further information) with a supporting narrative.
The supervisory consideration: AUSTRAC expects the Reporting Entity to operate a program that produces an appropriate volume of suspicious matter reports relative to its risk profile. An LLM-driven triage system that systematically biases toward closure can cause under-reporting; one that biases toward escalation can drown the financial intelligence system in low-quality reports.
The governance posture: LLM-driven triage requires periodic back-testing. A sample of LLM-recommended dispositions should be re-reviewed by a senior analyst, with metrics on agreement rates, false negatives (alerts the LLM closed that should have escalated), and false positives. Without this back-testing, the institution is operating its AML/CTF program partly blind.
4. Internal AML/CTF training and policy drafting
Many institutions use LLMs to generate AML/CTF training content, draft policy updates, and answer staff queries about AML/CTF obligations. This is generally lower-risk than the customer-facing uses, but it has a specific failure mode worth flagging: hallucinated regulatory citations.
LLMs are confident and articulate. They are also capable of generating plausible-sounding but inaccurate references to AML/CTF Rules, AUSTRAC guidance, and case law. Training content and policy updates that include such references can mislead staff and create unnecessary compliance risk.
The governance posture: LLM-generated content used in AML/CTF training or policy must have every regulatory citation verified against the primary source. This is straightforward but easy to overlook.
Cross-cutting governance considerations
Three principles apply across all four use cases.
Data residency and confidentiality
Most large language models are operated by US-based providers, with inference infrastructure that may not be located in Australia. The customer information passing through these systems is subject to Privacy Act obligations and, where the entity has banking secrecy obligations or specific contractual confidentiality, those as well.
The pattern emerging in major Australian institutions is to use LLM deployments with documented data residency commitments, audit logs, and contractual restrictions on use of customer data for model training. Reporting entities operating outside this pattern have a governance gap that AUSTRAC and the privacy regulator may both reach for.
Explainability and audit trail
AUSTRAC's risk-based supervision tests whether the Reporting Entity can explain why it took particular AML/CTF actions. An LLM-driven triage decision needs an audit trail that allows a supervisor (or an internal investigator following an external incident) to reconstruct what the LLM saw, what it recommended, what the human reviewer did, and why.
The technical capability for this audit trail varies considerably across LLM deployments. Reporting entities should test whether their current configuration produces an audit trail that would survive supervisory scrutiny.
Tranche 2 readiness
The Tranche 2 reforms expand AML/CTF obligations to legal, accounting, and real estate professionals through 2026 and 2027. Many of these new reporting entities are small or mid-sized firms that will reach for LLMs as a default compliance tool. The supervisory expectation will form quickly. Practitioners advising new reporting entities should build governance around LLM use into the AML/CTF program design from day one, not retrofit it after the program is live.
Privacy Act intersection
LLM use in AML/CTF programs intersects with the Privacy Act in several places. Customer data flowing through prompts is in scope of the Australian Privacy Principles. The use of inference logs containing customer data is a use and disclosure question. The cross-border transfer of customer data to an overseas-hosted LLM engages APP 8.
The OAIC's emerging position on AI is that the Privacy Act applies to AI processing in the same way it applies to other data processing. For AML/CTF programs, this means LLM deployment decisions need to satisfy both the AUSTRAC and OAIC frameworks. The two frameworks are aligned on the key points (transparency, control, accountability) but not always on the operational details. Where the LLM deployment cannot satisfy both frameworks, the entity has a compliance choice to make.
Sanctioned-jurisdiction LLM exposure
A specific risk category that has emerged: the use of LLM tools whose operating chain includes sanctioned-jurisdiction nexus. This is rare for the major enterprise LLM providers but worth a deliberate check. AML/CTF Rules expressly prohibit certain dealings with sanctioned jurisdictions, and the introduction of an LLM tool with such a nexus into the AML/CTF program would be self-defeating.
The practical action: the third-party assessment of any LLM tool used in the AML/CTF program should include an explicit sanctioned-jurisdiction nexus check, not only on the contracting entity but on the model provider, infrastructure provider, and any sub-processors.
Where LLMs are not yet appropriate in AML/CTF
A balanced view requires acknowledging the use cases where LLMs are not yet appropriate, regardless of governance maturity.
The first is the final certification of customer due diligence. The certification that the entity has met its CDD obligations is a regulated act. It must be made by a person with authority and accountability, not by an LLM. LLMs can support the underlying review; the certification itself sits with the analyst.
The second is the decision to file or not file a suspicious matter report. The decision to escalate (or not) is a judgement that the regulated entity makes, with the consequences sitting with the entity. LLM-driven triage can support the analyst's decision; the decision itself should sit with a human.
The third is high-risk customer relationship sign-off. Politically exposed persons, customers in sanctioned-jurisdiction nexus territory, and customers with complex beneficial ownership structures should have human-led customer due diligence, with LLM tools playing a supporting summarisation role at most. The risk of LLM error in these cases creates an asymmetric exposure that is rarely worth the efficiency gain.
These boundaries are not legally codified, but they reflect emerging practice in major Australian reporting entities. Practitioners building LLM governance frameworks for AML/CTF programs should articulate these boundaries explicitly rather than leaving them implicit.
Practical implications this quarter
For AML/CTF compliance teams, the four actions to prioritise:
- Map every LLM-assisted AML/CTF workflow against the AML/CTF Act, the Rules, and your program documentation. Each workflow needs a documented risk assessment and a documented control framework.
- Implement back-testing on LLM-driven triage and dispositioning. A monthly sample-based review is a sensible starting point.
- Verify the data residency and confidentiality posture of every LLM in use. Where the posture is inadequate, the use case should be paused or restructured.
- Build a citation-verification step into any LLM-generated training, policy, or guidance content. This is cheap to do and meaningfully reduces a common error mode.
Auditability and the AML/CTF program review
The AML/CTF Rules require periodic independent review of the AML/CTF program. For programs with significant LLM components, the independent review should explicitly cover the LLM use cases. The reviewer needs sufficient access to the workflows, prompts, outputs, and human review evidence to form a view on whether the program is operating as documented.
A practical pattern: the reviewer should be able to walk a sample of suspicious matters from initial alert through LLM-assisted triage and human review to disposition, with full audit trail. Where the audit trail is fragmented or relies on the LLM provider's logs (which the entity may not control), the reviewer should flag the gap.
This is consistent with broader good practice for AML/CTF programs, but the LLM dimension makes it materially harder to execute. The institutions that have built unified audit logging across human and LLM components of the workflow will pass the review more comfortably than those relying on the LLM provider's separate log infrastructure.
Direction of travel
AUSTRAC has not published AI-specific guidance, but the AUSTRAC Compliance Guide makes clear that the obligation is on the Reporting Entity to design a program proportionate to its risk profile and to operate it consistently with the AML/CTF Rules. The use of LLMs sits inside that obligation, not outside it.
The practical reality for AML/CTF teams is that LLMs are now operating at sufficient scale across reporting entities that AUSTRAC supervisory engagement on AI is increasingly likely. The institutions that will fare best are those that have documented governance, can show back-testing evidence, and can articulate a clear position on data residency and audit trail. None of that is technically difficult; it does require attention now, while the supervisory expectation is still forming.
Content disclaimer: This article is for general educational and informational purposes only. It does not constitute legal advice, regulatory guidance, or a substitute for professional compliance judgement. Regulatory obligations vary by entity type, licence, and circumstance. Always refer to primary source guidance from APRA, ASIC, or the relevant regulatory authority.
TheAICommand. Intelligence, At Your Command.