AI and the Scams Prevention Framework
Australia has moved from naming and shaming scam losses to legislating against them. The Scams Prevention Framework Act 2025, which received Royal Assent on 20 February 2025, creates an enforceable, principles-based regime that obliges designated businesses to actively prevent, detect, disrupt and respond to scams targeting their customers. The first sectors brought into scope are banks, telecommunications providers, and a defined slice of digital platforms, with the framework's obligations taking effect for those sectors from 1 July 2026.
For a regulated entity, this is a structural change, not a marketing campaign. The framework converts what used to be a reasonable-efforts reputational expectation into a set of duties with regulator oversight, mandatory reporting, external dispute resolution through the Australian Financial Complaints Authority, and civil penalties that reach into the tens of millions of dollars per contravention. The compliance bar is no longer "did you have an anti-scam page on your website". It is "can you demonstrate, on the evidence, that your systems detected and disrupted scam activity across your customer base in close to real time".
That phrase, in close to real time, is where artificial intelligence stops being optional. The volume and velocity of modern scams sit well beyond what a manual review team can process. This guide maps the six SPF principles to where AI genuinely earns its place, where it introduces fresh legal exposure under the Privacy Act and the Australian Consumer Law, and how a midsize regulated entity should sequence the build.
Why the framework exists: the loss numbers
The policy driver is straightforward. According to the National Anti-Scam Centre's Targeting Scams report for 2024, Australians reported combined scam losses of $2.03 billion in 2024, drawn together from Scamwatch, ReportCyber, the Australian Financial Crimes Exchange, IDCARE and ASIC. That figure was down 25.9 per cent on 2023, which the ACCC attributes to coordinated effort across government, law enforcement and industry. The report also recorded 494,732 scam reports for the year.
Two things follow from those numbers. First, losses are falling, which tells regulators that coordinated intervention works and is worth mandating. Second, two billion dollars in reported losses, with under-reporting widely acknowledged, is a large enough harm that government is prepared to legislate sector duties rather than rely on voluntary codes. The SPF is the legislative expression of that judgement.
The six SPF principles
The framework is built on six overarching principles that apply to every regulated entity, supported by sector-specific codes for banking, telecommunications and digital platforms that translate the principles into concrete obligations. Understanding the six is the foundation for any technology decision.
The Six SPF Principles
Each principle imposes a distinct operational demand. Govern is a board and executive accountability duty: there must be a senior owner and a documented strategy. Prevent and detect are intelligence and monitoring duties. Disrupt is an intervention duty, the ability to act on a suspicious transaction or message before the money or data leaves. Respond covers complaints handling and, critically, the prospect of compensation where an entity has failed to meet its obligations. Report is the intelligence-sharing duty that allows, for example, a bank to flag actionable scam intelligence to the ACCC, which can then alert other banks and platforms to disrupt the same scam elsewhere.
Note where the legal weight sits. The respond principle ties compensation to whether the entity met its obligations under the other five. That makes your detect and disrupt evidence not just an operational nicety but the determinant of your liability when a customer is defrauded.
Where AI does real work, principle by principle
AI is not a single product you bolt on to satisfy the framework. It is a set of capabilities that map to specific principles. The table below separates where machine learning genuinely outperforms rules and humans from where it does not.
AI Fit Across the SPF Principles
| Metric | Manual or Rules-Only | AI-Augmented | Improvement |
|---|---|---|---|
| Transaction monitoring | Static thresholds, high false positives | Behavioural anomaly detection per customer | Detect |
| Scam message analysis | Keyword blocklists, easily evaded | Pattern and intent classification at scale | Detect |
| Real-time intervention | Post-event review, money gone | Inline risk scoring before settlement | Disrupt |
| Intelligence sharing | Manual report compilation | Automated structuring of actionable signals | Report |
| Governance evidence | Periodic spreadsheets | Continuous control logging | Govern |
The detect principle is the strongest AI use case. A bank must monitor transactions for suspicious activity that may indicate a scam and identify actionable scam intelligence. Static rules, such as flagging every transfer above a fixed dollar amount, drown analysts in false positives and miss the scams that hide inside ordinary-looking amounts. Behavioural models that learn a customer's normal patterns and score deviation against known scam typologies, such as a sudden first-time payment to a newly added payee following a series of inbound calls, are materially better at separating real risk from noise. This is the same anomaly-detection discipline covered in our guide to AI-powered fraud detection for Australian business, applied to a now-mandatory context.
For telecommunications providers, the detect obligation focuses on analysing call and message traffic for patterns and indicators of a scam. AI classifiers that identify the linguistic and structural signatures of scam SMS, or the calling patterns of number-spoofing campaigns, operate at a volume no human team can match.
For digital platforms, a sector-specific obligation is to verify that advertisers hold appropriate licences before they advertise high-risk products such as financial services. AI assists with document verification and anomaly detection in advertiser behaviour, but the verification decision itself remains a controlled, auditable step.
The disrupt principle is where AI converts detection into prevented loss. The value of a risk score is realised only if it can be acted on inline, before a payment settles or a message is delivered. That demands low-latency scoring integrated into the transaction or messaging pipeline, not a nightly batch job. The architecture matters as much as the model.
A worked, hypothetical view of the detect-to-disrupt loop
Consider a typical midsize mutual bank with several hundred thousand customers and a small financial crime team. Treat the following as an illustrative model, not a claim about any specific institution.
Detect-to-Disrupt Loop
The economic logic is that intercepting a single high-value scam payment can offset a meaningful share of the model's running cost, while the framework makes the alternative, doing nothing demonstrable, a source of both penalty exposure and compensation liability. The point of the hypothetical is the shape of the loop, not a promised return.
What the Detect-to-Disrupt Loop Protects
The risks AI introduces: privacy, fairness and false positives
Deploying AI to satisfy one law does not exempt you from the others. Three risks deserve board attention before any model goes live.
The first is privacy. Behavioural scam detection works by profiling customer activity, which is personal information under the Privacy Act 1988. Building and running these models must be reconciled with the Australian Privacy Principles, particularly around the purpose of collection and the handling of data shared between entities under the report principle. The reforms to the Privacy Act now moving through implementation, including transparency requirements for automated decision-making, raise the bar further. Our guide to automated decision-making and privacy policy in Australia sets out what disclosure you will likely owe customers when a model influences a decision that affects them.
The second is fairness and false positives. A model that wrongly holds a legitimate payment, or that disproportionately flags certain customer cohorts, creates its own consumer harm and potential exposure under the Australian Consumer Law. Every customer interaction your systems drive, including an automated scam warning, is a representation that must not be misleading. We unpack that interaction between automated systems and the ACL in the ACCC, consumer law and AI. A scam-prevention control that blocks lawful commerce or misinforms customers trades one problem for another.
The third is governance and explainability. The govern principle requires senior accountability, and a regulator will expect you to explain why a model held a payment or cleared one. Opaque scoring with no audit trail fails that test. This is the same human-accountability and audit-logging discipline we describe in our framework for deploying AI agents responsibly: a human owns the decision, and the system records why.
Is Your AI Scam Control Defensible?
Build, buy or partner
Few midsize regulated entities will build a scam-detection stack entirely in-house, and few should. The realistic choices are a specialist vendor platform, an internal build on cloud machine-learning services, or a hybrid where a partner accelerates the build while your team retains the governance. The decision turns on data sensitivity, in-house capability and how much of the model logic you need to control and explain.
Whatever the path, data sovereignty is a live consideration. Scam-detection models train on highly sensitive financial and behavioural data, and where that data is processed and stored matters for both the Privacy Act and customer trust. The principles in our guide to data sovereignty and offline AI for Australian organisations apply directly: understand what leaves your environment, and never feed regulated customer data into a tool whose data-handling terms you have not verified.
Tuning the false-positive problem
The single biggest practical failure mode of an AI scam control is not missing scams. It is stopping legitimate activity. A model that holds too many genuine payments, or that warns customers so often the warnings become background noise, erodes trust, generates complaints, and trains customers to click through every alert, which defeats the disrupt principle entirely.
Tuning is therefore not a one-off calibration but an ongoing operational discipline. A defensible programme tracks the rate at which holds turn out to be genuine scams against the rate at which they turn out to be legitimate, segments that performance across customer cohorts to check for unfair impact, and feeds confirmed outcomes back into the model so it improves. The govern principle effectively demands this: you cannot claim a senior-accountable anti-scam regime if no one owns the question of whether the system is helping or harming.
This is also where the interaction with the respond principle becomes concrete. When a customer disputes a scam loss, the framework ties any compensation to whether you met your obligations. A well-tuned, well-evidenced detection system is your defence. A blunt one that you cannot explain is the opposite, because it shows you were monitoring without genuinely controlling. The quality bar is not activity, it is demonstrable effectiveness.
A realistic sequencing plan
The framework's obligations apply to the first sectors from 1 July 2026, so for designated entities the timeline is immediate rather than aspirational. The following phased approach assumes a midsize entity starting from a rules-based baseline.
Sequencing an SPF-Ready AI Capability
The order is deliberate. Govern comes first because it sets accountability and tells you which risks matter most. Detect precedes disrupt because intervening on a poorly tuned model creates customer harm. Reporting comes last not because it is least important, but because the intelligence you share is only as good as the detection that produces it.
What this means for entities not yet designated
If you operate a superannuation fund, a cryptocurrency exchange or another business that sits adjacent to the first designated sectors, the framework is built to expand. The government can bring further sectors into scope as scam patterns shift. Treating the six principles as a forward-looking design standard now, rather than waiting for a designation, positions you to comply on a shorter runway and signals to customers and partners that you take scam prevention seriously. The same applies to the broader financial-services compliance posture we cover in our guide to AI compliance under APRA and ASIC.
The bottom line
The Scams Prevention Framework turns scam prevention from a reputational nicety into an enforceable set of duties, and it does so at a scale and speed that only AI-augmented systems can meet. The detect and disrupt principles are where machine learning genuinely outperforms rules and human review. But deploying AI to satisfy the SPF does not suspend the Privacy Act, the Australian Consumer Law or the need for explainable, human-accountable governance. The entities that come out ahead will be the ones that treat the framework as a single, coherent control problem: detect well, disrupt in real time, and be able to prove, on the evidence, that they did.