The AI that helps a bank stop a scam before the payment

A payment scam does not always look unusual. A customer may authorize the transfer because someone persuaded them by phone, message, or fake advertisement. That makes bank prevention difficult: it is not enough to look for a stolen card or a wrong password. The bank must identify risk signals without indiscriminately blocking legitimate payments.

Commonwealth Bank of Australia (CBA) is a useful case because it has brought that detection into a large-scale operation. In 2024, the bank said it processed and analysed more than 20 million payments a day. Across that flow, it combines verification technologies, detection models, customer alerts, and teams that investigate cases. AI is part of the system, not the only defence and not the sole final decision-maker.

From payment to preventive warning

The deployment uses several layers. NameCheck checks account details before a transfer and aims to prevent money from reaching the wrong account or one linked to a scam. CBA also introduced a model to detect potential investment-scam transactions in real time, identifying cases for its teams to investigate.

In November 2024, CBA said generative AI was helping it flag thousands of suspicious transactions and send 20,000 proactive warnings a day to retail customers through its app. A warning is not an accusation that a customer or merchant has committed fraud. It is a signal for a person to reconsider a payment and for the bank to apply its controls according to risk.

The value is timing. An authorized transfer can move from a victim’s account through intermediary accounts within minutes. If a system links information such as the payee, a payment pattern, or signs of a known campaign and warns before funds leave, the bank gains an opportunity to prevent a loss. It is not a promise to stop every scam. It is a way to focus attention where it may be needed most.

Results, and what they cover

CBA reported that customer scam losses fell by more than 50% in FY24 compared with the previous year. It also estimated that NameCheck prevented about A$40 million in scam payments and more than A$370 million in mistaken payments from July 2023 to June 2024; its results presentation groups these as more than A$410 million in prevented payments. The tool was used 57 million times during that period.

In its November update, the bank linked a 30% drop in customer-reported fraud to measures such as its generative-AI-powered suspicious-transaction alerts. This is a meaningful operational improvement: rather than waiting for a complaint, the bank tries to intervene before or during payment.

But attribution matters. CBA does not say that one model alone caused the lower losses. Its reporting includes specialist teams, customer education, intelligence sharing with other institutions, identity checks, policies for payments to certain cryptocurrency exchanges, and telecoms collaboration. The 50% figure describes the outcome of that package.

Balancing security and friction

A system designed only to detect the greatest possible number of risks could stop too many legitimate payments. Those false positives carry a cost: they delay a purchase, complicate an urgent payment, and may affect customers with atypical transaction patterns unevenly. A useful alert therefore needs a comprehensible explanation, review routes, and criteria monitored against real-world data.

CBA says its investment-scam model identifies potential cases for human teams to investigate. It also says it implemented 11 guardrails for customer-facing generative AI, covering risks from malicious inputs and misleading outputs. These are relevant measures, but they do not replace independent assessment of bias, privacy, and effectiveness required for responsible financial systems.

Privacy is another condition. Detection works precisely because it analyses payment and behavioural signals. That requires limiting data use to prevention, applying security, and ensuring that customers can understand what happens if a transaction is delayed or challenged.

A lesson for banking

The value of AI in fraud prevention is not assigning a sophisticated-looking label. It is improving the moment of intervention: before money becomes unrecoverable, without turning every payment into suspicion. CBA illustrates a useful architecture: models that prioritize risk, clear warnings, verification steps, and staff who decide how to act.

For other institutions, the lesson is not to copy a headline figure. It is to measure prevented losses, false positives, response time, and outcomes across customer groups separately. An anti-fraud tool earns trust when it reduces real harm and explains how it protects the customer without preventing them from doing business.

From warning to avoided loss

A prevention chain has several rates: payments analysed, warnings sent, warnings opened, transactions paused, cases confirmed and losses avoided. The CBA update provides volume and aggregate changes, but it does not publish a full false-positive matrix on that page. Without that denominator, a fall in fraud cannot become the precision of one model.

NameCheck also answers a different question: whether the entered name matches the account. Generative alerts analyse suspicious transaction signals. Human teams investigate. Grouping these layers as “an AI” erases where the benefit arose and where friction may appear. An audit should preserve the path from signal to decision.

Four measures a bank should publish

The first is detection among cases later confirmed. The second is the share of legitimate payments interrupted. The third is extra customer time. The fourth is net loss avoided after recoveries and operating costs. The NameCheck figures separate mistaken payments from scams, a necessary distinction if error correction is not to be presented as fraud detection.

A stable time reference is also needed. More alerts in one day do not imply greater risk when a threshold has changed. Comparing periods requires payment volume, channel mix, scam definition, model version and manual rules. Otherwise, an apparent improvement may come from less fraud, more friction or a new classification.

Human control cannot remain a ritual phrase either. Reporting should state what an analyst can do: release, hold, contact a customer, escalate or override; how much time is available; and what evidence appears. A score without operational reasons can turn review into automatic acceptance. Measuring disagreements and their outcomes shows whether the person adds judgement or merely signs the output.

Customer appeals belong in the measurement. A legitimate payment stopped briefly and one that leaves a person without access to money should not count as the same alert, because doing so hides harm. CBA’s public description needs resolution time, overrides, repeat cases and appeal outcomes beside it. That evidence lets a bank balance security and friction without rewarding a system merely for blocking more payments.

The transferable skill is to read the whole funnel. “Twenty thousand warnings” measures activity; “lower losses” measures a joint outcome. Confirmations, false positives, time and attribution are required between them. AI deserves credit only for the segment that evidence can connect to it. The remainder belongs to the complete prevention system.