CommBank's agentic AI fraud analyst: an in-house agent that drafts detection rules, built in three months, now behind three quarters of its card fraud rules

Last Updated: September 21, 2026

The bottleneck was never detection. It was writing the rule.

Every large retail bank already runs machine learning over payments. CommBank has been doing it since its Customer Engagement Engine launched in 2015. The bank's fraud protection systems monitor more than 80 million signals each day, including transactions, card and online payments, and interactions with digital banking channels.

The constraint sits one layer down. Anomaly scores do not stop money moving; rules do. And rules are artefacts written by people — a fraud analyst notices a new typology, characterises it, drafts a condition, tests it against historical traffic, argues about the false-positive cost, and ships it. That cycle is measured in days or weeks. Scam typologies mutate faster than that. The gap between a pattern appearing in the data and a control existing to intercept it is where customer losses accumulate.

That framing explains why CommBank aimed its agent at rule production rather than at yet another classifier.

What CommBank actually built

The bank describes an agentic system that runs continuously over transaction and payments data and does four things in sequence: identify a suspicious pattern, assess its severity, analyse the surrounding context, and propose a new detection rule. The proposal then enters the existing human approval path.

When suspicious patterns are identified, the system quickly assesses their severity, analyses context, and proposes new detection rules to help intercept them.

— James Roberts, Executive General Manager, Fraud and Scams, Commonwealth Bank, CommBank Newsroom

Two details separate this from a demo. First, it is always on — Roberts states that the agent operates around the clock, continuously monitoring activity and adapting to emerging threats. Second, it is gated: CommBank says new detection rules are reviewed and approved by the bank's fraud analytics team prior to implementation, a process it names explicitly as human-in-the-loop oversight.

The build was internal. CommBank states its in-house data science and engineering teams developed the system in three months, with ongoing testing demonstrating strong fraud detection outcomes. The investment sits inside the bank's stated $1 billion annual commitment to protecting customers from fraud, scams, cyber threats and financial crime.

The numbers that exist — and the ones that do not

CommBank's headline claim is adoption, not attribution: the agent has contributed to developing or updating three quarters of the bank's card fraud rules, which are used to identify potential fraud. Separately, the bank reports that its fraud detection technology played a role in helping to reduce fraud losses by over 20% in the first half of FY2026 compared with the first half of FY2025, a figure it repeats in its Our Approach to Adopting AI report announcement and in a Business Council of Australia case story.

Those are two different claims, and it matters that they are not the same claim. Read literally, the loss reduction is credited to the whole detection estate over a half-year that ended in December 2025 — before the agent was publicly announced in April 2026.

For scale context on how far agentic patterns have spread inside the bank, CBA's FY26 results presentation reports an 86% agentic messaging resolve rate — the share of customer conversations initiated through the agentic chatbot channel resolved without a human-assisted servicing pathway — and says roughly 80% of staff actively engage with AI platforms including ChatGPT Enterprise or Copilot.

The governance is the product

The most transferable part of this deployment is the thing that sounds least impressive: the agent proposes, humans dispose. That single choice changes the risk profile from 'autonomous system alters live payment controls' to 'assistant accelerates the slowest analyst task'. The worst-case failure becomes a wasted review, not a wrongly blocked customer.

CommBank's public position on this is consistent across documents: AI models used across the bank are governed by its risk management frameworks and policies, with clear human accountability for outcomes. Its Group AI Policy sets six principles covering environmental and social impact, fairness, transparency, privacy and data protection, reliability and security, and accountability.

As Australia’s largest bank, trust is fundamental to how we use AI. Our approach is focused on our risk management foundations and guided by our AI principles.

— Alex Matthews, Executive General Manager and report lead, Commonwealth Bank, CommBank Newsroom

Why three months was possible

A three-month build for a production fraud capability at a systemically important bank is not a story about model quality. It is a story about the decade of platform work that preceded it. The agent had somewhere to run, data to read, a rule format to write into, and a review team to write for.

The Anthropic announcement is a useful marker of intent even though it does not confirm what runs under this agent.

Our teams look forward to working closely with CBA's engineers and data scientists to explore innovative applications of our technology, particularly in critical areas like fraud prevention and customer service enhancement.

— Krishna Rao, Chief Financial Officer, Anthropic, CommBank Newsroom

What the public record does not tell us

Being candid about the gaps is more useful than inflating the win. CommBank has not published: the incremental loss reduction attributable to the agent; the precision of agent-drafted rules versus analyst-drafted rules; how many proposals analysts reject; whether candidate rules are simulated against historical traffic before human review; or any latency figure for pattern-to-control time, which is arguably the metric the whole system exists to improve. The FY26 full-year presentation does not restate a loss-reduction percentage at all.

Scope is also narrower than headlines suggest. This agent works on card and payments fraud signals. It is not an origination-fraud control: coverage of the April announcement noted the bank was at the same time investigating a suspected home loan fraud scheme reported at up to $1 billion, involving falsified borrower information. Agentic detection in the payments rail says nothing about document fraud at the point of application — a distinction worth holding onto before generalising 'AI solved fraud'.

What an integration team should take from this

Pick the artefact, not the department. CommBank did not build an 'AI fraud platform'; it built a generator for one specific, reviewable artefact that already had an owner and a deployment pipeline. That is the pattern that ships in a quarter. Our case study library and notes on workflow automation keep returning to the same shape.

Instrument the gate. The approval step is not overhead — it is your training signal, your audit trail and your quality metric. Log it from day one.

Check your data latency before your model choice. An agent proposing controls against stale data proposes stale controls. CommBank's core banking migration to cloud is cited in the same release as the thing enabling seamless access to rich, real-time data; see our research notes on why substrate beats model selection in operational deployments, and model comparisons for where the remaining differences actually bite.

Finally, publish the denominators. The most credible thing in CommBank's disclosure is the specificity of the small claims — three months, three quarters of card fraud rules, review before implementation. The least credible is the one that reads biggest. If you are building the equivalent internally, decide now which numbers you will be able to defend. If you want a second pair of eyes on that, start here.