A Banker Reads the Moment
Nuno Matos runs ANZ Group Holdings, one of Australia's four big banks, with around 40,000 employees. On Tuesday he stood at the Australian Financial Review Asia Summit in Sydney and said something chief executives of major banks rarely say out loud: the technology his industry is adopting may be moving faster than the people who built it understand, and he cannot promise his workforce what happens next.
Matos said AI is creating risks at a much higher pace than its own builders and developers were thinking, and pointed to the sudden public caution from the technology's leaders. The fact that Elon Musk, Dario Amodei and Sam Altman say they think they have to pull down the pace should tell us something very clear, he said. They are saying that the risk of this technology is well above what they expected.
Asked directly whether AI could bring large-scale job cuts at ANZ, Matos declined to rule it out. I don't know, and I think whoever says something that is certain will lie to you, he said. Nobody knows.
The Labor Math Banks Are Running
Banking has always been one of the more automatable industries, and the current AI wave is landing first on the work that sits between customers and systems: document checks, loan assessment, service calls, the processing layers where headcount lives.
Matos framed the risk in starker terms than most of his peers, saying that without enough guardrails and limitations, AI will lead to dangerous outcomes, and that it will attack infrastructure. The choice of verb matters. A bank CEO who says the technology will attack infrastructure is not talking about a productivity tool. He is describing a systemic risk that the industry's own adoption is helping to create.
The same summit supplied the counterpoint in numbers. Westpac, ANZ's rival, presented investor materials claiming five AI agents operating across mortgages and consumer finance are estimated to remove 250,000 manual activities and free up about 150,000 hours of banker capacity a year. The bank put the annual benefit at roughly 100,000 hours from automating payslip and bank statement checks in home loans and about 50,000 hours in consumer finance, figures it described as estimates.
Read together, the two presentations are the industry's internal debate in miniature. Westpac sells the arithmetic of hours saved. Matos warns about what happens when the saving compounds across an economy.
The Week AI Safety Moved to the Center
The summit remarks did not happen in a vacuum. In the days before Matos spoke, Musk, Amodei and Altman each made public statements about the pace of AI development, in some cases floating the idea of slowing model development, and President Donald Trump rebuked the caution, framing any slowdown as ceding ground to China. A senior Google DeepMind researcher warned in an exit post that AI may kill us all. US senators are weighing legislation that would require AI companies to commit to preventing catastrophe.
Matos's reference to those three names is the part of the speech most worth slowing down on. A bank CEO citing the founders' sudden caution is not agreeing with them; he is extracting from their statements the thing that matters to a lender: if the builders are surprised by the speed, the people deploying the technology downstream have even less basis for confidence. That is the risk-management reading of an AI-safety debate that has mostly played out in Silicon Valley and Washington.
The Infrastructure Line, Taken Seriously
The most unusual sentence in Matos's remarks was the prediction that AI will attack infrastructure. A banker using that word is not describing a chatbot writing emails. He is describing a world where automated systems, financial and otherwise, act at a speed and scale that outrun the controls built around them, the same concern that central bankers have raised about cyber risk and model risk converging.
For a bank, infrastructure means payment rails, market plumbing, the shared systems that competitors and counterparties depend on. ANZ is one of the institutions that runs that plumbing in Australia. A chief executive who warns that the technology his own bank deploys could attack the systems his bank operates is making a claim about systemic risk from inside the system, which is why the line traveled.
The Certainty Gap
Matos's refusal to put a number on AI's labor impact is the most honest part of the whole episode, and it is worth contrasting with the standard corporate script. Bank executives typically pair every automation announcement with a headcount plan and a reskilling program, as if the technology were a scheduled renovation. Matos said the opposite: nobody knows, and anyone who claims certainty is lying.
There is a practical consequence to that admission. Labor markets, unions and regulators plan around what employers say. A bank that says it cannot predict the size of AI-driven cuts is also saying it cannot promise its workforce a stable headcount, which is precisely the kind of uncertainty that employment in a heavily regulated industry is supposed to be insulated from. The gap between what banks know and what they can promise is now on the public record, in the words of one of the industry's own chief executives.
What Banks Can Actually Do About the Risk
Matos's solution sketch, guardrails and limitations, is the same phrase every bank now uses, and it conceals as much as it reveals. For a bank, AI guardrails are concrete things: model risk management frameworks, human approval thresholds for automated decisions, limits on what an agent may do without review, and the regulatory capital and audit machinery that surrounds all of it. The Australian Prudential Regulation Authority has made clear it expects banks to treat AI risk like every other operational risk, which means documented controls and a board that owns them.
The uncomfortable part of the CEO's message is that guardrails slow deployment, and deployment is where the savings are. A bank that pauses to validate every model is a bank that captures the productivity gains more slowly than its competitors, which is why the industry's AI race has the same shape as its risk debate: everyone wants to be second-fastest, and nobody wants to be slowest. Matos is telling his shareholders that ANZ will run the race with its seatbelt on, which is the prudent position and the expensive one at once.
For the rest of the economy, the Australian episode is worth watching as a preview. Banking is a concentrated industry in a mid-sized economy with a highly educated workforce and strong regulators, the cleanest laboratory for AI's labor effects available. If the Westpac hours materialize and ANZ's cuts do not, the difference will be measurable in two annual reports. If the reverse happens, the guardrail argument loses its strongest test case.
Primary sources
- Business Standard on Matos's remarks at the AFR Asia Summit for the quotes, the Westpac figures and the context, from Bloomberg's Richard Henderson.
- Westpac investor centre for the AI productivity estimates presented at the same summit.