Australia's APRA warns banks of gaps in AI agent oversight

Australia's prudential regulator APRA has warned financial companies that AI agents are being rolled out faster than controls over them are being built. The warning targets banks and pension fund managers, which already use such systems both in internal processes and in work with customers.

The APRA signal

APRA said that in late 2025 it conducted a targeted review of several large regulated organizations to assess how they are adopting AI. This is formally not about banning the technology or pausing experiments. The regulator looked at something else: how well companies understand where exactly their AI systems operate, who is responsible for them, and what control mechanisms apply after launch.

The very fact of a separate review shows that the topic has moved out of the pilot stage and become a question of operational and supervisory risk.

The warning comes as financial institutions expand their use of AI in two directions at once. The first is internal operations: document analysis, employee support, automation of routine decisions. The second is customer scenarios: assistants, responses to queries, service support and other touchpoints where a mistake affects not only back-office efficiency but also money, trust and rule compliance. For the financial sector, that is enough to stop AI governance from being a purely technical matter.

Where the gaps were found

APRA's key point is simple: practices for governing AI agents and verifying their reliability are currently weak. In other words, companies launch new tools but do not always manage to put clear rules around them.

By assurance, it makes sense here to mean not a marketing promise but a set of checks: how the system is tested before release, how errors are tracked after launch, how the model's limitations are recorded and how disputed decisions are escalated. Without this layer, even a useful agent quickly turns into an opaque risk.

In essence, the regulator is steering the market toward several basic questions:

• who approves the scenarios in which an AI agent may make or propose a decision at all;
• how the company verifies the model's accuracy, robustness and limits before moving it into production;
• which actions the agent is barred from taking without a human, especially in sensitive processes;
• who is responsible for monitoring, logging and reviewing errors after launch.

For banks and pension managers, this is not box-ticking bureaucracy. If AI takes part in handling customer requests, preparing recommendations, reviewing documents or routing transactions, weak control creates a chain of risks: a wrong answer to a customer, a breach of internal policies, a broken audit trail and, in the worst case, action from the supervisor. The closer an agent is to money and obligations to the customer, the costlier unclear responsibility becomes.

Why now

The reason for the firm tone is clear: the market is moving quickly from experiments to scaling. While AI was used sporadically, process gaps could be masked by manual checks and the enthusiasm of individual teams. But once such systems are built into permanent operations, governance problems stop being local. They touch model risk, compliance, customer protection and business resilience.

The regulator is essentially telling companies that speed of adoption can no longer excuse weak control.

This matters especially for AI agents, as opposed to plain analytical models. An agent does not just generate text or a forecast; it increasingly takes actions: it initiates steps, talks to the user, proposes the next scenario and passes data between systems. The higher the autonomy, the more important it is to define in advance the limits of authority, the requirements for checking results and the conditions under which a human must step in. For financial companies, this is no longer a question of convenience but of the controllability of the entire digital chain.

What this means

APRA's signal is unlikely to stop AI adoption in finance, but it will certainly make the market more cautious. The winners will not be those who bolted an agent onto a process fastest, but those who can show transparent rules, quality control and clear accountability for outcomes. For the rest, 2026 may be the year when interest in AI is, for the first time, measured not by the number of pilots but by the maturity of governance.