25 of 27 decision-makers in banking say they use AI, but only 6 have set a risk appetite for it.

The Actuaries Institute and the UTS Human Technology Institute published AI Risk Management in the Financial Services Sector in September 2026.

What it says

The framework is deliberately not a new scheme. It sits on top of existing enterprise risk management, ISO 31000 and APRA prudential standards. It targets what the authors call the AI-delta: the ways AI amplifies existing risks and introduces new ones.

It is built around four questions:

  • Governance: the function deploying AI should not be the sole arbiter of the risk it creates.
  • Classification: AI risk is managed inside existing risk domains and tagged so aggregate exposure can be reported.
  • Quantification: specific loss events are estimated as financial distributions rather than ratings on a heat map.
  • Controls: controls are calibrated by use case and reversibility, including an authorization boundary for irreversible actions that the model cannot override.

Why it matters

APRA, Australia's prudential regulator, noted that most entities recognize existing prudential standards apply to AI risk, and few have operationalized governance in practice. ASIC reached a similar conclusion in 2024. It reviewed the AI governance arrangements of 23 licensees and found weaknesses in both governance and risk management practices.

The survey points the same way: 8 of 27 respondents place AI risks on the risk register. The authors treat the sample as directional, and the pattern matches both regulators.

Part of the gap is structural. Responsibility for AI operating safely tends to sit with technology executives, named by 19 of 27 respondents. The value of the technology sits with the business units that use it. Risks that depend on knowledge of a specific business application then carry too little weight in deployment decisions.

The gap also reaches the board. CPS 220 requires the board to set risk appetite and to satisfy itself that the risk management framework addresses material risks. When AI risk is managed only inside existing domains, with no tagging across registers, neither the board nor the regulator can see aggregate AI exposure.

The obligation already exists. What is missing is evidence that the framework operates on AI systems specifically: on the register, in the risk appetite statement, in the controls.

Key takeaways

  • AI-related risks can stay on existing registers, tagged so the board sees aggregate exposure without a separate AI register.
  • In AI triage, the highest dimension score sets the risk tier, not the average.
  • Human review needs a measurable output. A review process that produces no overrides is almost certainly not working.

Source:
· Actuaries Institute and the UTS Human Technology Institute — AI Risk Management in the Financial Services Sector — actuaries.asn.au, September 2026

What evidence would convince your board that your existing risk framework actually covers AI?

Mathieu Flamant
Founder · Tech Leader · mathieuflamant.com