APRA names AI as its top concern.
APRA's System Risk Outlook names AI its first concern. The same week, Anthropic and OpenAI both move agent execution inside the customer's data perimeter.
Hi folks,
This week the conversations we've been having privately with GPs since January started showing up in regulator-grade language. APRA's quarterly Outlook landed Thursday with AI named as the first thing the regulator is focused on, ahead of the Middle East and everything else.
The one thing
APRA names AI as its top concern.
- APRA's System Risk Outlook landed Thursday 21 May. Chair John Lonsdale: the regulator is "most focused on rapid developments in AI, which are outpacing the ability of many entities to manage the risks", ranked ahead of Middle East geopolitics.
- The Outlook flags "inadequate assurance over dynamic AI systems" as the supervisory gap, signalling "increase in supervisory intensity, including the potential for enforcement action."
- This is APRA's second AI-focused communication in three weeks, after the 30 April letter on AI risk management. ASIC's 8 May cyber letter to licensees also flagged frontier AI as intensifying the risk environment. AU-regulator cadence on AI is accelerating.
What this means in plain terms.
- The GP question is no longer "should we use AI?" It is "how do we govern it?"
- Boutique funds running copilots informally are now in scope alongside the majors. An AI policy as a one-page Word doc is no longer sufficient. The framing is about assurance, not policy.
In the mix
- Anthropic and OpenAI both moved agent execution inside the customer's perimeter. Anthropic shipped self-hosted sandboxes for Managed Agents on 19 May. OpenAI announced the Dell on-prem Codex partnership on 18 May. "We can't put the loan book in a US cloud" is being engineered away by both labs in the same week. (Anthropic, OpenAI + Dell)
- Gemini 3.5 Flash launched at Google I/O on 19 May. GA same day at $1.50 / $9.00 per million tokens, ~25% under Gemini 3.1 Pro. Outperforms 3.1 Pro on coding and agentic benchmarks. A cheaper option for batch summarisation and non-frontier agent steps. Worth a cost benchmark against Claude. (Google blog)
- OpenAI joined C2PA + added invisible watermarks on 19 May. Content credentials ship by default on every OpenAI-generated image. Public verification tool in preview. "Can you prove which artefacts were AI-generated?" will be a compliance-team question within 12 months. (OpenAI)
From my week
This week we spent on data-perimeter and governance research, the layer the labs are not going to solve for you. Three things that keep coming up in the work.
- Structural gate, not model capability. Across six well-publicised AI failures over the past 18 months, none were caused by the model being wrong. All were missing structural gates in the architecture. The model is rarely the risk; the architecture around it is.
- Every document needs an owner and a lifecycle. A fund's deal folder is full of docs at different states of truth: superseded term sheets, the latest IC memo, drafts never approved. Drop AI on top without lifecycle metadata and it confidently cites whichever version it retrieves first. Fix: named owner per artefact, explicit state (draft / approved / superseded / archived). Same discipline civil engineers apply to drawings.
- The labs removing the perimeter objection makes data hygiene the binding constraint. Both Anthropic and OpenAI moved agent execution inside the customer's perimeter this week (see above). The historical objection is being engineered away. What replaces it: whether your data is actually structured, owned, and lifecycled well enough for AI to do useful work over it. Garbage in, garbage out.
Levercon is the AI team for investment funds. We embed alongside your team to automate workflows, build custom tools, and compound operational gains. Ultimately we help Australian based investment funds leverage AI.
Book a strategy call.
See where AI fits in your fund.
We'll show you where AI lifts efficiency and performance across your organisation.