APRA AI Governance for Banks and Insurers: April 2026 Letter and What It Requires

Treat the April 2026 APRA letter the way you'd treat a warning shot, not a new rulebook: APRA is explicit that CPS 230, CPS 220, CPS 510, and the Financial Accountability Regime already apply to AI, and the letter's four gaps describe where supervised entities are falling short of standards that were already enforceable before anyone built an AI governance function. . The practical read for a CRO: "we're waiting for an AI-specific prudential standard" was never a valid excuse under APRA's framework, and after this letter it is a documented one.

Board-level AI visibility is a CPS 220/CPS 510 gap, not a new AI requirement

A board that cannot name its material AI systems, hasn't approved an AI-specific risk appetite statement, and hasn't assigned a single governance body clear oversight of AI cannot discharge the oversight CPS 220 and CPS 510 already require of it — generic enterprise risk appetite language that never mentions AI does not count. This is the gap I'd expect APRA to lead with, because it's the easiest one for a board to fix cheaply (an AI inventory and a one-page risk appetite addendum) and the hardest one to explain away in an enforcement conversation if it's missing.

Model risk frameworks built for regression don't ask the right questions about AI

Running a generative or agentic AI system through a model risk framework designed for logistic regression validates the wrong things — it checks for things like coefficient stability and misses hallucination rate, prompt-injection susceptibility, and whether the model's explanation of its own decision is even accurate. The uncomfortable truth here is that this isn't an Australia-specific problem: if SR 26-2's scope genuinely excludes generative and agentic AI from US model-risk guidance, international frameworks have the identical gap, and APRA is telling its regulated population not to wait for that gap to close somewhere else. .

Vendor AI oversight cannot be contractual — CPS 230 already says so

APRA's finding that entities test AI they build in-house but lean on vendor documentation for AI they buy is the single most actionable gap here, because CPS 230 already defines what active oversight of a material service provider looks like, and a signed warranty isn't it. The second-order problem APRA flags — that an AI vendor may itself depend on a foundation model provider the bank has zero visibility into — is the one most banks genuinely haven't solved yet; demanding your vendor's own upstream dependency disclosure is not standard practice at most institutions today, and it should become one.

AI failure scenarios belong in operational resilience testing, not just IT outage drills

CPS 230 requires that critical operations stay within impact tolerance even when a supporting system fails, and "the AI system fails or degrades" is a materially different scenario from "the data center loses power" — different failure modes, different detection signals, different fallback processes. A bank that has a tested manual fallback for a core-banking outage but has never actually pulled the plug on its AI fraud-detection model to see what happens next has a resilience framework that predates its own AI adoption.

Indian banks with Australian operations are already supposed to be doing this

State Bank of India, Bank of Baroda, and Bank of India operate Australian branches or subsidiaries subject to APRA's framework, and the "parallel governance" problem is more concrete than "map against both standards" — RBI's outsourcing directions already require Indian banks to apply risk-based, ongoing oversight (not contractual reliance) to any material outsourced service, AI vendors included, which is functionally the same test CPS 230 applies. The genuinely new piece for these banks is the Financial Accountability Regime's personal accountability mapping — FAR requires named "accountable persons" at the entity, and an Indian bank's Australian subsidiary needs to know explicitly which named individual owns AI governance failures under FAR, because RBI's framework doesn't impose an equivalent individual-liability structure today.

What a bank (or insurer) should actually do

1. Build a board-level AI inventory and an AI-specific risk appetite addendum this quarter — don't wait for a dedicated AI prudential standard to formalize what CPS 220/510 already require.
2. Re-run model risk assessments on every generative or agentic AI system using AI-specific failure modes (hallucination, prompt injection, bias, explanation accuracy) — not the checklist built for regression models.
3. Replace vendor-warranty reliance with independent testing and monitoring evidence for every material AI vendor, and ask vendors to disclose their own upstream model dependencies.
4. Add named AI-failure scenarios (not generic "IT outage") to operational resilience and business continuity testing, with documented fallback processes.
5. If operating in Australia, name the FAR accountable person for AI governance explicitly and reconcile that assignment against RBI's outsourcing-oversight obligations for the same AI systems.