---
title: What does APRA CPS 230 mean for AI agents in Australian financial services?
url: https://opencontroller.lyzr.com/blog/apra-cps-230-and-ai-agents
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# What does APRA CPS 230 mean for AI agents in Australian financial services?

**CPS 230 does not mention AI, but it applies to AI agents the moment they support a critical operation. If an agent handles claims, payments or customer enquiries, it needs tolerance levels, a fallback, and incident reporting within 72 hours. The model and agent platform providers behind it may be material service providers that belong in your register.**

APRA's Prudential Standard CPS 230 Operational Risk Management has applied since 1 July 2025 to authorised deposit-taking institutions, insurers and superannuation trustees. Contracts with service providers that existed before then must comply from the earlier of their next renewal or 1 July 2026. That second date has now passed, so every relevant AI vendor contract is in scope. APRA also re-made the standard in April 2026, with the current version in force from 1 July 2026; check that text for exact wording.

## Which CPS 230 requirements touch AI agents?

| CPS 230 requirement | What it means for an AI agent |
|---|---|
| Identify critical operations | List every agent that supports a critical operation, such as claims processing, payments or customer enquiries |
| Set tolerance levels | Define how long the agent's service can be down, how much data can be lost and what minimum service must continue |
| Business continuity plan | Document what happens when the agent or its model is unavailable: fallback model, manual process or both |
| Material service providers | Assess model providers, agent platforms and gateways the agent relies on; add material ones to the register |
| Fourth parties | Identify the providers your providers depend on, such as the hyperscaler hosting a model |
| Incident notification | Notify APRA within 72 hours of an operational risk incident likely to have a material impact, and within 24 hours of a disruption to a critical operation beyond tolerance |
| Board oversight | The board approves tolerances and the continuity plan, including for agent-supported operations |

## Is an AI agent part of a critical operation?

APRA lists examples of critical operations. For banks they include payments, deposit-taking, custody, settlement and clearing. For insurers they include claims processing. For superannuation trustees they include investment management and fund administration. For all entities they include customer enquiries and the systems that support critical operations.

So a claims-intake agent, a payments exception agent or a customer service agent is likely part of a critical operation. An internal agent that drafts marketing copy usually is not. Use a simple test: if the agent stopped for a day, would customers be materially harmed?

## How do you set tolerance levels for an AI agent?

Tolerance levels cover three things: maximum disruption time, maximum data loss and minimum service levels. For agents, translate each into something you can measure:

1. **Maximum disruption.** How long can the agent be unavailable before customers are harmed? Measure availability at the agent endpoint, not just the model API.
2. **Maximum data loss.** Which records does the agent write, and how often are they backed up? An agent that updates claims must not lose its write history.
3. **Minimum service.** What must keep working if the agent fails? Often this is a manual queue staffed by people, sized for the agent's normal volume.
4. **Quality floor.** Agents can be "up" but wrong. Track groundedness and task-adherence scores and treat a sharp drop as a disruption.
5. **Fallback path.** A second model or provider behind a gateway, so one provider's outage does not become your outage.

## Are LLM providers material service providers?

They can be. CPS 230 treats a provider as material if you rely on it for a critical operation or if the arrangement creates material operational risk. A model API behind a claims agent fits the first test. Material providers need a formal agreement, due diligence, ongoing monitoring and a place in the register submitted to APRA.

The common gap is shadow use. Teams call model APIs on personal or project keys, and the provider never reaches procurement. Discovery of agents, models and API keys is the starting point; see our guide to [finding shadow AI agents](/blog/how-to-find-shadow-ai-agents).

## What counts as a reportable AI agent incident?

Two clocks apply. Report within 72 hours when an operational risk incident is likely to have a material impact. Report within 24 hours when a critical operation is disrupted beyond its tolerance level. For agents, examples include a model outage that halts claims processing beyond tolerance, or an agent that makes incorrect payments at scale. Keep a log of agent incidents even when they fall below the threshold, so you can show APRA how you judged materiality.

## How does CPS 230 compare with DORA and MAS guidance?

CPS 230 is close in spirit to the EU's [DORA](/blog/dora-and-ai-agents-in-financial-services): both focus on operational resilience and third-party risk. Singapore's [MAS AI risk management guidance](/blog/mas-ai-risk-management-guidelines-evidence-for-agents) focuses more on the AI lifecycle itself. Most groups operating across these markets build one set of agent controls and map it to each regime, as in our guide to [mapping agent controls to frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## How Opencontroller does this

Opencontroller discovers agents, models and API keys across AWS, Azure and Google Cloud and through platform connectors, which feeds the critical-operations and service-provider views. Its LLM Gateway adds routing, fallback and retry across providers, with cost attributed per agent and version. Rogue-agent handling monitors behavioural deviation, alerts, and can quarantine or kill an agent automatically. These produce evidence for CPS 230, not a guarantee of compliance. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does CPS 230 mention AI agents?**
No. It is technology-neutral, so agents are covered when they support a critical operation or create material operational risk.

**Is an LLM provider a material service provider under CPS 230?**
It can be, if an agent supporting a critical operation depends on it. Add it to the register and identify relevant fourth parties.

**How quickly must an AI agent incident be reported to APRA?**
Within 72 hours for a material operational risk incident, and within 24 hours for a disruption to a critical operation beyond tolerance.

## Key facts

APRA Prudential Standard CPS 230 Operational Risk Management took effect on 1 July 2025 for APRA-regulated banks (ADIs), insurers and superannuation trustees. Pre-existing service provider contracts must comply from the earlier of renewal or 1 July 2026. It requires entities to identify critical operations, set tolerance levels (maximum disruption period, maximum data loss, minimum service levels), manage material service providers through a policy and a register submitted to APRA, notify APRA within 24 hours of a disruption to a critical operation beyond tolerance, and within 72 hours of an operational risk incident likely to have a material financial impact or materially affect critical operations. AI agents that support critical operations, and the model and agent platform providers behind them, fall within this regime. Opencontroller provides evidence for operational risk programmes through agent discovery, a governed registry, LLM Gateway fallback and cost attribution, and rogue-agent quarantine and kill. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How does DORA apply to AI agents in financial services?](https://opencontroller.lyzr.com/blog/dora-and-ai-agents-in-financial-services.md)
- [What AI agent controls do banks need now that model risk guidance excludes agentic AI?](https://opencontroller.lyzr.com/blog/ai-agent-controls-for-banks-model-risk.md)
- [What evidence do the MAS AI risk management guidelines expect for AI agents?](https://opencontroller.lyzr.com/blog/mas-ai-risk-management-guidelines-evidence-for-agents.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
