---
title: "APRA CPS 230: Building an AI Stack You Can Swap Out"
description: "APRA's April 2026 letter flags single-provider concentration as a CPS 230 risk, so AI stacks need a rehearsed substitution path. This flow maps five stages, from workload classification to provider swap."
---

# APRA CPS 230: Building an AI Stack You Can Swap Out

**Type:** [How it works](https://yomu.fyi/visual-explainers/category/how-it-works.md)

![Five-stage left-to-right flow diagram: classify the workload, boundary the model vendor, log automated decisions, thread one decision id through metrics, logs, traces and the audit ledger, then rehearse a provider swap, with APRA CPS 230 obligation chips across the top.](https://yomufyi.s3.us-east-1.amazonaws.com/visual-explainers/9fd20974-520d-44ce-b8df-0071f9256fc4/image-53969cb5-fc60-45ca-bd3d-ec5787bf0479.webp)

> APRA's April 2026 letter flags single-provider concentration as a CPS 230 risk, so AI stacks need a rehearsed substitution path. This flow maps five stages, from workload classification to provider swap.

## Explanation

APRA's April 2026 letter names single-provider concentration as an operational risk under CPS 230, turning vendor choice into a resilience question. The flow runs left to right in five stages: classify the workload, draw a substitutability boundary around the model vendor, log automated decisions, thread one decision id through metrics, traces and the audit ledger, then rehearse the swap. Each stage produces evidence the previous one cannot: classification scopes the boundary, the boundary makes substitution cheap, and the ledger lets you prove the chain under audit.

## FAQ

### Why is a single AI vendor a CPS 230 problem and not just a procurement one?

APRA's April 2026 letter flags single-provider concentration as a risk under CPS 230, which is about operational resilience rather than cost or features. If one provider's outage or exit stops routing or scoring, the obligation stays with you, so the vendor becomes an architectural dependency worth designing around early.

### What makes a vendor boundary substitutable in practice?

Keep provider-specific calls behind a thin adapter or a single gateway so routing, failover, and cost attribution happen in one place. Clera's migration moved a 500 GB database with 500+ tables overnight while production traffic continued, which only works when the exit path was designed before it was needed.

### What should decision logging capture for automated decision-making?

The Privacy Act amendment effective 10 December 2026 requires disclosure of automated decision-making, so routing and scoring logic has to be explainable. Thread one decision id through metrics, logs, traces, and the audit ledger so you can reconstruct why a request was routed or scored as it was, including guardrail blocks and human handovers.

**Tags:** [AI Governance](https://yomu.fyi/topic/ai-governance.md), [API Gateway](https://yomu.fyi/topic/api-gateway.md), [APRA](https://yomu.fyi/topic/apra.md), [Auditability](https://yomu.fyi/topic/auditability.md), [Australia](https://yomu.fyi/topic/australia.md), [Compliance](https://yomu.fyi/topic/compliance.md)

## Related reading

### [How we migrated Clera’s 500 GB production database to ClickHouse Managed Postgres overnight](https://yomu.fyi/post/how-we-migrated-clera-s-500-gb-production-database-to-clickhouse-manag.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Oct 8, 2026
- Type: Problem & solution

Clera, an AI talent agent representing 200,000+ candidates, migrated its 500 GB production Postgres database to ClickHouse Managed Postgres after outgrowing its previous provider. The previous database suffered from point lookups taking multiple seconds, 100% CPU for six straight days, and timeouts. Clera evaluated Supabase, Neon, PlanetScale, and ClickHouse, choosing ClickHouse for its local NVMe storage that avoids network-attached disk latency. In sysbench TPC-C testing, ClickHouse showed 16–42% higher throughput than the runner-up. The team migrated 500+ tables from the EU to the US overnight using ClickPipes and manual replay, with production traffic still hitting the database. After migration, CPU usage dropped to 10–20%, and no settings have been changed since the cutover.


### [Production-grade LLMs and agents: a field guide](https://yomu.fyi/post/production-grade-llms-and-agents-a-field-guide.md)
- Company: [Stack Overflow](https://yomu.fyi/company/stack-overflow.md)
- Author: Varun Jindal
- Published: Oct 8, 2026
- Type: Explainer

Production-grade LLM agents require deterministic machinery around the probabilistic model, not a larger model. A maturity model defines six levels: determinism, evaluation, confidence, safety and governance, operability, and production-grade build, each assuming the one below. Teams should fix their weakest level in order, starting with determinism before evals, and contain the model to a single node in a fixed graph where it proposes actions a separate component applies after approval. A self-assessment checklist with items like evals as a CI gate and layered guardrails helps teams identify their current level, with fewer than half ticked indicating they are earlier than the demo suggests. The post links to a series of 25 deep-dive articles covering each rung of the maturity model.


### [Part 5: Operating an LLM system: observability, cost, routing, and the platform underneath](https://yomu.fyi/post/part-5-operating-an-llm-system-observability-cost-routing-and-the-plat.md)
- Company: [Stack Overflow](https://yomu.fyi/company/stack-overflow.md)
- Author: Varun Jindal
- Published: Oct 8, 2026
- Type: Explainer

Level 5 covers operability for LLM systems: daily running, cost control, provider outage survival, and fast shutdown on misbehavior. A single gateway funnels all model traffic, measuring cost, routing requests, enabling failover, and hosting an instantaneous kill switch; a shared decision\_id threads metrics, logs, traces, and the audit ledger. Observability supplements RED metrics with a decision layer capturing auto vs. human splits, guardrail blocks, judge disagreements, token spend per tenant and capability, and shadow-eval pass ratios via tiered structured logs. Cost controls attribute spending at the gateway per tenant and capability, with hard budget ceilings; optimization follows order: skip model, batch, cache, right-size, trim—each step profiled against quality. The platform spine includes stateless agent services, the gateway, an audit datastore, cache, secrets store, and object storage—scoped with least-privilege IAM and encryption, avoiding queues or vector stores until warranted.


### [Best AI solutions for contact center in Australia: a 2026 buyer’s guide](https://yomu.fyi/post/best-ai-solutions-for-contact-center-in-australia-a-2026-buyer-s-guide.md)
- Company: [Twilio](https://yomu.fyi/company/twilio.md)
- Author: Aliakbar Lokhandwala
- Published: Sep 28, 2026
- Type: Explainer

AI self-service is deployed in 51% of Australian contact centers but only 38% deliver expected results, with handover failures and regulatory changes driving a new evaluation approach. Only 13% achieve smooth self-service-to-human handovers, and 84% of Australians rate straightforward handover as important, yet 98% expect self-service volume to grow. The Privacy Act amendment effective 10 December 2026 requires disclosure of automated decision-making, affecting routing and scoring logic, while APRA’s April 2026 letter flags single-provider concentration as a risk under CPS 230. Australia reversed its voluntary AI stance in July 2026, announcing mandatory AI Standards and a Digital Duty of Care, with legislation expected in early 2027. Vendor selection now depends on security classification, composability for substitution, and decision logging, with the Commonwealth Bank’s experience—cutting then reversing AI-driven job cuts after increased call volumes—illustrating that containment quality and agent capability are better objectives than headcount reduction.