flowchart TD
R["AI Risk & Review Forum<br/>(CIO + DPO + Compliance + AI Product Owner)"] --> G
G["AI-use intake & classification<br/>low / medium / high risk"] --> A
A{"Approval gate"}
A -->|low| P["Proceed — Level 1-2 HITL"]
A -->|medium| H["HITL + fairness + monitoring plan"]
A -->|high| F["Full model-risk assessment<br/>+ exec sign-off"]
P --> M["Model registry + monitoring"]
H --> M
F --> M
M --> I{"Incident?"}
I -->|yes| IR["AI incident process<br/>→ consider NDB notification"]
I -->|no| R2["Periodic review & retirement"]
AI Governance Considerations
ISYS6020 — should we, responsibly?
A positive NPV is necessary but not sufficient. This page covers the obligations and risks that apply to CloudCore’s AI ambitions regardless of payback — because the company that asks these questions is the same one that, in September 2025, detected a breach of ~250,000 records only after prolonged unauthorised access. For CloudCore, AI governance is not hypothetical: the failure mode it guards against has already happened.
1. Why governance is the gate, not the afterthought
Two facts frame everything below:
- CloudCore processes personal and commercial data for ~500 clients, much of it in regulated sectors (healthcare, finance) visible in the customer dataset. AI that touches this data inherits every obligation CloudCore already carries — plus new ones.
- The company’s controls were recently found wanting. The public breach account cites phishing and compromised credentials; the internal account is a six-factor failure including delayed detection. Either way, a governance framework that the company could not reliably operate under load is a fragile foundation for automated decision-making.
The discipline this page applies is simple: govern the decision, not just the model. An AI system that is technically excellent but makes an unlawful, unfair or unexplainable decision about a customer is a liability, not an asset.
2. Australian privacy law & the Notifiable Data Breaches scheme
CloudCore is an Australian APP entity under the Privacy Act 1988 (Cth) and must comply with the 13 Australian Privacy Principles (APPs). The principles most load-bearing for AI are:
| APP | Obligation | Why it bites AI |
|---|---|---|
| APP 1 | Openly manage personal information; have a clear privacy policy | An AI system that uses personal data is part of “information handling” and must be reflected in policy |
| APP 3 | Only collect personal information reasonably necessary; collect lawfully and fairly | Scraping or re-purposing client data to train a model is a collection act |
| APP 5 | Notify individuals of collection, including purpose | “Train a churn model” is a purpose that may need disclosure |
| APP 6 | Use/disclose personal information only for the purpose collected, unless an exception applies | Secondary use of data to train AI is a key risk — re-purposing needs a lawful basis |
| APP 7 | Direct-marketing rules | A churn model that triggers retention offers can stray into direct-marketing rules |
| APP 8 | Cross-border disclosure accountability | Routing inference through an overseas SaaS path is a disclosure — central to the RAG build-vs-buy decision |
| APP 11 | Take reasonable steps to protect personal information | Security of the model, its training data and its outputs |
| APP 13 | Correct inaccurate personal information | A wrong prediction or profile must be correctable |
The Notifiable Data Breaches (NDB) scheme. Under Part IIIC of the Act, an “eligible data breach” — unauthorised access to or loss of personal information likely to cause serious harm — must be notified to the OAIC and affected individuals. An AI system that exposes personal data (a leaky model, a prompt-injection on a RAG system, a mis-routed ticket) can itself create an eligible data breach. CloudCore has already lived one NDB event; an AI programme multiplies the surface area.
Beyond the Act, two voluntary-but-influential frameworks shape what “good” looks like in 2026:
- Australia’s AI Ethics Principles (8 principles: human/social wellbeing, fairness, privacy, reliability/safety, transparency/explainability, contestability, accountability, sustainability).
- The Voluntary AI Safety Standard (10 guardrails, 2024) and the progressing policy debate on mandatory guardrails for high-risk AI. A defensible CloudCore strategy treats the voluntary standard as the minimum and anticipates the mandatory direction.
3. Human-in-the-loop — a spectrum, not a switch
The reflex “keep a human in the loop” is too blunt. The right question is at which point a human intervenes, and with what authority. A graduated model:
| Level | Automation | Human role | CloudCore example |
|---|---|---|---|
| 1 — Inform | AI suggests only | Human decides, may ignore | RAG policy search (cites sources; human reads) |
| 2 — Recommend | AI recommends; human acts | Human reviews before action | Support-triage routing (agent confirms) |
| 3 — Concur | AI acts, human confirms in batch | Human samples/audits | Churn-retention prioritisation (manager signs off the list) |
| 4 — Act with override | AI acts; human can intervene | Human monitors, can halt | (Not advised for client-affecting decisions yet) |
| 5 — Autonomous | AI decides and acts | None | Do not use for any decision affecting a customer’s data, money or service |
Mapping the three candidate use-cases:
- Support triage → Level 2. The model routes and drafts; the agent owns the response. Low risk, but ticket text must be screened for personal data before any model that learns from it.
- Churn prediction → Level 3, never above. The model prioritises which at-risk clients to contact; a human account manager decides the offer and makes the call. The model must never auto-cancel, auto-price or auto-communicate. A false positive that triggers a “we notice you might leave” message to a satisfied client is reputational damage, not a tuning problem.
- RAG policy search → Level 1, with a hard refusal rule. The assistant cites a source document for every answer and, below a confidence threshold, declines to answer rather than hallucinate. In a regulated context a confident wrong policy answer is worse than no answer.
The contestability requirement (APP & ethics). Any AI-influenced decision that affects a customer must be explainable (what did the model consider?) and contestable (can a human override it?). This is why Level 4–5 is excluded for client-affecting decisions at CloudCore’s current maturity — the company cannot yet guarantee either property.
4. Model risk management
A model is an asset that decays. Four risk classes must be governed through the lifecycle:
- Validation before release. Defined training/test splits, documented features, measured performance against a pre-agreed acceptance threshold, and a record of what the model is bad at (failure modes), not just its average score.
- Drift after release. Customer behaviour, product mix and threat patterns change; a churn model trained on 2026 data silently degrades. Monitoring for data and concept drift is mandatory, with a retraining cadence.
- Bias and fairness. A churn or pricing model that systematically disadvantages a region, industry or company size is both an ethics failure and a legal exposure. Fairness testing across the dataset’s
regionandindustrydimensions is a baseline control. - Robustness & security. Prompt-injection and data-poisoning are real attack surfaces — especially salient for a company that was breached via a third-party vulnerability. Models handling free text (
ticket_text,review_text) need input sanitisation and adversarial testing.
Make the model-risk policy explicit. CloudCore already has a security-architecture / operating-model policy. Extending it with an AI-specific addendum — acceptance criteria, monitoring, a model registry, and a retirement process — is item 5 in the technology-landscape roadmap and is cheap relative to the risk it contains.
5. Use-case risk register
| Use-case | Top governance risk | Primary control | Residual risk |
|---|---|---|---|
| Support triage | Personal data in ticket text leaking into a vendor model | Mask PII pre-inference; AU-resident or DPA-covered vendor; Level 2 HITL | Low |
| Churn prediction | Unlawful secondary use of customer data; unfair targeting; direct-marketing breach | APP 5/6 disclosure; fairness testing by region/industry; Level 3 HITL; explicit retention-offer governance | Medium |
| RAG policy search | Hallucinated policy advice acted upon; access-control bypass | Mandatory citation + refusal below threshold; access controls mirror source docs; Level 1 HITL | Medium |
| (Future) Log anomaly | Alert fatigue or missed critical alerts; autonomous response | Detection only — never auto-response; human SOC acts | Medium |
6. Vendor & third-party model risk
Most of CloudCore’s first AI will be bought or API-sourced, not built. That shifts risk to procurement and contracts:
- Data residency (APP 8). Does inference run in Australia? Post-breach, an overseas inference path for internal policy or customer data is a governance regression. This is the decisive factor in the RAG build decision.
- Training-data provenance. Will the vendor train its foundation model on CloudCore’s data? This must be contractually excluded.
- Model changes. A vendor that silently updates its model invalidates the validation CloudCore performed — contracts must require change notification and re-validation rights.
- Audit & exit. SOC 2 / ISO coverage, the right to audit, and a clean data-extraction path on termination.
A vendor that is cheaper but fails any of these is not cheaper — it is unbudgeted risk.
7. A lightweight AI-governance operating model
Governance that is not operated is decoration (the breach proved this). A minimum viable model for a ~47-person company:
- One accountable owner per use-case (model risk is never “the team’s”).
- An intake question: does this decision affect a person’s data, money or service? If yes → high-risk path.
- A model registry recording purpose, data, owner, HITL level, acceptance metrics and review date.
- An AI incident process wired into the existing breach-response process, with a standing question: could this be an eligible data breach under the NDB scheme?
8. The through-line
Every governance choice above reduces to one principle: for a company that was breached because controls existed but were not reliably operated, the test of an AI initiative is not whether it is built, but whether the company can prove — to a regulator, a client and its own board — that the decision it makes is lawful, fair, explainable and correctable.
That test is what separates an AI investment that compounds trust from one that simply expands the attack surface.