Governing Intelligence · Part III. Putting it to work
10The AI assurance lifecycle and gates
The five domains describe the capability an agency needs. The assurance lifecycle is how that capability is applied to each AI use case, consistently, from the moment it is proposed to the moment it is retired. It is the agency’s single operating process for AI, and it is deliberately built on the Commonwealth’s own structure, the AI technical standard’s three phases of Discover, Operate and Retire, so that following the lifecycle is the same as meeting the standard.
10.1The lifecycle, stage by stage
Every in-scope use case travels eight stages. The early stages decide whether and how to proceed; the later stages assure it in operation and close it out responsibly. Gates between the stages are go/no-go decisions taken by the AI Governance Committee, on evidence.
| Stage | What happens | Key artefact |
|---|---|---|
| 1 Intake and triage | Register the use case; capture purpose, owner and a first view of risk. | Use-case register entry |
| 2 Impact assessment | Assess impact against the eight ethics principles using the DTA tool. | AI Impact Assessment |
| 3 Risk tiering | Classify as low or high risk; set the assurance depth that follows. | Risk tier and rationale |
| 4 Design assurance | Specify agents, autonomy gates and human-participation points; secure and safe by design. | Agent specs; design sign-off |
| 5 Build, test and simulate | Build under change control; test, and rehearse against realistic and adversarial scenarios. | Test and simulation evidence |
| 6 Deploy (gated) | Pass the pre-deployment readiness gate; deploy progressively (shadow, then ramp). | Readiness gate decision |
| 7 Operate and monitor | Monitor drift, trust and outcomes; keep the kill switch and rollback ready. | Monitoring and assurance reports |
| 8 Retire | Decommission responsibly; archive records and provenance. | Decommission and records plan |
10.2The readiness gates
Gates are where governance bites. The framework defines five go/no-go checkpoints; an AI use case advances only when the gate is passed, and the decision is recorded. The gates scale with risk: a low-risk assistive tool passes lightly, while a high-risk system affecting people’s entitlements faces every gate in full.
| Gate | Question it answers | Pass condition |
|---|---|---|
| G1 Design complete | Is the system fully specified and governable? | Agents, constraints, gates, fallbacks and human roles defined. |
| G2 Simulation passed | Does it behave under stress and adversarial conditions? | Critical scenarios pass; constraints, escalation and fallback observed. |
| G3 Governance validated | Is it accountable and compliant? | Impact assessment, risk tier, owner, transparency and audit trail in place. |
| G4 Shadow proven | Does it match expected behaviour against live conditions? | Shadow outcomes align; override rates within bounds. |
| G5 Organisationally ready | Can the agency operate and oversee it? | Owners briefed, escalation live, monitoring and rollback configured. |
10.3Rehearsing before reality: simulation
One discipline deserves particular emphasis because it is unfamiliar in government delivery: simulation. An AI system, especially an autonomous one, must be rehearsed before it is trusted, in a safe-to-fail environment that exposes how it behaves under boundary conditions, contradictory goals, corrupted data and adversarial pressure. Simulation is not prediction; it is rehearsal, and it is where an agency discovers the failure modes that a document review never will. For higher-risk use cases, the framework treats a passed simulation, gate G2, as a precondition of deployment, and recommends a shadow period in which the AI runs alongside the existing process without acting, so its behaviour can be compared to reality before it is trusted with it.
10.4Operating safely: monitoring, drift and the kill switch
Deployment is the beginning of assurance, not the end. In operation, the agency monitors the things that matter for AI: drift in the data and in the model’s behaviour, the rate and pattern of human overrides, trust scores, and outcomes against intent. Two capabilities are non-negotiable and must be designed in: a kill switch that can halt the AI immediately, and a rollback that can return to the prior state cleanly. The AI6 practices, test and monitor, and maintain human control, are operationalised here. An AI that cannot be stopped and reversed by a human is not ready for government use.
Facts checked against official sources on 7 October 2026. Confirm obligations against the source instruments, which can change. This framework is not legal advice.