This pillar requires access
Pillars 02–06 are available with the full AI6 Guidance — Compliance App. Purchase access or enter your password below.
Humans remain responsible for decisions and outcomes. Define exactly where humans must remain in the loop, ensure operators have the training and authority to override AI, and maintain viable manual alternatives for every critical function an AI system performs.
No AI system should make final decisions on matters that carry significant consequences for people, operations, safety, or legal standing — without a human reviewing the output and taking accountability for the decision. This is the principle of meaningful human control. It is not a technical constraint. It is a governance requirement, and it must be operationalised through documented protocols, not assumed through good intentions.
There are two modes of human oversight. "In the loop" means a human reviews and approves every AI output before it is acted on. "On the loop" means a human monitors AI decisions and has the authority to intervene — but does not approve every individual output. Which mode is appropriate depends on the risk level of the system and the consequence of each decision.
Equally important: alternative pathways. Every AI system that performs a critical function must have a documented, maintained, and tested manual alternative. If the AI system fails or is decommissioned, operations must be able to continue. The manual pathway cannot be theoretical — it must be practiced.
Your safety monitoring AI recommends permitting a task to proceed based on sensor data. The site supervisor believes conditions are unsafe — but the AI system's output is displayed as a clearance, not a recommendation. The supervisor is uncertain whether they have authority to override it. No Human Oversight Protocol exists to clarify this. There is no documented deference hierarchy between AI outputs and site supervisor judgement. The decision to proceed results in an incident. The investigation finds no documented override protocol and no training on when operators should disregard AI outputs.
Your AI-powered workforce scheduling system crashes mid-project. It manages rostering for 340 site workers across three shifts. The vendor's servers are unavailable for 36 hours. Your organisation has no documented manual rostering process — the previous system was decommissioned when the AI tool was deployed. Operations are disrupted, safety-critical shift handovers are delayed, and the principal is notified. The audit finding: no alternative pathway was maintained for a safety-critical operational function.
The appropriate oversight mode must be defined for each AI system based on its risk level, decision frequency, and consequence. Both modes require documented protocols and trained operators.
A human reviews and approves every AI-generated output before it is acted on. The AI is an input to a human decision — not a decision-maker. Required for: safety-critical recommendations, employment or disciplinary decisions, outputs that affect contractual obligations, and any decision with significant or irreversible consequences.
Examples: AI-generated safety clearances, AI-assisted performance assessments, AI-generated tender pricing
A human monitors AI decisions in aggregate and has the authority and capability to intervene at any point. Suitable where decision volume makes individual review impractical, but where human oversight of patterns and outcomes is still required. The intervention mechanism must be documented, accessible, and tested.
Examples: AI-generated roster scheduling, AI-assisted document compliance checking, AI-driven site observation analysis
Every AI system must have a documented override process. This is the minimum four-step workflow. Adapt it to your operational context and document it in the Human Oversight and Intervention Protocol.
A worker, supervisor, or operator initiates a formal challenge to an AI-generated output or decision. The challenge mechanism must be accessible, named, and specific to the AI system — not a general complaint process. The operator must understand they have the right and authority to raise a challenge without penalty.
The AI system provides a technical capability to suspend implementation of the challenged decision. For high-risk systems, this pause capability must be testable and documented. It cannot rely on the operator informally "ignoring" the AI output — the system must support a formal hold state.
A named human reviewer — with documented authority and AI literacy — evaluates the AI output against real-world context. This person must understand the AI system's known limitations and failure modes. They must have access to the plain-language system explanation from Pillar 04. The review must be completed within a defined timeframe.
If the reviewer determines the AI output is incorrect or inappropriate, the decision is manually overridden and the intervention is recorded in the Contestability and Override Log. The log entry must capture: what decision was challenged, who reviewed it, what was decided, and why. This data feeds back into the monitoring framework from Pillar 05.
Maintaining human control requires coordination between operations, IT, HR, and risk functions. The oversight protocol is only as strong as the people trained to use it.
| Function / Activity | Executive Accountable Official | Operations Management | IT & Systems | HR & Training | Risk & Compliance |
|---|---|---|---|---|---|
| Oversight protocol design | A | R | C | C | C |
| Override mechanism implementation | I | C | R | I | A |
| Operator training delivery | I | C | I | R | A |
| Decommissioning plan maintenance | A | R | C | I | C |
| Contestability & override logging | I | R | A | I | C |
These four artefacts form the minimum documentation set for Pillar 06 compliance. Together they constitute the operational evidence that your organisation maintains meaningful human control over its AI systems.
A per-system document that defines: the oversight mode (in-the-loop or on-the-loop), the named overseer and their qualifications, the specific trigger conditions that require human review, the intervention mechanism, and the decision authority hierarchy. Must be reviewed and updated when the system changes or the operational context changes. Site managers and operators must be trained on this protocol.
Documented criteria and procedures for safely retiring an AI system. Must include: the specific conditions that trigger decommissioning (performance failure, unmitigable bias, privacy breach, strategic change), the technical steps for secure data handling and model suspension, the alternative pathway that replaces the system's functions, and the communication plan for affected stakeholders. Must exist before the system is deployed.
A running record of every instance where a human operator challenged, paused, reviewed, or overrode an AI-generated output. Must capture: the date and system involved, the nature of the challenge, the reviewer identity, the decision made, and the outcome. This log is operational telemetry — reviewed as part of the monitoring framework from Pillar 05 to identify patterns of AI failure or operator concern.
Documented manual procedures for every critical operational function performed by an AI system. Must be maintained, not archived — the manual process must be practiced periodically to confirm it remains viable and that competent personnel exist to execute it. For safety-critical functions, alternative pathways must be tested at least annually. Theoretical alternatives that have not been rehearsed are not compliant.
Your progress is saved automatically. Print this page to include your checklist status in a compliance submission.
Tick each item when you have the required documentation in place.
These are the findings that appear most frequently when contractors are assessed against Pillar 06.
Organisations assume that because a human sees AI outputs before they are actioned, meaningful oversight is in place. It is not. Meaningful oversight requires that the human reviewing the output has the training to evaluate it, the authority to challenge it, a documented process to escalate it, and a record of their review. Seeing an output is not the same as overseeing it.
Organisations deploy AI systems without decommissioning plans. When a system fails, is discontinued by a vendor, or is found to produce unacceptable outputs, there is no documented procedure for shutting it down safely, handling the data it holds, or transitioning to manual processes. Decommissioning plans must exist before deployment — not be improvised after a failure.
When an AI system was deployed, the manual process it replaced was discontinued — staff retrained, spreadsheets archived, procedures deleted. When the AI system fails, there is no functional manual alternative. This is the most operationally dangerous gap. For any safety-critical or operationally critical AI function, the manual alternative must be maintained, documented, and practiced in parallel with the AI system.
Site supervisors and operational personnel feel they do not have the authority to disregard or override AI outputs — particularly when those outputs are presented as approvals, clearances, or recommendations. No Human Oversight Protocol has been communicated that gives them explicit authority to exercise judgement. This is a governance failure that creates safety risk: it trains operators to defer to AI outputs even when their experience tells them the output is wrong.
The full AI Governance Compliance App delivers every document template across all six pillars, a live compliance dashboard that tracks your progress against each requirement, and a structured audit trail — purpose-built for Australian heavy industry contractors and ready for submission to principals, insurers, or compliance auditors.