By Jo Kitney
AI adoption by Australian businesses increased from 1% in 2022–23 to 12% in 2024–25, with uptake reaching 35% among large businesses. As AI moves quickly into everyday operations, the next safety system risk is not a missing form or an outdated dashboard. It is an AI agent taking action across the business before anyone has decided what it is allowed to do.
Organisations and WHS teams are progressively replacing spreadsheets, paper forms and disconnected processes with digital systems. This uplift has improved visibility, consistency and reporting. Now, another shift is underway. Artificial intelligence is moving beyond summarising information or drafting content. AI agents can monitor data, identify patterns, trigger workflows, allocate tasks, communicate with people and, in some cases, make or recommend decisions.
That changes the governance question. WHS leaders can no longer only ask, “Does the system work?” They also need to ask, “What can it decide, what can it change, who remains accountable, and how will we know when it is wrong?”
AI is part of the WHS risk landscape
Safe Work Australia’s guidance is direct: AI and digital technologies can create opportunities to improve WHS, but they can also create new hazards or increase existing physical and psychosocial risks. PCBUs must manage those risks as they would any other WHS risk, by identifying hazards, consulting workers, and eliminating or minimising risks so far as is reasonably practicable.
This matters because an AI tool sits inside the system of work. If it influences rostering, workload, task allocation, monitoring, incident triage, corrective actions or access to WHS information, it can change how work is organised and experienced. A technically accurate tool can create unsafe outcomes if it increases work intensity, removes human judgement, produces unfair recommendations or gives people false confidence in incomplete information.
Agents create a different level of risk
A digital WHS system generally records, displays or routes information according to predefined rules. An AI agent may take a goal and determine the steps needed to achieve it. It may search several data sources, draft an answer, create an action, notify a manager and follow up, all without a person managing each step.
That capability is useful. An agent could identify overdue critical-control verifications, prepare a summary for an officer, prompt a supervisor about an unclosed corrective action or help a worker find the right procedure. But the same agent could rely on poor data, misunderstand context, expose confidential information, create excessive notifications or make an unsafe recommendation look authoritative.
The point is that AI autonomy must be matched by controls. The more an agent can influence work, people or decisions, the clearer its boundaries, oversight and evidence trail must be.
Five controls safety leaders should establish now
- Keep an AI register. Record every AI system and agent in use, its purpose, owner, users, data sources, decisions influenced and level of autonomy. Include tools introduced through software updates as well as AI projects.
- Set decision boundaries. Define what the agent may recommend, what it may do automatically and what always requires human approval. High-consequence actions should have a named human decision-maker.
- Assess the impact on work. Review physical and psychosocial risks before implementation and when the system changes. Consider workload, monitoring, fairness, accessibility, competence, consultation and unintended interactions with other systems.
- Require evidence and traceability. Safety-critical outputs should show their sources, assumptions and limitations. Keep records of significant actions and decisions so they can be checked, challenged and learned from.
- Monitor after deployment. Treat implementation as the start of governance, not the end. Track errors, near misses, overrides, worker concerns and changes in system behaviour, and pause use when controls are not working.
Governance must be shared
Effective AI oversight needs executive leadership, operational knowledge, worker consultation, privacy, cyber security, legal, procurement, people and culture, and WHS. ISO/IEC 42001 offers a useful management-system structure for organisations developing, providing or using AI, including defined responsibilities, risk assessment, impact assessment, monitoring and continual improvement.
Safety leaders have an important role because they understand consultation, systems of work, risk controls and assurance. They can help organisations move beyond broad statements about “responsible AI” and translate them into practical controls that operate where work is actually done.
What changes on Monday morning?
Ask three questions: Where is AI already influencing work? Which of those uses could affect health, safety or fairness? Who has authority to stop, override or correct the system?
Do not wait for a fully autonomous WHS agent to appear. Start with the AI already embedded in your organisations platforms, pilots and vendor tools. Build the register, agree the boundaries and involve the organisation now.
The next phase of digital WHS will not be defined by who adopts AI fastest. It will be defined by who can show that AI is governed, understood and safe to use.
Jo Kitney is a WHS specialist and business leader specialising in WHS, risk, governance, management systems and digital transformation. She works with organisations to strengthen WHS systems and responsibly integrate Microsoft 365, data and AI into the way work is managed.
