AI rollouts are often presented first as technology decisions.

The questions arrive in a familiar order:

  • Is the system accurate?
  • Is the data secure?
  • Does it comply with privacy requirements?
  • What productivity gain can we expect?

They are sensible questions. But they are not enough.

The moment an AI system allocates work, changes priorities, recommends a decision, monitors performance or alters the information available to a worker, it has changed the work system.

How do we know the redesigned work remains safe when real people use it under real operational conditions?

This is becoming difficult to ignore.

SafeWork NSW says the 2026 reforms define a digital work system as an algorithm, artificial intelligence, automation or online platform. The reforms clarify that the existing primary duty of care applies to risks arising from these systems and create a further duty concerning risks arising from the allocation of work by a digital work system.[1] The Act received assent on 18 February 2026.[2] However, as at 22 July 2026, those substantive amendments had not commenced. SafeWork NSW says they will commence on a date announced by the NSW Government, at least one month after the supporting guidelines are published.[1]

The enacted but uncommenced section 21A specifically identifies excessive or unreasonable workloads, performance metrics, monitoring or surveillance, and unlawful discriminatory practices or decision-making as matters a PCBU must consider.[2]

The Australian Institute of Company Directors’ 2026 guidance says workplace AI can create risks of physical or psychological harm. It says the primary duty applies when organisations develop, procure or deploy AI systems, and that measures may include WHS risk assessment, consultation and training.[4]

Meanwhile, a new Australian Government report finds no evidence to date of broad AI-driven labour-market upheaval in Australia. It reports that occupations more exposed to potential automation by generative AI have grown more slowly than other occupations, while expressly describing the results as suggestive rather than definitive.[3]

The important point for risk leaders is not whether AI is good or bad. It is that implementation is already changing the conditions in which work is performed.

The risk is often in the interaction

An AI system can perform well in a technical test and still create unsafe work.

Consider this hypothetical example: a scheduling tool optimises travel and task completion. Its recommendations may be mathematically efficient. But what happens when the schedule removes recovery time, creates unrealistic transitions, increases lone work or encourages people to bypass a control to remain “on target”?

Or consider a hypothetical automated triage tool that helps workers prioritise cases. What happens when its confidence is overstated, its rationale cannot be examined, or workload makes challenging the recommendation practically impossible?

In a third hypothetical example, consider a performance dashboard that gives leaders unprecedented visibility. What happens when workers begin optimising what the system measures while important but less visible work is deferred?

None of these examples requires the technology to “fail”. The system may be operating exactly as designed. The failure may be that the design boundary excluded the realities of work.

Consultation is not change communication

An organisation may say it consulted workers when it explained the new tool, delivered training and opened a channel for questions.

That is communication. On its own, it may not satisfy WHS consultation requirements. Under the model WHS framework, consultation includes sharing relevant information, giving affected workers a reasonable opportunity to express views and contribute to decision-making, taking those views into account before decisions are made, advising workers of the outcome, and involving any relevant health and safety representative.[5]

Meaningful consultation asks the people who perform and support the work:

  • What has become easier—and what has become harder?
  • Where has discretion narrowed?
  • What new workarounds are appearing?
  • Which signals are trusted, and which are being ignored?
  • When the system is wrong or unavailable, who can intervene?
  • What pressure is created by the way output is measured?

The answers are operational evidence. They reveal how the system functions outside the implementation plan. People are often the first to detect weak signals in a changing system. If they are treated only as users to be trained, the organisation loses an important source of intelligence.

Governance needs verification, not reassurance

Boards and officers do not need to understand every line of code. Officers do, however, have a due diligence duty under the model WHS laws. Safe Work Australia describes this as including active monitoring and evaluation, ensuring appropriate resources and processes exist and are used, and checking that reporting processes are followed.[6]

That should include evidence beyond a project status report or vendor assurance pack.

  1. What work changed? Identify changes to allocation, timing, decision rights, supervision, workload, monitoring and escalation—not just the software installed.
  2. Who was consulted before and after deployment? Include workers, contractors, supervisors, health and safety representatives and people who support exceptions.
  3. What assumptions does the system make about normal work? Test them against variability, degraded conditions, competing goals and foreseeable misuse.
  4. Where can a person challenge or override the system? An override that is technically available but punished by performance metrics is not a meaningful control.
  5. What evidence would show emerging harm? Look beyond injury outcomes to workload, fatigue, error recovery, workarounds, escalation patterns, complaints and the quality of human review.
  6. Who owns the risk after go-live? Technology teams can support the system, but assigning technical ownership does not transfer or remove applicable WHS duties.[6][7]

Projects finish. Work systems continue to adapt. If governance ends at deployment, the organisation is verifying the technology at the point when it should be learning about the work.

Start with the work

AI can remove drudgery, improve access to information and help people make better decisions. Treating it as a work-design issue is not an argument against adoption. It is how adoption becomes more reliable.

The organisations that learn fastest will not be those with the longest acceptable-use policy. They will be those that can see how technology, people, incentives, workload and operational conditions interact—and can adjust before weak signals become harm.

Before approving the next AI rollout, ask: what will this change about the way work is really done, and how will we verify the answer?

This article provides general information, not legal advice. Organisations should obtain advice about the obligations applying to their circumstances and jurisdiction.

References

  1. SafeWork NSW, Development of the Digital Work Systems Guidelines, accessed 22 July 2026. Primary regulator source; commencement status should be rechecked when relied upon.
  2. NSW legislation, Work Health and Safety Amendment (Digital Work Systems) Act 2026 No 5—as made and legislative history, accessed 22 July 2026. Primary legislation sources.
  3. Australian Government Department of Employment and Workplace Relations, The AI and employment in Australia report, 8 July 2026. Primary government research summary; reported associations do not establish causation.
  4. Australian Institute of Company Directors, A Director’s Guide to AI Governance, 2026, Appendix A, p. 50. Authoritative governance guidance, not legislation.
  5. Safe Work Australia, WHS duties—Consultation, accessed 22 July 2026. Model-law guidance; exact requirements depend on the relevant jurisdiction.
  6. Safe Work Australia, Officer duties, accessed 22 July 2026. Model-law guidance; officer status and reasonable steps depend on the facts.
  7. Safe Work Australia, Principles that apply to work health and safety duties, accessed 22 July 2026. Model-law guidance on the non-transferability of WHS duties.