A Green Home-Office Score Does Not Clear the Risk.

A completed checklist is one account of one home-working moment. Use AI to route credible flags, missing evidence and change signals to people who can act. Never let a model clear the arrangement, diagnose a worker or turn a green result into silence.

A green score does not prove that a home-working arrangement is safe. It proves only that a particular set of inputs did not cross the thresholds built into a particular scoring method.

AI can help a financial-services employer sort self-assessments, identify missing answers and route follow-up. People who understand the work must examine the evidence, consult the worker and decide the response.

The better workflow is route, verify, consult and close. Green is a routing state with an expiry date, not an automated clearance and never a transfer of the duty.

Why does a completed checklist not clear the risk?

Home is still a workplace while work is performed there, but it is not a standardised branch floor. Safe Work Australia's PCBU information sheet on working from home says homes and living arrangements vary, so risk controls will vary. It identifies hazards including workstation, sedentary-work, fatigue, physical-environment and psychosocial hazards.

A checklist can reveal some of that picture. It cannot observe everything. The worker may not know that monitor height, sustained laptop use or task pattern matters. A photograph may omit the space under the desk. One response may miss sustained contact-centre or after-hours work.

That is why the Safe Work Australia sheet recommends a combination of methods. These may include surveys, tools and checklists, prompt reporting, consideration of tasks and resources, monitoring the work environment and reviewing incident, complaint and consultation information. It also says a health and safety professional may be needed where the potential risks warrant specialist assessment.

The current Commonwealth duty is broader than the form. Under the Work Health and Safety Act 2011, current compilation dated 1 July 2024, a PCBU must ensure, so far as is reasonably practicable, the health and safety of workers it engages, causes to be engaged or influences or directs while they are at work. Section 14 states that a duty cannot be transferred. Sections 47 to 49 require consultation with directly affected workers on specified WHS matters, including when identifying hazards, deciding controls and proposing changes that may affect health or safety.

The Work Health and Safety Regulations 2011, current compilation dated 25 March 2025, require relevant controls to be maintained so they remain effective and to be reviewed in specified circumstances. A green model output cannot establish that a chair is adjusted correctly, a break control is used, a work-demand control remains effective or a reported change has been resolved.

No mandated checklist or universal physical-inspection rule appears in the Commonwealth and Victorian sources reviewed for this article. Safe Work Australia's guidance describes checklists as one possible method. WorkSafe Victoria says in its current working-from-home guidance that a checklist may help, an employer may inspect, and an inspection need not be a physical visit because virtual viewing, video or photographs may be used. Do not demand home images automatically. Start with text. Explain any additional collection, minimise the view, use an approved access-controlled system with retention rules, and offer a non-image route where practicable.

Victoria must be treated separately. The Occupational Health and Safety Act 2004, current version effective 6 August 2025 applies instead of the model WHS Act. WorkSafe Victoria says the section 21 employer duty extends to work from home and section 35 consultation applies to matters that directly affect, or are likely to directly affect, employees.

The most dangerous design mistake is to treat self-report accuracy as the worker taking over the duty. Section 28 of the Commonwealth Act gives workers their own obligations, including reasonable care and cooperation with notified reasonable WHS policies or procedures. Their answers are evidence for consultation and risk management. They are not a warranty that the work is safe.

Nor should a triage system make a medical or employment decision. A report of discomfort may require a conversation, an ergonomic assessment or another authorised pathway. The model must not diagnose an injury, infer disability, reject a flexible-work arrangement, direct office attendance or rank the worker by supposed vulnerability.

What should AI triage actually do?

Replace the clearance score with a routing contract. Every assessment state should specify what the system may do, what forces human review, who owns the next action and what evidence closes it. No score closes its own case.

Use this routing checklist:

• Source state: default to text and record whether each input is a worker answer, approved system fact, free-text comment or missing information.
• Rule trace: show the exact approved rule that created each flag. Do not use an unexplained model confidence number as the reason for action.
• Override triggers: force human review for reported pain or discomfort, a relevant prior injury or individual need, missing or ambiguous answers, a move or renovation, changed equipment or tasks, an incident, or a worker or HSR concern.
• Human route: name the WHS practitioner, trained workstation assessor, manager, facilities contact or specialist authorised to consider the issue.
• Consultation point: state what will be discussed with the worker and HSR, where applicable, before controls are decided or changed.
• Closure evidence: record the human decision, control, responsible person, due date, worker feedback and review trigger. Never store only the colour.

Green cases still need a challenge mechanism. Select a documented sample for human review, independently of worker performance. Compare it with equipment-delivery status, unresolved hazard reports and consultation feedback. If reviewers find missed hazards, revise the questions and routing rules.

That is a green-challenge sample. It tests the screening method, not the worker. The same trigger discipline belongs in the wider risk register, where AI signals route to named control owners.

Safe Work Australia's current AI and digital-technologies risk guidance says AI risks should use the ordinary risk-management process, supported by worker consultation. It identifies human oversight and audits as controls to consider, and requires control measures to be maintained and reviewed. The model can sort evidence, but the surrounding work system must be tested as a control in its own right.

Use this prompt to propose routes from approved, de-identified assessment data. An appropriately skilled WHS lead must verify the rules and routes before use; a trained assessor reviews only cases routed for that expertise.

Act as a home-working risk triage assistant, not a duty holder, clinician or
decision-maker.

Jurisdiction: [JURISDICTION]
Approved questions and routing rules: [RULES]
De-identified responses: [RESPONSES]
Known work pattern and supplied equipment: [WORK_AND_EQUIPMENT]

Return a table showing source state, exact rule triggered, missing or conflicting
information, proposed human route, consultation question and required evidence.
Do not diagnose, clear the arrangement, decide work location or infer risk from
silence. Default to text-only inputs and do not request home images. Mark any
result without an approved rule as [HUMAN_REVIEW].

Fictional worked example: [BANK_NAME] asks [WORKER_ID] to complete an approved home-working self-assessment. All multiple-choice items return green. A free-text answer says the worker uses a laptop at a dining table for three contact-centre days each week and has noticed recurring shoulder discomfort.

An unsafe system ignores the text because the scored fields passed. A slightly more polished unsafe system labels the worker high risk and recommends an attendance change. Both responses exceed the model's role.

The governed system applies a documented override to route the case to [TRAINED_ASSESSOR]. That person confirms the tasks, duration, equipment and worker's concerns, then consults on practical options. An appropriately skilled human reviewer decides whether ordinary controls are enough or the case needs an OHS or ergonomics professional. No diagnosis is entered in the triage record.

Privacy is part of the workflow. Home images can reveal household members, possessions and location details. Free text may contain health information. The OAIC's guidance on commercially available AI products says the Privacy Act applies to all uses of AI involving personal information by covered organisations, recommends privacy by design and advises against entering personal or sensitive information into publicly available generative AI tools. The Australian Privacy Principles frame what may be collected and how it must be handled.

Collect only what the approved process needs. Offer private, non-image routes. Never send personal information, health details or home images to a public generative AI tool. Keep required health information in the authorised system.

Use this prompt to prepare a follow-up packet after a human confirms a flag. An appropriately skilled human reviewer must check the facts and proposed route with the worker; the WHS lead retains the control and closure decision.

Prepare a human follow-up packet from these approved records:
[DE-IDENTIFIED_RECORDS]

Worker reference: [WORKER_ID]
Confirmed flag: [FLAG]
Jurisdiction: [JURISDICTION]

Separate confirmed facts, worker-reported information, unknowns and conflicts.
List consultation questions, current controls, evidence still needed, possible
control categories and review triggers. Do not diagnose, prescribe, determine
fitness, approve home working or recommend an employment decision.

Comcare's current working-from-home guidance illustrates why follow-up cannot stop at furniture. It addresses workstation setup, lighting, trips, ventilation, connection, support, work expectations and boundaries around hours. A scalable triage process must be able to route both physical and psychosocial signals without pretending one composite number measures them equally.

Do this Monday

1. Remove the word cleared. Replace it in one self-assessment workflow with routed, pending human review, action open or verified closed. Give every state an owner and expiry date.
2. Publish the override list. Consult workers and HSRs on the facts that force human review, including discomfort, individual needs, missing answers, moves, renovations, task or equipment changes and requests for help. A one-page committee decision packet is a clean vehicle for that consultation.
3. Trace ten results. Check whether a reviewer can reconstruct the source answer, rule, route, consultation, decision and closure evidence without relying on the colour.
4. Challenge the green lane. Select a documented sample unrelated to performance, use an appropriately skilled human reviewer and record any missed hazard or weak question.
5. Cut the data footprint. Default to text and provide a non-image route. Confirm the approved system, access controls, retention period and private disclosure route.

Do not start by scoring the entire hybrid workforce again. Start with one business unit, one approved question set and one accountable follow-up team. Measure whether credible flags reach a person and whether actions close, not how many workers the model labels green.

Bottom line

A self-assessment is evidence, not a clearance certificate. AI can route gaps and change signals, but an appropriately skilled human reviewer must verify the case, consult the worker and decide the response. Keep medical judgement and individual employment decisions outside the model. A green state is safe governance only when it remains open to challenge, review and human action.

This article is general information and education only. It is not legal, compliance, financial or professional advice. Obligations vary by organisation and circumstance. Verify current requirements against the primary sources cited and seek advice specific to your situation.

References

1. Safe Work Australia, PCBU information sheet: Working from home: https://www.safeworkaustralia.gov.au/doc/pcbu-information-sheet-working-home
2. Federal Register of Legislation, Work Health and Safety Act 2011, current text: https://www.legislation.gov.au/C2011A00137/latest/text
3. Federal Register of Legislation, Work Health and Safety Regulations 2011, current text: https://www.legislation.gov.au/F2011L02664/latest/text
4. WorkSafe Victoria, Office health and safety: Working from home: https://www.worksafe.vic.gov.au/office-health-and-safety-working-home
5. Victorian legislation, Occupational Health and Safety Act 2004, current version: https://www.legislation.vic.gov.au/in-force/acts/occupational-health-and-safety-act-2004/045
6. Safe Work Australia, Artificial intelligence and digital technologies: Managing risks: https://www.safeworkaustralia.gov.au/safety-topic/hazards/digital-technologies-ai/managing-risks
7. Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products: https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products
8. Comcare, Working from home: https://www.comcare.gov.au/office-safety-tool/spaces/work-areas/working-from-home