These examples are fictional. All people, workplaces and events below are invented for education. They do not describe real organisations, and the appropriate response will depend on the facts, jurisdiction, workplace instruments and applicable law.
How to work through each scenario
- Notice the decision: Identify what the AI is doing and who is affected.
- Pause where needed: Contain immediate harm and preserve relevant records.
- Check the context: Identify the jurisdiction, award or agreement, policies, data, supplier terms and applicable law.
- Respond and learn: Give a person authority to review the outcome, correct controls and monitor whether the change works.
Recruitment: a ranking tool overlooks suitable applicants
Workplace: A medium-sized aged-care provider in Queensland recruiting 18 scheduling coordinators.
What happens
Mia, an HR adviser, receives 280 applications. A new screening tool ranks candidates using employment history, keywords and document formatting. Applicants with career breaks and CVs created with assistive technology appear disproportionately near the bottom. Managers plan to interview only the top 30. One experienced applicant asks whether a person reviewed her application.
What could go wrong
- The ranking criteria may not reflect the actual job.
- Formatting or career history may act as a poor proxy for capability.
- Accessibility and unequal effects may go unnoticed.
- An unexplained score may become a de facto rejection.
A practical response
- Pause automatic rejection and review a representative sample across the ranking.
- Document job-relevant criteria and test the tool against them.
- Offer an accessible contact and human-review route.
- Record the decision, testing results, limitations and accountable approver.
Questions to ask
- What evidence shows that each input predicts job performance?
- Which applicants are most likely to be incorrectly screened out?
- Can a reviewer see the original application and change the result?
Related guidance: Australian Human Rights Commission AI and recruitment checklist.
Rostering: efficient shifts create fatigue and disruption
Workplace: A distribution centre in Western Australia introducing demand-based rostering.
What happens
A rostering system predicts busy periods and publishes shifts automatically. Priya is repeatedly given a late finish followed by an early start, while casual workers receive frequent last-minute changes. The system meets its staffing target, but it does not recognise travel time, agreed availability, caring responsibilities or local fatigue controls.
What could go wrong
- The roster may conflict with an award, agreement, contract or approved availability.
- Fatigue and psychosocial risks may be hidden by an efficiency score.
- Workers may have no practical way to correct a bad output.
A practical response
- Stop unsafe shift combinations before publication.
- Check the applicable award, agreement, consultation process and WHS requirements.
- Build hard rules for rest, availability and notice into the workflow.
- Give a trained manager authority to override the system and explain changes.
Questions to ask
- Which constraints are mandatory rather than preferences?
- Are particular groups receiving more unstable or undesirable shifts?
- How quickly can a worker reach a person before the roster takes effect?
Related guidance: Fair Work Ombudsman consultation and cooperation guide.
Meeting transcription: the record captures too much
Workplace: A Melbourne engineering contractor trialling an AI meeting assistant.
What happens
The assistant joins a project meeting and produces a transcript, summary and action list. The discussion includes a worker's health information, a supplier's confidential pricing and an inaccurate statement attributed to the project manager. A share link is then sent to the wider project channel, where contractors can open it.
What could go wrong
- Participants may not understand that audio and text are being processed.
- Sensitive or confidential information may be retained or shared too broadly.
- An incorrect transcript may be treated as the official record.
A practical response
- Restrict access and remove the link while the incident is assessed.
- Check notice, consent, privacy, surveillance, confidentiality and contract requirements.
- Exclude unsuitable meetings and provide a clear way to pause recording.
- Require a participant to verify minutes before they become a record.
Questions to ask
- Where are recordings and transcripts stored, and for how long?
- Who can access, download, correct or delete them?
- Which meetings should never use automated transcription?
Related guidance: OAIC Australian Privacy Principles guidance. Surveillance and recording rules vary by jurisdiction.
Customer chatbot: a confident answer misstates the policy
Workplace: An online homewares retailer based in Sydney.
What happens
The retailer launches a chatbot to answer delivery and returns questions. It promises same-day dispatch for an item that is out of stock and tells a customer that sale items can never be returned. The customer cannot find a human-contact option and assumes the chatbot's statements are the retailer's final position.
What could go wrong
- The overall impression may mislead customers.
- The bot may invent terms that do not match consumer guarantees or company policy.
- Customers with accessibility needs may be unable to escalate.
A practical response
- Correct the affected advice and make human escalation visible.
- Limit answers to approved, current source material.
- Test common, unusual and adversarial questions before release.
- Review logs for repeated harmful answers and assign an accountable owner.
Questions to ask
- Does the interface clearly identify automation?
- Can the chatbot make promises or only explain approved information?
- What triggers immediate transfer to a person?
Related guidance: ACCC advertising and promotions guidance.
Performance monitoring: a proxy score becomes a disciplinary shortcut
Workplace: A contact centre in New South Wales assessing an AI-assisted quality dashboard.
What happens
The dashboard combines call length, silence, customer sentiment and script adherence into an "engagement" score. Jordan's score falls after moving to a noisy home workspace and handling more complex calls. A manager starts a performance plan without listening to the calls or checking whether the score is reliable for different accents, disabilities or call types.
What could go wrong
- A weak proxy may be mistaken for performance.
- Monitoring may engage workplace-surveillance, privacy, consultation or industrial requirements.
- Workers may not know what data is used or how to correct it.
A practical response
- Pause consequential use of the score and review the underlying calls and context.
- Check the NSW surveillance framework and applicable workplace instruments.
- Test accuracy and subgroup effects before relying on the metric.
- Explain limitations and provide a documented correction and review route.
Questions to ask
- What does the score measure, and what important context does it miss?
- Was monitoring introduced with the required notice or consultation?
- Can a manager disagree without being penalised for overriding the tool?
Related guidance: AI monitoring and workplace surveillance.
Safety recommendations: a green dashboard hides a real hazard
Workplace: A food-processing plant in regional Victoria using computer vision to flag hazards.
What happens
The system detects missing protective equipment but misses a recurring wet-floor hazard because glare obscures the camera view. After the dashboard rates the area "low risk", a supervisor closes a near-miss report without inspecting the site. Workers say they had already raised drainage and cleaning-timing concerns.
What could go wrong
- A false negative may create unjustified confidence.
- The dashboard may displace worker knowledge and competent assessment.
- The underlying hazard may remain uncontrolled.
A practical response
- Reopen the report and inspect the area using the applicable OHS process.
- Consult affected workers and health and safety representatives.
- Control the hazard using established risk-management principles.
- Record the missed detection, retraining or configuration change, and validation result.
Questions to ask
- Which hazards cannot be reliably observed by the system?
- Who can stop work or escalate despite a low-risk score?
- Are false negatives tracked as carefully as successful detections?
Related guidance: Safe Work Australia AI and digital technologies guidance. Victoria operates under its own OHS framework.
Generative AI: a fast draft exposes information and invents facts
Workplace: An Adelaide engineering consultancy preparing a competitive tender.
What happens
Leo uploads a client's brief to a public generative-AI tool and asks it to draft the tender. The response is polished, but it invents two project statistics, names a supplier the firm has not approved and reproduces confidential details from the brief. A director is asked to sign the document an hour before the deadline.
What could go wrong
- Confidential or personal information may leave approved systems.
- Fabricated claims may enter a contractual document.
- Copyright, attribution and record-keeping questions may be missed.
A practical response
- Stop submission and assess what information was entered and retained.
- Verify every material claim against an authoritative source.
- Use only approved tools and apply data-classification rules.
- Keep a human approver accountable for the final tender.
Questions to ask
- Is the tool approved for this information and purpose?
- Can every claim, calculation and citation be independently verified?
- What record of AI assistance should be retained?
Related guidance: Australian Cyber Security Centre Essential Eight and the organisation's confidentiality, privacy and records requirements.
Agentic AI: an automated purchasing assistant acts beyond its limit
Workplace: A regional wholesaler in Tasmania piloting an AI agent connected to inventory and supplier systems.
What happens
The agent is allowed to monitor stock, draft supplier emails and create purchase orders. A corrupted inventory feed makes 400 units appear unavailable. The agent sends an urgent order worth $82,000 and emails two suppliers before a staff member notices. Its service account also has access to customer records that are unrelated to purchasing.
What could go wrong
- Excessive permissions can turn a data error into financial and privacy harm.
- Actions may occur faster than a person can review them.
- Accountability may be unclear across the business and vendor.
A practical response
- Use the kill switch, preserve logs and assess affected orders, suppliers and data.
- Restrict the agent to the minimum systems, data and actions required.
- Require human approval above clear transaction and communication thresholds.
- Test failures in a sandbox and rehearse rollback before reconnecting.
Questions to ask
- What can the agent read, change, send, approve or purchase?
- Which actions require one-person or two-person approval?
- Can the organisation reconstruct and reverse every material action?
Related guidance: Australian Cyber Security Centre Essential Eight. Least privilege, transaction limits and kill switches are recommended governance controls, not universal statutory requirements.
Turn a scenario into action
Use the templates to document the system, assess impacts, plan consultation and record an incident or near miss.
Sources and currency
Jurisdiction: Australia. Currency: Sources checked 27 October 2026. Legislation is binding where it applies; regulator and government material is guidance unless stated otherwise. Standards and recommended governance practices are not legislation. Check the instrument, jurisdiction and facts that apply to you.
- Safe Work Australia AI and digital technologies guidance
- Fair Work Ombudsman consultation and cooperation guide
- OAIC Australian Privacy Principles guidance
- Australian Human Rights Commission AI and recruitment checklist
- ACCC advertising and promotions guidance
- Australian Cyber Security Centre Essential Eight