A practical framework for testing image-based safety guidance without removing professional judgement from workplace inspections.

Professional warehouse inspection workflow showing a facility manager reviewing an AI-generated safety suggestion on a tablet while a warehouse aisle is visible in the background.

Artificial intelligence can make workplace inspections faster, but speed alone does not make a safety process reliable. For Singapore facility managers, warehouse operators and SMEs, the more useful question is how to introduce AI into an existing inspection workflow without allowing an automated suggestion to become an unchecked safety decision.

The Workplace Safety and Health Council introduced SAGE in July 2026 as an AI-powered prototype that can analyse workplace photographs, identify common safety risks, recommend relevant guidance and provide language support for migrant workers. Testing is expected before a wider launch in early 2027. Its intended role is to complement professional judgement, not replace it.

This creates an opportunity for businesses to prepare a controlled pilot. The pilot should focus on evidence quality, human approval, escalation, privacy and follow-through rather than treating an AI image result as a final inspection outcome.

What SAGE could contribute to a safety inspection

A photo-based tool may help a team review visible conditions such as blocked access routes, poor housekeeping, unsafe storage arrangements, missing safeguards or other common hazards. It may also help retrieve relevant workplace safety guidance and make information more accessible across different languages.

However, a photograph has limits. It may not show the full work activity, the condition behind an obstruction, the history of a recurring defect, or the risks associated with equipment operation. Lighting, camera angle, image quality and clutter can also affect the result. A photo may suggest a potential issue, but a competent person still needs to determine whether the issue is real, how serious it is and what action is appropriate.

Start with a narrow, controlled pilot

Do not begin by asking an AI system to review every safety issue across an entire site. Select one or two repeatable inspection areas where photographs can provide useful evidence.

  • Warehouse aisles and loading areas: review access routes, housekeeping, stacking conditions and visible obstructions.
  • Plant rooms and service areas: review access, signage, visible leaks, housekeeping and general condition.
  • Common areas in commercial buildings: review trip hazards, temporary obstructions and maintenance-related conditions.
  • Refrigeration or cold-room operations: use photographs as part of a wider inspection process, while separately managing equipment, maintenance, environmental and disposal requirements.

Define the pilot boundary in writing. Record which locations, hazard categories, users, devices and working hours are included. Avoid using the pilot to make decisions about worker performance or disciplinary action. Its purpose should be hazard identification and workflow improvement.

Design the workflow around human approval

A practical SAGE-enabled inspection workflow can use the following stages:

  1. Prepare: define the inspection route, approved device, photo guidance and escalation contacts.
  2. Capture: take clear photographs from a safe position. Record the location, date, time, inspection type and relevant work activity.
  3. Analyse: submit the image for AI-supported hazard identification and guidance retrieval, where the pilot arrangement permits it.
  4. Review: require a supervisor, safety representative or other designated competent person to classify the result as confirmed, not confirmed, unclear or requiring further investigation.
  5. Act: assign a corrective action, responsible person and target completion date.
  6. Verify: record whether the action was completed and capture follow-up evidence where appropriate.
  7. Learn: review repeated false positives, missed hazards, delayed actions and recurring site conditions.

The approval step should be visible in the record. A simple status such as AI suggestion, human reviewed, action assigned and closed provides better governance than storing an image without an outcome.

Set clear escalation rules

Not every result should be handled as a routine maintenance ticket. Create a written escalation matrix before the pilot starts.

  • Immediate control: stop or isolate the affected activity when there is a credible and urgent risk.
  • Supervisor review: refer unclear images, repeated hazards or conditions requiring operational context.
  • Safety review: involve the responsible safety professional or designated competent person for higher-risk work or significant uncertainty.
  • Engineering or maintenance action: route defects involving equipment, access, building systems or physical safeguards to the appropriate technical owner.
  • Management review: escalate repeated overdue actions, systemic housekeeping issues or risks requiring budget and operational changes.

AI confidence should not be used as the only escalation trigger. A low-confidence result may still describe a serious condition, while a high-confidence result may be wrong because the image lacks context.

Improve the quality of inspection evidence

Good evidence is more than a clear photograph. Use a consistent capture checklist:

  • Take an overview image and, where necessary, a closer image of the condition.
  • Avoid placing workers in an unsafe position merely to obtain a better photograph.
  • Record the exact area, asset or route rather than relying on a general site label.
  • Include enough context to show access, surrounding equipment and likely consequences.
  • Do not edit or crop images in a way that removes relevant information.
  • Record what the inspector observed directly, separately from the AI output.

For warehouse operations, link the inspection record to a zone, bay, dock, rack area or work process. For facility management, link it to the building, floor, plant room, asset or contractor activity. This makes trends and recurring defects easier to identify.

Manage privacy and information handling

Photos can contain faces, identity cards, vehicle details, screens, documents, security arrangements or commercially sensitive information. Before the pilot, decide what may be photographed, who may access the images, how long records are retained and where data is stored or processed.

Use approved devices and accounts. Avoid capturing personal information that is not needed for the inspection. Where practical, position the camera to focus on the hazard rather than people. Establish a process for handling accidental captures and restrict sharing to the people who need the information for safety or corrective action.

Businesses should also review their internal privacy, security and data-handling requirements and obtain appropriate advice for their operating model. A pilot should not create an uncontrolled image repository.

Connect AI findings to corrective-action management

The value of photo analysis is limited if findings remain in a chat, email or unstructured folder. Connect the pilot to an inspection or maintenance workflow that can assign ownership and track closure.

At minimum, each confirmed finding should include the location, hazard description, risk priority, interim control, action owner, due date, verification status and supporting evidence. A workflow may be managed through an existing digital inspection platform, maintenance system or structured form. The technology choice matters less than having a reliable handover from observation to action.

Where relevant, AI-supported inspection findings can sit alongside other digital WSH tools such as electronic permit-to-work processes, video analytics or IoT environmental monitoring. These systems should complement one another, not create multiple disconnected lists of unresolved issues.

Measure the pilot without overstating AI performance

Use practical measures that help management decide whether the workflow is useful. For example, track inspection completion, time from finding to human review, time to corrective action, overdue actions, repeated findings and the proportion of AI suggestions that were confirmed, rejected or marked unclear.

Do not report these figures as proof that AI is safe or accurate in every situation. They are pilot measures for understanding performance in a particular site, image set and operating process. Review examples manually, especially false positives and missed hazards. If the tool performs poorly in a location or hazard category, narrow the scope or improve the capture method.

A sensible next step for Singapore SMEs

A small operator does not need a large transformation programme to begin. Start with one site, one inspection route, a limited hazard list and a named human approver. Document the process, test it with real operating conditions and involve the people who will take corrective action.

BCA’s Smart FM guidance emphasises staged adoption, data-driven operations and measurable outcomes. That principle is equally useful for an AI safety pilot. Build a workflow that is understandable, auditable and useful even when the AI result is wrong or unavailable.

ISS can help businesses assess inspection workflows, connect facility and engineering requirements, structure corrective-action processes and plan practical AI automation. Contact ISS to discuss your engineering, facility management or AI automation requirements.

Reference points