Practical steps for preparing inspection data, human review and safety workflows before AI-supported workplace testing begins.

Professional Singapore facility inspection scene showing a worker using a tablet to review a workplace photograph, with simple visual callouts for photo standards, human review and corrective action.

Singapore is preparing for a new stage of technology-supported workplace safety. On 23 July 2026, the Ministry of Manpower announced SAGE, an AI-powered safety advisory and guidance engine developed with the Singapore Institution of Safety Officers and industry partners.

Based on the announcement, the prototype is intended to analyse workplace photographs, identify common safety risks, recommend relevant safety guidance and support communication in different languages. Workplace safety professionals are expected to test the prototype over the coming months, with a full launch expected in early 2027.

SAGE is not yet a commercially available replacement for inspections, risk assessments or professional judgement. However, its development gives facility managers, warehouse operators, building owners and SMEs a useful opportunity to prepare. The organisations most likely to benefit from future AI safety tools will be those with consistent inspection data, clear accountability and disciplined corrective-action processes.

What SAGE could mean for workplace teams

Photo-based AI can potentially help a safety or facilities team review visible conditions such as blocked access routes, poor housekeeping, unsafe storage arrangements or other common hazards. A guidance-retrieval function could help users locate relevant safety information more quickly, while multilingual support may make instructions easier to communicate to a diverse workforce.

These capabilities should be viewed as decision support. A photograph may not show the full operating context, the task being performed, the condition of equipment or the level of risk. AI may also miss hazards, misinterpret an image or identify a condition that requires further verification. The responsible model is therefore human-led: AI can help surface observations, but trained personnel should assess significance, decide on controls and confirm closure.

1. Standardise the way inspection photos are captured

Before an organisation tests any image-based safety tool, it should improve the consistency of its visual records. Random photographs taken at different angles, distances and lighting conditions are difficult for people and software to compare.

Develop a simple photo protocol covering:

  • When photographs should be taken, such as during routine inspections, after a reported issue or before work begins.
  • Which areas, assets and work activities require coverage.
  • How inspectors should capture an overall view, a closer detail and the surrounding context where relevant.
  • How files should be named using a consistent site, location, date and inspection reference.
  • How repeat images should be captured so that corrective actions can be compared over time.

Teams should also avoid relying on photographs alone. Each image should be linked to basic context, such as the location, inspection type, observed condition and immediate action taken. This information can make future AI outputs more useful and easier for a human reviewer to validate.

2. Define human review and escalation responsibilities

AI-generated observations should enter an existing responsibility structure rather than create a parallel process. Before testing, define who receives an AI observation, who checks it, who decides the risk level and who authorises corrective action.

A practical review model may include:

  • An initial reviewer who confirms whether the image is clear and whether the observation is relevant.
  • A competent safety, engineering or operations representative who assesses the condition and selects appropriate controls.
  • A manager or permit owner who confirms whether work should pause, change or proceed.
  • A close-out owner who verifies that the corrective action has been completed.

Organisations should also define escalation rules for high-risk or time-sensitive conditions. An AI tool should not delay immediate controls, emergency communication or professional assessment. If a photo suggests a serious risk, the existing safety response should take priority while the observation is reviewed.

3. Protect sensitive workplace information

Workplace photographs can contain more information than expected. Images may show employee faces, identity cards, access-control details, vehicle registration plates, security layouts, customer information, equipment settings or confidential project materials.

Before uploading or sharing images with any AI system, businesses should identify what information is being captured and whether it is necessary for the safety purpose. Practical controls may include:

  • Using approved devices and storage locations.
  • Restricting access according to job responsibilities.
  • Removing or masking unnecessary personal or commercially sensitive details where feasible.
  • Defining retention and deletion rules for inspection images.
  • Recording where images are sent, processed or stored before a pilot begins.
  • Briefing staff on what may and may not be photographed.

Facility managers should involve the appropriate internal IT, security, privacy or management stakeholders before connecting a new tool to operational systems. A small, controlled test area is usually easier to govern than an immediate site-wide rollout.

4. Connect AI observations to existing safety workflows

The value of an AI observation depends on what happens after it is generated. If the output remains in a separate dashboard or chat window, teams may struggle to track accountability and completion.

Map the intended process against the systems already used by the organisation. For example, an observation related to active work may need to be reviewed alongside the permit-to-work process. A recurring housekeeping issue may belong in an inspection or facilities ticketing workflow. A serious event or near miss may need to enter the incident-reporting process. A confirmed issue should be assigned a corrective action, target date, responsible person and verification step.

Before testing, agree on a common status structure such as identified, under review, action required, action completed and verified. Define the minimum information that must be retained, including the original image, reviewer decision, action taken and close-out evidence. This creates an auditable trail without treating the AI output itself as proof that a hazard exists.

5. Start with a focused, measurable pilot

Businesses do not need to begin with every site, activity or hazard category. A focused pilot can test whether the tool fits real operating conditions while limiting disruption and data exposure.

Suitable pilot questions may include:

  • Can inspectors capture images consistently during normal operations?
  • Are AI observations understandable to supervisors and workers?
  • How often do reviewers confirm, reject or reclassify the observations?
  • Does the tool help teams retrieve guidance more efficiently?
  • Are multilingual outputs clear enough for the intended audience?
  • Does the observation move into the existing corrective-action workflow without duplicate data entry?

Measure process usefulness rather than assuming that a higher number of alerts means better safety. Excessive false positives can create alert fatigue, while missed hazards require careful investigation. Feedback from workers, supervisors, safety professionals and facilities teams should shape the pilot design.

6. Keep AI adoption aligned with broader technology controls

MOM’s workplace safety technology guidance identifies several technology-enabled controls, including electronic permit-to-work systems, vehicular safety technologies, environmental IoT sensors and heat-stress solutions. SAGE should be considered as one possible layer within a broader safety technology strategy, not as a standalone answer.

For example, a warehouse may need to address vehicle and pedestrian interaction through physical controls, operating procedures and technology systems. An AI review of photographs could help identify visible conditions, but it cannot replace traffic management, supervision or equipment safeguards. Similarly, an image-based tool may complement environmental monitoring but cannot measure every exposure from a photograph.

Preparing now for responsible testing

Singapore businesses can begin with practical groundwork: create a consistent inspection-photo standard, establish ownership for human review, classify sensitive information, document escalation rules and connect observations to permit-to-work, incident-reporting and corrective-action processes.

The upcoming professional testing of SAGE is a reminder that successful AI adoption is mainly an operational and governance challenge. The organisations that prepare their workflows first will be better placed to evaluate the technology realistically when suitable testing or deployment opportunities arise.

ISS supports businesses exploring engineering, facility management and AI automation requirements. Contact ISS to discuss how your organisation can prepare its inspection, reporting and digital workflow environment for practical technology adoption.

Sources