AI System for Routing Site Photos to Safety Review

The problem

Images requiring review may be noticed late when daily volumes are high. When photographs labelled with their area and time, the health and safety checklist and open observations are scattered across different screens, documents or team records, issues do not become visible in time.

This fragmented data delays preparation of the safety review queue and requires staff to search for supporting records again by hand. Before making a decision, the authorised health and safety specialist has to complete missing fields and verify conflicts one by one.

A common exception is that the camera angle or a control outside the frame may be misleading. If an incomplete or incorrect workflow runs without checking this context, a hazard may be missed or a safe area may be stopped unnecessarily; the output therefore cannot be applied directly.

What we set out to improve

• Rate of referrals verified by the specialist
• Rate of outputs corrected or rejected by the authorised health and safety specialist
• Time between the source record and the review output

How the system works

1. 01 Define the review scope and the authorised health and safety specialist role
2. 02 Receive the inputs: photograph labelled with its area and time, health and safety checklist and open observations
3. 03 Check source identity, date, version and data freshness
4. 04 Match the photograph to the site area
5. 05 Pre-classify visible indicators
6. 06 Check the exception: the camera angle or a control outside the frame may be misleading
7. 07 Place low-confidence or conflicting records in a separate human review queue
8. 08 Write the source identity, rule, output and time to the audit trail
9. 09 Have the authorised health and safety specialist approve, correct, defer or reject the referral

Methods we used

• Image pre-classification limited to visible conditions
• Source, format, version and required-field checks for photographs labelled with their area and time, the health and safety checklist and open observations
• A missing-data warning instead of a prediction when inputs are absent or conflicting
• A separate check for the exception outside the main rule
• Joint storage of the source identity, rule result, safety review queue and human decision

Where people stay involved

Data and security

Access to photographs labelled with their area and time, the health and safety checklist and open observations is restricted to the minimum permissions required for each role. Identifiers and sensitive fields are masked where possible, and only the necessary section of a file is sent to the model rather than the complete raw file. Access, output and the authorised health and safety specialist's decision are logged; production use does not begin until the organisation has approved the retention period and data location.

Who this suits

Frequently asked questions