Using AI to Build Leading Safety Indicators on Active Sites

The Metric That Arrives Too Late

Total recordable incident rate is how most construction programs measure safety performance. It gets reported in weekly reviews, benchmarked across industry, and referenced in contract prequalification packages. By design, it is a record of what already went wrong.

TRIR tells you the incident happened, and little about whether it was coming. On a program moving fast, that delay has real cost. A crew absorbing cumulative musculoskeletal exposure during three weeks of overhead mechanical installation will not generate an incident report until week six or eight, after the critical activity has closed and the conditions that caused the injury are gone. By the time the metric moves, the intervention window has passed.

The useful move is upstream of TRIR, toward the leading indicators OSHA describes: measures that flag where exposure is accumulating before it becomes an incident.

What a Leading Indicator Requires

Leading indicators get discussed in most pre-bid safety plans and applied inconsistently on many active programs. The challenge is usually measurement more than understanding. A leading indicator deserves the name only if it produces a signal before the incident occurs, based on observation the program controls and can act on.

Near-miss reporting rates are the most common substitute. The problem is structural, and it is measured: Probst and Estrada found about 2.5 unreported accidents for every one reported across five high-risk industries, with under-reporting worst where supervisors enforced safety least. On a 500-worker site with a tight TRIR target, the workers most exposed to acute risk are the least likely to report a near-miss. They are in SIMOPS environments, under production pressure, with supervisors whose performance metrics are already under scrutiny. The data the program most needs comes from the workforce least likely to provide it voluntarily.

Remove the voluntary element, and the indicator stops filtering through willingness to report. Camera-based monitoring, wearable telemetry, and posture inference systems observe continuously. They produce an observation baseline without depending on anyone filing a report. That baseline is the precondition for a leading indicator program that reflects what the crews are doing and not only what the reporting system captured.

What Computer Vision Covers That Walkthroughs Cannot

A safety supervisor on a 600-worker site conducting two structured walkthroughs per shift is observing each active workfront for a fraction of the working day. With 20 or more active workfronts, direct observation per workfront averages a few minutes. The rest of the shift is unobserved.

Computer vision deployed on site cameras covers that gap. A model processing live feeds across three to four camera positions can classify posture risk in real time, flag workers in sustained awkward postures, identify congestion at access control points, and track PPE compliance without anyone being physically present. The data is continuous, and it runs whether the supervisor is on that side of the site or not.

A coverage gap that spans most of the working day is hard to build a real leading indicator program around, because most of what happens on a site happens in the space between observations.

Building the Program: Start Narrow

The fastest way to make an AI-assisted leading indicator program fail is to start with too much. A site team that receives 40 alerts per shift from a broadly configured system tends to stop acting on the alerts within two weeks. Alert fatigue starts when the process cannot absorb the number of signals the system produces.

Start with one risk category. Pick the highest-consequence exposure on the current program. On a data center mechanical installation, that is probably sustained overhead work by the MEP crews. Define two or three specific observable indicators: percentage of task cycles with arms above shoulder height, duration of sustained static postures over a defined threshold, and frequency of access to the elevated platform. A weekly review with the field supervisor and safety lead anchors the data to decisions rather than to reports.

The first four weeks are calibration. The system is learning which signals are noise and which are real. The supervisor is learning to trust the data. The safety lead is learning which indicators directly correlate with the risk they care about. After a month of consistent review, the signal-to-noise ratio improves and the team has field evidence for whether the indicators are moving in the right direction.

Why the Field Team Has to Be Part of It

When AI-assisted safety monitoring does not stick, the pattern is usually the same. The technology gets implemented without field team ownership. Supervisors who feel surveilled instead of supported disengage. Crew members who do not understand what is being measured comply defensively and change little about how they work.

How the program gets introduced shapes how it lands. Supervisors who understand they are getting earlier information instead of additional oversight tend to use the data as a planning tool instead of managing around it. When a supervisor can see that their crew's above-shoulder exposure increased 35 percent in the past week, and trace it to the scope change that added ductwork to the install sequence, they have a specific problem to bring to the project engineer. That conversation happens before someone files a medical report.

The durable capability is not the tool. It is a field team that knows how to act on exposure data before an incident closes the intervention window. The posts that follow in this series examine what that monitoring looks like when detection is calibrated to the specific exposure patterns that drive risk on elevated-scope programs.