Ask a safety director how their program is performing, and the answer usually arrives as injury rates. It is a precise number, it is comparable year on year, and it describes events that have already hurt someone. Measuring a safety program by its failures is a strange way to prevent incidents.
Safety leading indicators exist to solve this. They measure the conditions and behaviors that precede harm, on the theory that improving safety works best when you intervene on precursors rather than count casualties. The theory is sound. The practice has been held back for decades by a problem that has little to do with health and safety and everything to do with data collection.
Key Takeaways
- Lagging indicators measure outcomes that already happened, such as injury rates and property damage. Leading indicators measure precursors you can still act on.
- The academic evidence on leading indicators is mixed, largely because the data depends on voluntary human reporting.
- Continuous computer vision changes the measuring instrument rather than the theory, producing a census of precursor events instead of a sample.
- Five indicators are measurable continuously: near misses, PPE compliance, zone incursions, obstruction duration, and control effectiveness.
- Report at zone and shift level, never per named individual, for both worker consent and EU AI Act reasons.
Idea to AI vision app in seconds.
Leading vs Lagging: What Each Actually Tells You
Lagging indicators measure outcomes that have already occurred, such as recordable injury rates, lost time, property damage, and compensation claims. Safety leading indicators measure precursor conditions, such as near misses, hazard observations, and compliance rates, which occur before harm and can still be acted upon.
The distinction is borrowed. Leading and lagging indicators originated in business performance measurement, where customer satisfaction is the textbook leading indicator and quarterly revenue the lagging one. Safety adopted the framework for the same reason finance did, because managing exclusively on outcomes means managing too late.
| Lagging | Leading | |
|---|---|---|
| Measures | Injury rates, lost time, property damage, claims | Near misses, unsafe conditions, compliance rates |
| Available | After harm occurs | Before harm occurs |
| Data quality | High, because reporting is mandatory | Variable, because reporting is voluntary |
| Use | Accountability and benchmarking | Intervention and prevention |
Both matter. Research from NIOSH examining employer safety data found associations between the two categories and concluded that both leading and lagging measures are necessary for genuine safety improvement. The difficulty has never been whether leading indicators are conceptually right. It is whether anyone can measure them reliably.
Why Safety Leading Indicators Have Underdelivered
The evidence base is more contested than most vendor material admits, and the reason is instructive.
Salas and Hallowell, studying 261 contractors, found near-miss reporting and safety audits among the indicators empirically predictive of improved safety performance. Yet a systematic review of indicator use across utility industries notes that earlier work found the relationship between safety management practices and near misses difficult to establish, specifically because of the probability of underreporting. Other researchers have gone further, questioning whether leading indicators measured at one point predict outcomes later, and warning that the cyclical patterns they observed were unlikely to produce long term sustained improvement.
Read those findings together, and a pattern appears. Leading indicators depend almost entirely on humans choosing to record things. A near miss becomes data only when someone notices it, recognizes it as significant, and files a report that may reflect poorly on their own crew. Hazard observations happen when an inspector walks a route, which samples a fraction of a shift across a fraction of the site, leaving a gap between policy and practice that no dashboard reveals.
The weakness in leading indicators is not the concept. It is that the measuring instrument has been human attention, applied intermittently, with a built-in disincentive to report.
What Continuous Vision Changes About Measurement
Computer vision alters the instrument rather than the theory. AI-powered safety systems running on camera infrastructure you already own can observe every shift, in every monitored zone, without depending on anyone deciding to file anything.
Three properties follow, and they are the reason this matters for data-driven safety programs:
- Census rather than sample. Safety audits and inspections observe a slice of activity. Continuous monitoring observes the population, which removes sampling error from the denominator.
- No reporting incentive. Automated logging does not care whose crew was involved, so the political friction that suppresses near-miss reporting disappears.
- Consistent definitions. A proximity threshold applied by software means the same thing on Tuesday as it does on a Friday night shift, and the same thing across sites, which makes benchmarking meaningful.
That last property is what turns indicators into comparable management information rather than anecdote.
Metrics That Vision Systems Track Continuously
- Near-miss frequency and severity. Proximity events between people and vehicles, logged with distance and closing speed rather than a subjective account. This is the canonical leading indicator and historically the least reliable, which makes it the biggest gain.
- PPE compliance rate. Continuous verification produces a rate rather than a pass or fail from a spot check, letting you see drift before it becomes a violation.
- Restricted zone incursions. Zone breaches counted per shift, which reveals whether controls are respected or routinely worked around.
- Escape route availability. Cumulative minutes that a walkway or exit spends obstructed, converting a binary audit finding into an exposure measure.
- Control effectiveness over time. Whether an intervention actually moved the numbers, measured against a baseline the system established before the change.
What One Construction Program Revealed
A construction firm monitoring PPE compliance and worker-equipment proximity recorded 145 safety risks over six months at its first site, most involving workers entering the blind spots of active excavators. A second site logged more than 300 risk events in a single month.
The gap between those two figures is itself the finding. It exposed variance between subcontractor crews that no manual process had surfaced, because no manual process was watching consistently enough to compare. More useful still, risk clustered in the first hour of operations each day, which pointed at handover routines as the specific target for revised safety training.
None of that is available from injury rates. It emerged from aggregated precursor data, which is exactly what a leading indicator is supposed to provide.
Indicator to Intervention
Measurement earns nothing on its own. Programs that reach genuinely predictive safety tend to run a tight loop: establish a baseline, target the highest-frequency precursor, deliver a specific intervention, then measure whether the indicator moved. The visibility gained matters only if it changes what happens on the floor.
This is also where the financial argument becomes concrete. Teams that track leading indicators can model avoided cost rather than assert it, which is the difference between a safety case and a measurable return. Tying safety performance to business performance stops being rhetorical once precursor frequency is a number you can trend. The underlying stakes are not small: the International Labour Organization estimates that nearly three million workers die each year from work-related accidents and diseases, at a cost the ILO places at roughly 4% of global GDP.
Governance and the 2026 Regulatory Position
Continuous measurement of worker behavior carries obligations, and the timeline moved this summer. The Digital Omnibus on AI, adopted by the European Parliament in June 2026 and confirmed by the Council shortly after, deferred high-risk obligations for employment and worker management systems from August 2026 to December 2027. Transparency duties still apply from August 2026, so the deferral buys runway rather than relief.
Two design constraints follow. Emotion inference in workplace settings has been prohibited since February 2025 outside a narrow medical and safety exception, which draws a firm line between measuring conditions and inferring worker states. And indicators should be reported at aggregate level, by zone and shift rather than by named individual, since a system that produces per-person scorecards will lose worker consent long before it loses a regulator’s approval. Processing on-premise or at the edge, with retention scoped tightly, keeps most of this manageable.
Where Measurement Goes Next
Current systems count events against defined rules. The direction of travel is toward systems that interpret situations, using vision language models to reason about context, and eventually toward agentic computer vision that decides on a proportionate response without waiting for review. Indicators become richer as the perception layer gets better at judgment.
To identify which leading indicators your existing cameras could produce, our AI Safety Audit assesses current coverage and highlights the fastest routes to a usable baseline. Teams weighing the operating case may also find our view on what safety leaders ask first useful.