The quarterly safety review at most manufacturers is a ritual conducted in the past tense. The recordable rate is up or down against last year, the lost-time count is compared to plan, the workers' compensation trend gets a slide, and everyone in the room understands, without saying it, that every number being discussed describes harm that has already happened to someone. The uncomfortable arithmetic underneath the ritual is that a single plant may go quarters between recordables, which sounds like success and is also a measurement problem: when the numerator of your key metric is nearly zero, the metric cannot tell you whether you are good or lucky, and it certainly cannot tell you what to fix this month. This article is about building the other kind of safety Key Performance Indicator (KPI), the leading kind, out of continuous detection data, and it belongs to our full guide on AI for workplace safety in manufacturing.
OSHA's case for leading indicators
The regulator itself has been pushing this direction for years. The Occupational Safety and Health Administration's (OSHA) guidance, Using Leading Indicators to Improve Safety and Health Outcomes, describes leading indicators as "proactive and preventive measures that can shed light about the effectiveness of safety and health activities," in contrast to lagging indicators, which alert an employer to failure only after it has occurred. The national statistics show why the distinction matters at the level where Environmental, Health and Safety (EHS) directors live. The 2024 injury survey put private industry's total recordable case rate at 2.3 per 100 full-time equivalent workers, the lowest on record, and the 2024 fatality census still counted 5,070 deaths, 353 of them in manufacturing. Rates that low are genuinely good news and genuinely useless for steering a single site month to month, because at 2.3 recordables per hundred worker-years, a 400-person plant expects nine recordables a year, and no statistician on earth can extract a trend from nine events scattered across four quarters and a dozen root causes.
So the case is settled in principle, and nearly every mature program has tried to act on it. The problem has been the data. Traditional leading indicators, training completion rates, audit scores, observation counts, near-miss reports, all measure activity of the safety department more than exposure of the workforce, and the ones that do measure exposure depend on people writing things down. Near-miss reporting, the best of them, is filtered through every pressure a production floor generates, which is the subject of our companion article on why self-reported safety data understates exposure. What has been missing is a leading indicator with the properties the lagging ones have: counted automatically, continuously, and identically across sites, with no human deciding in the moment whether an event was worth recording.
Continuous detection is the first data source that meets that description, and turning it into KPIs is mostly a design exercise, which is what the rest of this article walks through.
Four indicators worth reporting
A site running the detection stack described across this series accumulates a stream of timestamped, located, classified events: Personal Protective Equipment (PPE) absences by zone, hazard-zone entries, vehicle-pedestrian near passes, dwell violations. Raw, that stream is operational noise. Aggregated with a little care, it becomes four indicators worth standing behind in a quarterly review.
Exposure rate per zone per shift is the foundation. Take the count of safety-relevant detections, normalize it by the hours the zone was active, and you have a number that behaves like a recordable rate but moves weekly instead of yearly: hazard-zone entries per hundred operating hours, PPE absence rate in the press bay by shift. The normalization matters more than it looks, because a raw count rises with production volume and punishes busy periods, while a rate isolates the thing you actually manage.
Trend against the site's own baseline is the second, and it is where the discipline lives. The first weeks of detection data, gathered silent before anyone reacts to it, become the reference line, and every indicator thereafter is read as movement against it. This sidesteps the trap that kills most new-metric programs, which is comparing sites or zones against each other before anyone has verified the cameras, coverage, and thresholds are comparable. A zone improving against its own baseline is a fact. A zone "worse than" another zone might just have better camera angles.
Near-miss ratio, detected versus reported, is the third, and it is the one that changes meetings. When the same vehicle-pedestrian conflict shows up 40 times in the detection record and twice in the reporting system, the gap is a measurement of reporting culture itself, zone by zone, and it converts the eternal argument about whether people report into a number that can be tracked and improved.
Time-to-correction is the fourth, and it turns the indicator program back on management. A detection is an observed condition; the interval between the condition appearing and the condition being fixed, the guard restored, the aisle cleared, the barrier installed, is a measure of the organization's response capacity, and it is the leading indicator most predictive of whether the program is real. OSHA's leading-indicator guidance is explicit that indicators should measure the effectiveness of prevention activities, and nothing measures it more directly than how long a known exposure stays open.
The quarterly review, rebuilt
It helps to picture the artifact this program produces, because the format disciplines the content. The safety section of a quarterly business review built on these indicators runs four slides, and each one answers a question an executive actually has.
The first slide is the exposure map: the site's zones ranked by exposure rate, this quarter against last, with the movers highlighted. Its job is to answer "where is our risk," and for the first time the answer is measured rather than remembered. The second is the correction ledger: every exposure above threshold, the fix applied or planned, and the time-to-correction trend. Its job is to answer "are we responding," and it is the slide that converts the safety function from a reporter of problems into a demonstrator of throughput. The third is the reporting-culture ratio, detected versus reported events by area, trending. Its job is honesty about the site's own information quality, and it is the slide that stops the near-miss count being presented as if it were a fact of nature. The fourth is the lagging outcomes, recordables and Days Away, Restricted, or Transferred (DART), presented last and explicitly as the trailing consequence of the first three. Over enough quarters, the causal story assembles itself on that slide, exposure down then injuries down, and that sequence is the entire argument for the program.
The same four numbers change the insurance conversation, which is worth planning for deliberately. Underwriters price what they can verify, and a manufacturer that walks into a renewal with continuous exposure data, a documented correction ledger, and a falling trend is presenting evidence of control rather than assertions of culture. No carrier commits in advance to what that is worth, and the experience of sites that have tried is consistent on one point: the conversation changes when the data is continuous and third-party-verifiable rather than assembled by hand the week before renewal. The same ledger, as the recordkeeping article in this series notes, also documents what stayed unfixed, which is one more reason the correction side must be resourced to match the measurement side.
Scaling to multiple sites
The corporate EHS function that sees one plant's indicator program work will want it fleet-wide by the next planning cycle, and this is where a quiet technical discipline decides whether the fleet gets a measurement system or a leaderboard built on sand. Comparability across sites is not automatic. Two plants with different camera coverage, different zone definitions, and different thresholds produce exposure rates that are not the same quantity, even when the column header matches, and ranking sites on non-comparable numbers is worse than not ranking them, because it rewards the plant with the worst coverage.
The fleet rollout that works standardizes the definitions rather than the numbers: a common event taxonomy, a common normalization basis of operating hours, a documented commissioning standard for coverage and thresholds, and a rule that cross-site comparison begins only after a site's configuration has passed that standard. Until then, every site reports against its own baseline, which is the honest unit of progress anyway. Corporate gets its consolidated view, but the consolidation is of trends, each site against itself, rather than of absolute rates, and the plants stop fearing that a camera upgrade will make them look worse on a chart. It is slower than the leaderboard. It is also the version that survives its second year.
Rollout mistakes to avoid
A word of caution from the sites that have run this transition, because the failure modes are organizational and they are predictable.
Do not wire the new indicators to individual consequences, and say so up front. The moment zone-level exposure rates feed anyone's performance review, the incentive to game coverage appears, cameras get bumped, thresholds get argued upward, and the data quietly rots. The indicators measure conditions and systems, the same posture the whole monitoring program depends on, and the written no-discipline commitment covered in our article on works councils and worker privacy applies to metrics exactly as it applies to clips.
Do not report an indicator you have not validated. Every threshold change, camera move, and zone redraw shifts the counts, so the first quarter is for stabilizing configuration and the baseline, not for publishing. A leading indicator that visibly jumps because someone re-tuned a confidence threshold discredits the whole family, and the audience that matters, plant managers who have seen metric programs come and go, will not extend second chances.
Do not let the indicator count grow past what a monthly meeting can act on. Four good numbers reviewed monthly beat a dashboard of thirty, because the purpose of a leading indicator is to trigger a decision, and a wall of metrics triggers nothing. The right test for adding a fifth is naming the decision it would change.
And keep the lagging indicators exactly where they have always been. Recordable and DART rates remain the outcome measures the program answers to, the numbers OSHA, insurers, and boards speak in. The leading indicators exist to move them, and the story a mature program tells runs in one direction: exposure rates fell, then the injury rates followed, and here are both curves.
How VIDIZMO supports the measurement layer
VIDIZMO AI Live Insight generates the event stream this whole approach depends on, from detection running on the plant's existing cameras, processed on the customer's own hardware on site, with every event carrying its timestamp, camera, detection type, severity, and confidence. Surveillance reporting aggregates camera state and alert volumes with their dispositions, which is what turns the operational feed into something a safety committee reviews on a cadence, and reports filter and export to CSV, so the exposure rates and ratios land in whatever the site already uses for its scorecards rather than in one more portal. Where events should flow onward automatically, into an EHS system of record or a BI stack, the REST API and webhooks carry them, a path covered in detail in our article on routing safety events into the EHS system of record. The clips behind every count stay in the Nexus portal the deployment works with, under access control and retention, so any number on the quarterly slide can be traced back to the evidence behind it.
One practical note on freshness: analytics accrue on an interval rather than in real time, which is the right design for a measurement layer, since a KPI that twitches by the minute invites exactly the wrong kind of attention. The live alerting path is where the seconds matter; the indicators are where the months do.
The starting point is smaller than most programs assume. One zone, one quarter, the four indicators, reviewed beside the incident numbers the site already trusts, and the leading half of the safety review stops being aspiration and starts being a page with numbers on it.
