Safety is the first slide of every utility meeting, and the most under-tooled department in the building. One month with a safety program that finally has a brain: four years of incidents, near-misses, and observations mined for the patterns nobody could see (61% of near-misses cluster around backing and rubber-goods handling), monthly safety meeting decks generated per district from that month's actual data, a policy copilot that answers "what's the minimum approach distance?" with the cited page, and the fleet wall: bucket-truck dielectric tests, boom lubrication by engine-hours, glove retest dates, grounding-set certs, tracked like the life-safety equipment it is. Two trucks come off the road before someone trusts a boom that's out of its dielectric window.
| Without | With GridCORTEX | Δ |
|---|
| Without | With GridCORTEX | Δ |
|---|
The Zero Day follows one month at a fictional utility. The safety department is three people covering six districts and 1,400 field employees, sitting on four years of reports nobody has time to read. The stakes are injuries. A "recordable" is an injury serious enough that federal safety rules require it to be officially logged, and each one means a hurt worker, an investigation, and higher insurance costs. The month opens at 312 days since the last recordable and 91% equipment compliance, meaning 9 percent of trucks and protective gear are overdue for a required safety check. On Day 1 the software reads everything at once: 4,100 near-miss reports (close calls written up by crews in their own words), 214 injury investigations, 48 monthly meeting decks (mostly recycled), the full safety manual, and the truck maintenance records. By Day 2 it has found the pattern no person could see across that much text: 61% of the close calls involve just two activities, backing trucks up and handling the insulating rubber gloves and sleeves that protect line workers from live wires. Tuesday afternoons and two districts stand out. Recommendation 1 asks the safety director to approve a program built on that finding: monthly safety decks generated for each district from its own real events, safety talks matched to the live pattern, and a policy assistant on every crew tablet that answers rule questions with the exact page cited.
Approve it and the month starts moving. Day 3.5: six district decks go out, each built from that district's own events, with names removed. Day 5: the policy assistant answers 47 questions on its first day. Most ask how close a worker may safely get to a live wire, and every answer cites the exact manual section and the matching federal safety table. Day 7.5 brings Recommendation 2, a single tracking wall for trucks and safety equipment: 14 items are overdue or due within 30 days, including bucket trucks BT-01 and BT-04, both past the deadline for their insulation test. That test proves the boom, the arm that lifts a worker toward live wires, will not conduct electricity into the bucket. Approving the recommendation pulls both trucks off the road immediately. BT-01 passes its retest and returns. BT-04's test uncovers a boom defect needing repair, on a boom that carried a line worker the week before. Equipment compliance climbs from 91% to 99%.
Day 14.5 is the hinge of the story. A crew reports a backing close call at the NORTHGATE substation gate by voice, in 90 seconds, and the software matches the wording to seven earlier reports: same gate, same time of day, five different crews. Recommendation 3 proposes a targeted fix: require a spotter (a person guiding the driver) at the 4 matching locations, add backing cameras to 12 trucks, and issue a safety talk built from the eight reports. Day 21 brings Recommendation 4: the software has checked the week's 1,240 pre-job safety briefings against each job's known hazards and flagged three briefings missing the step that verifies power lines are grounded, plus one specifying the wrong class of insulating glove, all before the crews left the yard. The ask is to make that check permanent. The month closes on Day 30 with zero recordables, 341 days since the last one, and close-call reporting up 40%. Rising reports are the good outcome here, because they mean workers flag hazards before someone gets hurt. Next month's six district decks are already drafted. Leave the recommendations unapproved and the demo plays the conventional month instead: the recycled deck, BT-04 quietly working past its insulation-test deadline, and on Day 26 an injury at the same NORTHGATE gate, the ninth event in a cluster nobody could see.
The failure is not carelessness; it is unread data. The backing pattern sits invisible inside 4,100 free-text reports, so the close call on Day 15 simply joins seven identical ones in the database. The monthly safety meeting is last quarter's deck with the date changed, so crews tune out. BT-04's insulation-test deadline lapses inside a fleet system nobody cross-checks against safety, and the truck keeps working. On Day 26 the cluster produces its ninth event, this time an injury serious enough to log. The 312-day streak resets to zero. The company's official injury rate, the score boards and insurers watch, ticks up. Compliance sits at 91% with BT-04 still on the road, and the investigation finds what the data already knew.
The software reads the four years of reports no human ever could and surfaces the cluster, 61% of close calls tied to backing and rubber-glove handling, in an afternoon. It builds each district's safety meeting from that district's own month. It puts every safety-test deadline for trucks, gloves, and grounding gear on one wall, and anything overdue comes out of service automatically. It checks every pre-job briefing against the job's known hazards before crews roll. Every move needs a person's sign-off: the safety director approves each of the four recommendations, and the safety professionals edit and deliver every deck, talk, and fix. The result is zero recordables, 99% equipment compliance, 214 policy answers with the page cited, and close-call reporting up 40%. The warning signs finally arrive before the injuries do.
| KPI | Without GridCORTEX | With GridCORTEX | Delta |
|---|---|---|---|
| Recordablesinjuries serious enough that federal safety rules require them to be officially logged | 1 (Day 26, during a backing maneuver) | 0 | the cluster was cut off |
| The backing patternthe 61% of close calls tied to backing trucks and handling insulating rubber gear | invisible inside 4,100 text reports | found, matched, acted on | 7 earlier reports surfaced |
| BT-04 dielectric windowthe deadline for the test proving the bucket truck's boom will not conduct electricity | lapsed, truck kept working | caught; boom defect found and fixed | a line worker trusts that boom |
| Safety meeting decksthe slides each district presents at its monthly safety meeting | the same generic deck, recycled | 6 districts × their own month's events | crews actually listen |
| Policy questionscrew questions about safety rules, answered with the source page cited | call someone, wait | 214 cited answers, ~5 sec each | answered at the truck, before work starts |
| Briefing gapspre-job safety briefings missing a required protective step | found by luck, or by an injury | 4 caught before crews left the yard | the last line of defense, checked |
| Near-miss reportingclose calls reported by crews; more reports mean hazards get flagged before injuries | flat: reporting is tedious and unanswered | +40%, filed by voice, always answered | more warnings, fewer injuries |
| Fleet & equipment compliancethe share of trucks and protective gear current on their required safety tests | 91%, records scattered | 99%, one wall; overdue means out of service | life-safety gear tracked as such |
| TRIR trajectorythe total recordable incident rate, the standard injury score that boards and insurers watch | up on the injury | down 2 quarters running | the trend the board sees |
| Safety staff timewhere the three-person safety team spends its hours | building decks by hand | coaching crews in the field | 3 people, multiplied |
| Investigation posturewhat an investigation can draw on after something goes wrong | "what did we miss?" | every pattern and flag logged with its reasoning | the answer was on file |
| Culture signalwhat the program tells crews about whether anyone is paying attention | same deck, date changed | a program that visibly knows them | how culture moves |
The exposure this addresses is the serious injury or fatality event that was preceded by a pattern already sitting in your own near miss reports, unread as a population. The honest limit is that the system flags and a human decides: a safety professional validates every pattern and a supervisor decides what to do about it, which is correct, because a flag is a hypothesis and a crew briefing is an intervention.
Counted in units you already track:
Hours spent coding and counting come back to the safety team and get spent in the field observing instead.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Safety analyst labor | coding and rollup hours avoided x your loaded safety analyst rate |
| Incident cost avoided | your own direct and indirect cost per recordable and per serious event x the share you believe an earlier, targeted intervention would have prevented, a share you set and defend, not one we supply |
| Workers compensation posture | your broker's own modeled effect of a changed loss history on your experience modifier, using their numbers and not ours |
| Targeted intervention delivery | stand down and observation hours delivered to flagged crews rather than to the whole workforce x crew size x your loaded field rate |
| Investigation effort | investigations per year x hours spent assembling history for each one x your loaded rate, since the history arrives already assembled |
You pay for the scoped engagement that builds and runs this and for the integration, but the largest first cost is getting years of narrative text out of your incident management system in a usable form, which is usually messier than anyone expects. Your safety team also spends real time validating early patterns, and you should budget the governance work of agreeing, in writing and with the union, what a flag may and may not be used for.
Payback here does not rest on labor, and pretending it does will cost you credibility. The analyst hours are a small, auditable floor; the case rests on your own cost per event and your own judgment of the preventable share, so state both assumptions on the first page of the business case rather than burying them.
This is the direct case. The failure modes it addresses are the ones that hurt people in electric work: approaching an energized conductor inside the minimum approach distance, working equipment believed to be clear that is not, and grounding at the wrong points or the wrong number of points. The assistant puts the current switching state and the applicable rule in front of the technician at the equipment, before contact, instead of in a manual in the truck.
Counted in units you already track:
Standing still time comes back to the crew, and interrupted time comes back to the control room operators and senior technicians who currently answer these questions by radio.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Crew standing time | lookup and call minutes avoided x crew size x your loaded technician rate |
| Control room operator interruption | call minutes avoided x your loaded operator rate, which is the most expensive interrupted hour in this list |
| Senior technician drive outs | trips avoided x hours per trip x your loaded senior technician rate, plus miles x your fleet cost per mile |
| Incident exposure | your own average total cost of a recordable electrical contact incident, including claim, investigation, and downtime, x the reduction you are willing to assume, which is your number and should be conservative |
| Documentation labor | hours per year avoided re-keying paper safety checklists x your loaded supervisor rate |
You pay for the augmented reality (AR) hardware and its ruggedization, for the GridCORTEX assistant, for integration into your advanced distribution management system (ADMS) so switching state is live rather than stale, and for your safety department's time to load and maintain your rulebook, because the answers must come from your manual and not a generic one. Your safety staff owning that content is an ongoing cost, not a one time one, and it is the cost people forget.
Payback is usually carried by crew standing time and control room interruption, because those are countable. Incident avoidance is the number the safety officer cares about and the number finance will discount hardest, so present it separately and let them set it.
The safety benefit here is indirect and specific: the crew steps out of the truck already knowing the hazards a previous crew recorded at that exact location, such as a vault that takes water, a missing animal guard, or a prior near miss. Today that information exists but nobody reads it, so the crew discovers it by encountering it.
Counted in units you already track:
Field crew time spent reading records in the cab comes back to the crew, and it comes back multiplied by the number of people sitting in the truck while one person reads.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Crew labor in the cab | record hunting hours avoided x crew size x your loaded crew rate |
| Dispatcher labor | call minutes avoided x your loaded dispatcher rate |
| Repeat truck rolls | return trips avoided x hours per trip x crew size x your loaded crew rate, plus miles x your fleet cost per mile |
| Overtime | hours avoided that would have fallen into overtime x your overtime premium, which is where the savings are largest during storm work |
| Customer minutes | your own value per customer minute of interruption x the minutes of restoration time returned per job x jobs per year |
You pay for the GridCORTEX briefing service, for read integration into your OMS, WMS, and asset history, and for delivery inside the mobile app your crews already use, because a brief in a separate app will not get read. Your own cost is mostly supervisor and crew lead time during the first weeks, correcting briefs that summarize the wrong thing so the model learns what your crews actually need at the top of the page.
Payback is dominated by crew minutes multiplied by crew size and job count, so the arithmetic is sensitive to crew size above everything else. Get that number right first and the rest of the case follows.
What is this, exactly? It is AI software: intelligent agents and models built and delivered by SoftServe, running on NVIDIA accelerated computing. It is not a hardware appliance and it does not replace the systems you run today. It deploys in your own cloud or on your premises, connects read-only to your existing systems, and recommends; your people approve every action, starting in shadow mode until it earns trust.
An analytics service for the safety organization. It mines years of incident and near-miss narratives alongside work data and surfaces precursor patterns and elevated-risk conditions as a briefing the safety team can act on. The demo above uses synthetic data; everything below describes what the real deployment needs from your organization.
| Your system | Typical products | How we connect |
|---|---|---|
| Safety incident management | Intelex, Cority, Enablon, VelocityEHS | scheduled file export (CSV or CIM XML) |
| Asset / work management (EAM/CMMS) | IBM Maximo, SAP PM, Oracle WAM | scheduled file export (CSV or CIM XML) |
| Field and crew systems | crew scheduling (ARCOS), mobile workforce tools | scheduled file export (CSV or CIM XML) |
| Weather and environment | National Weather Service feeds | read-only API |
| Document and knowledge stores | job hazard analyses, safety briefing forms | document upload |
| Advanced Distribution Management System (ADMS) | Schneider EcoStruxure ADMS, GE Vernova PowerOn | read-only API |
| Geographic Information System (GIS) | Esri ArcGIS Utility Network, GE Smallworld | scheduled file export (CSV or CIM XML) |
Runs in your own cloud account on GPU instances; incident data contains personal information, so records are anonymized on export and access is need-to-know. Connections are read-only through your existing data zone. Findings are advisory: the safety team decides every intervention.
The Approve button you just clicked in the demo above is the real workflow. This is what it looks like on the screen of the utility's safety director in the GridCORTEX console:
Accept files the briefing to your safety incident management system as an attached report via its API and queues stand-down talk assignments as pending tasks for supervisors to confirm. Supervisors schedule the talks through their own channels; GridCORTEX flags, it never directs crews.
Findings trigger automatically from incident and work data feeds; safety staff can attach context to a finding with a short note.
Incident and near-miss feeds sync nightly, work and weather data daily; each finding shows the as-of date of its data.
Lives in the GridCORTEX console for the safety team; new findings push to email, elevated-risk crew flags to supervisors' mobile. The console runs in a browser beside your existing screens on day one; embedding into your own systems is a roadmap step once the read-only phase has earned trust. Approve, Modify, and Decline are all captured in an audit trail your compliance team can pull, and GridCORTEX never blocks or overrides anything in the systems you run today.
The fair question from any safety director: "We have an incident database, a fleet system, an LMS, and safety professionals who care deeply, what's new here?" Here's the honest answer.
When someone asks "what did it actually calculate?", this is the list. In the simulation these factors drive the storyline; in a pilot they are computed from your incident, fleet, and training data.
Presenter's one-liner: "Four years of near-misses had the answer nobody had time to read: it was backing and rubber goods, Tuesdays, two districts. The AI built each district's safety meeting from its own month, put the dielectric clock on a wall where it couldn't hide, pulled two trucks before anyone trusted a stale boom, and answered two hundred tailgate questions with the page number. Zero recordables, and for the first time, the leading indicators actually led."