Lone-worker safety, AI that prevents incidents

The recordable-incident rate is a lagging indicator. By the time TRIR moves, the incident already happened. The leading indicator that actually matters is the number of unsafe dispatches that didn’t happen because the system caught the mismatch before the truck rolled. That is the safety metric AI can move, and the one that distinguishes AI that prevents incidents from AI that records them.

Recording vs. preventing

Most “AI in safety’ products in oil & gas are recording tools. AI-generated JSAs from a template library. Auto-summarized incident reports. Near-miss trend dashboards. Computer vision on body-camera footage flagging unsafe behavior in retrospect. These have value. They are not prevention. They document what happened. Prevention happens at dispatch, before the truck leaves the yard.

The distinction sounds semantic until you trace which AI artifact is in front of which human at which moment. A post-incident report is read by a safety committee in a review meeting two weeks after the event. A dispatch- enforcement decision is in front of the foreman at 5:45 AM when they’re assigning crews. The first is for organizational learning. The second is the only one that stops a person from getting hurt today.

Recording AI documents the incident. Preventing AI is the reason the incident wasn't on the schedule today.

The lone-worker exposure

Upstream and midstream operations run a meaningful share of their work as lone-worker activity. A pumper visiting forty wells across a half-million acres of leased land. A pipeline inspector walking a right-of-way alone. A meter reader on a distribution main. The compliance answer is hardware (G7, SoloProtect, phone-based check-ins) plus a process that requires GPS pings on a cadence and triggers escalation if a check-in is missed. That works for what it does, which is detect a man-down event quickly. It does not prevent the mismatch that put the lone worker on the wrong asset in the first place.

The agentic approach to lone-worker safety doesn’t replace the hardware. It wraps it in a work loop where the decision to dispatch a worker to a specific asset already considered the asset’s hazard profile, the worker’s qualifications, the weather, and the route’s communication coverage before the dispatch was approved.

Dispatch-enforced qualifications

The most operationally consequential safety control we ship is the simplest: the optimizer treats crew qualification as a hard constraint, not a weight. A crew that doesn’t have the OQ qualification for the specific pipeline class they would be inspecting does not get dispatched to that inspection. A contractor whose ISNetworld / Avetta / Veriforce / KPA record shows an expired training certificate does not get a ticket for a hot-work job. A junior pumper without the H2S certification doesn’t get routed to a sour-gas well, regardless of how much deferred production is on the table.

This is the boring-but-load-bearing piece. Most operators already have the qualification data. It lives in the contractor management platform, in the training records, in the OQ system, in the HR system, in spreadsheets the safety team maintains. Pre-WorkSync, that data is read by humans at audit time. Post-WorkSync, it is read by the optimizer at dispatch time, every shift.

Same for site-specific instructions. The hot-work permit requirement at this battery, the noise-restriction window in this neighborhood, the no-go status of that wellhead during a workover — all attached to the dispatch automatically, not handed over verbally and forgotten.

Dynamic hazard scoring

On top of qualification enforcement, the optimizer maintains a dynamic Hazardous Task Index per task per crew per moment. The score combines the asset’s static hazard profile (sour gas, high pressure, elevated work, confined space), current conditions (weather, lightning, heat index, road conditions), the crew’s competency for the specific task, and recent incident history at that asset or that asset class.

What the optimizer does with the score is the operationally interesting part. Below a threshold, the dispatch goes through. Above a threshold, the optimizer reroutes the work to a more-qualified crew or pairs the lone worker with a check-in cadence tighter than the default. Above a higher threshold, the dispatch is blocked entirely until the condition changes (weather clears, an additional crew member is available, a permit is updated).

The JSA generated for the work pulls from the same dynamic hazard profile. It isn’t a generic JSA filled in from a template. It is the actual hazards present today at this asset for this crew, written into the safety briefing the crew sees before they leave.

What it adds up to

The integrated HSE deployment with our top-25 private producer customer (the same operator running the upstream ranked work loop in Chapter 6) moved TRIR from 1.8 to 0.3. That is the lagging indicator that the leading indicators were doing their job. Each of the prevented dispatches that led to that result is a small thing. The aggregate was an 83 percent reduction in recordable incidents.

We do not claim the TRIR move was caused entirely by the agentic dispatch loop. Safety culture, leadership, training, and the operator’s safety team did the heavy lifting. What the agentic stack contributed was a mechanism that made unsafe dispatches harder to send and made the safety team’s decisions stick at the moment of execution instead of evaporating between the safety meeting and the morning dispatch. Both pieces had to be there. Neither is sufficient on its own.

Where this lives in the work loop

Field Safety as a product is sold standalone for HSE-led entry, and it is the same module that integrates with the ranked work loop in Chapter 6. The integration is what makes the safety program operational rather than parallel. Standalone, Field Safety is a competent lone-worker / contractor-competency / JSA tool that compares well against the silo’d incumbents. Integrated, it becomes the qualification-and-hazard layer of the optimizer that ranks every task. Same architecture, depending on which entry point the operator takes.

Chapter 10 picks up the conversation a CFO would want to have next: the ROI math. How a 15 percent free-cash-flow uplift on the same headcount and a TRIR move from 1.8 to 0.3 translate into a defensible internal number for the operator building a board case for AI investment.