Safety coaching from CCTV, incident and observation data

On a mine site, a camera catches a worker walking behind a reversing truck. Or a supervisor logs an observation, or someone reports an incident. Someone must check the event, then turn it into a lesson for the right person. Today a supervisor or a safety advisor does this by hand. They use the health, safety and environment (HSE) system and a toolbox talk. In this design an agent checks the event against the clip and the site rules, and drafts the coaching. The safety owner approves every item before it goes out. Real-time interventions, discipline and any decision about a person stay with people.

Experience behind this example

The design builds on the author's earlier risk-bot work, which found at-risk behaviour in CCTV streams.

The systems involved

The work touches seven of the fourteen layers. The full estate is on the energy and resources page.

How it runs today

The trigger is a video analytics event, an in-cab fatigue or proximity alert, a supervisor's observation, or an incident report. The work then runs in eleven steps.

Handoffs

The event passes from the monitoring centre to the supervisor and the safety advisor. Serious incidents go to an investigation team. HR and employee relations come in if conduct is involved. Then training, site leadership and the regulator.

Where reporting comes in

HSE dashboards and the monthly safety report count injuries, high-potential incidents and critical control checks. The board safety committee reads the same report. Serious incidents go to the regulator.

Cycle time and volumes

• A large site logs thousands of observations a month (typical, unsourced).
• A safety alert reaches other sites in days to weeks (typical, unsourced).

Pain points

• There is far more video than anyone can watch.
• Detectors raise false events.
• Coaching is generic.
• Lessons spread slowly between sites.
• Workplace surveillance law applies. New South Wales has the Workplace Surveillance Act 2005. Workers and unions expect consultation.
• Coaching gets mixed up with discipline. That kills trust.
• Completing a lesson is counted as learning.

How agents run it

The agent checks events and drafts coaching. It reads clips by ID only, through a restricted video tool. It writes draft coaching items to a review queue in the HSE system. It has no path to any machine, alarm, in-cab system or permit.

If the clip does not support the label, the case stays unconfirmed and nothing is drafted. A failed delivery stays visible as incomplete. It is never reported as completed coaching.

What stays with a person

• Real-time interventions for fatigue or proximity. They are safety actuation, a deterministic zone.
• Anything about discipline or a person's performance. The rule for decisions about people applies. A named person decides any significant decision about a person.
• Incident classification and regulator notification. They are legal records.
• Investigation findings. They are legal records too.
• Stopping work. That is a safety authority.

Where the evidence is kept

The original clip stays unchanged, under the video system's retention rules. The coaching item sits in the HSE system, with its approver and its links. The agent's log sits with the operator, and the agent cannot change it. Teaching videos are kept apart from the evidence, under their own retention rule.

What changes in the systems

• A new HSE record type, "coaching item", with a status and a link to the clip ID.
• A review queue for the safety owner.
• A video interface limited to fetching a clip by its ID.
• A retention rule for teaching videos.
• Notice and consultation under surveillance law, before the trial starts.

What does not change

In-cab alerts, fatigue monitoring and collision avoidance stay as they are. Incident notification and the investigation method do not change. The discipline process does not change, and the agent has no part in it.

Who does this today

No named mining or oil and gas deployment of a coaching agent was found by 15 October 2026.

• Protex AI with Bendix, a vehicle parts factory. Protex reports 84 percent fewer targeted at-risk behaviours within 60 days (GlobeNewswire, 14 January 2026) [vendor]. It is a factory, not a mine.
• Intenseye lists mining as a target market. It names no mining customer.
• HSE vendors have added AI features: Cority Cortex AI, Enablon incident management, and Intelex with COMET (27 October 2025) [vendor]. They are features, not agents.
• Forwood calls its critical risk management "AI-driven", across more than 800 sites [vendor]. The profile's sources date from October 2023. It gives no AI detail.
• DataRobot with Chevron (2 June 2026) is the nearest case of an agent entering the permit process. Its "Safe Start" agent aims to replace mission permit approval with continuous checks [vendor]. This design keeps control of work with a person.

Where this sits

The four questions place this agent as follows. Its role is create: it drafts coaching material. Its permissions are read only on detector events, clips by ID, the permit system, the roster and procedures. It writes draft coaching items to the review queue and nothing else. It has no path to any machine, alarm or control system. It runs at delegated tasks (A2) on the autonomy contract, as designed and as permitted. The safety owner approves every item. Its impact is on people. Footage shows identifiable workers, and coaching goes to a named person.

On the risk profile it is Tier 2 (enhanced). Personal data scores Medium. If coaching records ever feed discipline, the rule for decisions about people applies, and it becomes Tier 3 (strict). This design keeps them apart. It touches no deterministic zone. It starts at curated learning (L1). People choose the reviewed events it learns from.

Watch for automation bias. People stop checking a detector that is usually right. The safety owner's review must stay a real check.

Measure the result

Measure how often the detector is right, and missed events in a reviewed sample. Measure review effort, coaching relevance and delivery errors. Then measure whether the same behaviour comes back. Completing a lesson does not prove less risk.

Start with a trial

Consult workers and unions first, and give notice under surveillance law. Then take a month of past clips from one site. Run the agent on them, with nothing sent. Ask the safety owner to grade every draft. Then go live with one crew. Register the agent through the agent lifecycle, with a named sponsor.

Evidence and limits

No mining or oil and gas deployment of this agent was found. The best measured case is a factory, and its figure is the vendor's own.

The as-is is typical practice. The volumes are typical and unsourced. The surveillance law named here was not re-checked in this round.

The detector work comes from the author's own experience. The agent extension is a proposal.

Next step

Score this agent on the risk profile. Then read how the rule for decisions about people works on the deterministic zones page.

Explore this topic

Sources were read on 15 October 2026 unless a line says otherwise. [vendor] marks a vendor's claim about its own product or customer.

• GlobeNewswire, Protex AI and Bendix, 14 January 2026 [vendor].
• Protex Intelligence assistant, undated [vendor].
• Cority incident management, Cortex AI, undated [vendor].
• Enablon incident management, undated [vendor].
• Business Wire, Intelex and COMET, 27 October 2025 [vendor].
• Macquarie, investing in workplace fatality prevention software. A profile of Forwood. Its sources date from October 2023 [vendor].
• DataRobot and Chevron, agents for autonomous inspections, 2 June 2026 [vendor].
• NVIDIA video search and summarization blueprint, vendor documentation. Accessed 17 September 2026.
• Workplace Surveillance Act 2005 (NSW). Known from background, not re-checked in this round.