- Gartner projects that 40 per cent of enterprise applications will carry task-specific AI agents by the end of 2026, up from under 5 per cent just a year earlier.
Most conversations about agentic AI in the UAE still centre on customer service, software development, and back-office automation.
Gartner projects that 40 per cent of enterprise applications will carry task-specific AI agents by the end of 2026, up from under 5 per cent just a year earlier.
Yet almost none of that conversation has reached the safety function, even though environmental, health and safety management (EHS) may be one of the domains best suited to it.
That gap is worth examining, because the case for agentic AI in environmental, health and safety management (EHS) is, on paper, stronger than in most other departments.
EHS teams already operate at the centre of constant, high-volume data, including site inspections, permit-to-work records, incident logs, equipment sensors, and now, increasingly, live video analytics and IoT feeds.
What they have lacked is a system that can act on that data continuously, rather than waiting for a human to review it.
Transition from dashboards to decisions
Any safety technology designed over the past decade has largely been built to inform, not act. For example, a computer vision on a site CCTV flags a worker without a hard hat, a sensor reports a gas reading above threshold, and a dashboard shows which permits are open.
Agentic AI is a different proposition. Rather than simply surfacing information, an agentic system is designed to reason about it and initiate a response, escalating a confined-space entry without a permit, pausing a crane lift when a worker enters the exclusion zone, or triggering a heat-stress protocol when site conditions cross a defined threshold.
Much of this reasoning is now possible because of vision language models (VLMs), which can interpret what a camera or drone is actually seeing in plain-language terms rather than simply detecting a predefined object.
A conventional computer vision model might be trained to recognise unauthorised person” and “a hard hat” as separate objects; a VLM can describe the scene as a worker standing too close to an open excavation while operating machinery, in a way that captures context, not just presence.
That contextual understanding is what allows an agent to distinguish a genuine hazard from a harmless one, rather than flagging every instance of the same object.
In a UAE dairy and beverage facility, recurring breaches of hygiene standards were observed. The existing AI cameras could already detect each lapse, like a missing hairnet, an ungowned worker, a skipped handwashing step, but treated every instance as a standalone alert, with no way to tell a one-off lapse from a worker or checkpoint with a recurring pattern.
Adding the agentic reasoning layer on top of that detection let the system connect those individual alerts over time, recognising, for instance, that the same gowning station had produced three breaches in a week, and escalate accordingly rather than logging each one in isolation.
A 30 per cent improvement in compliance was visible within months, illustrating how the same reasoning layer described above plays out in practice.
The real value of agentic AI in EHS lies less in seeing more than before, and more in remembering and connecting what it has already seen, the same shift now showing up in how the UAE’s own safety regulations are evolving.
Regulatory environment
The UAE’s regulatory direction makes this a particularly relevant moment for EHS teams to consider. Dubai’s Law No. 3 of 2026 on the Quality and Safety of Buildings requires a centrally managed digital system logging inspections, maintenance and remediation across a building’s lifecycle, explicitly designed around continuous, auditable data rather than periodic paperwork.
Law No. 7 of 2025 similarly tightens contractor accountability, with penalties escalating sharply for repeat violations. The risk this creates is less about any single infraction than about patterns going unnoticed until they have repeated enough times to trigger a fine.
An agentic system is well suited to exactly this kind of pattern recognition, flagging that the same missing-PPE violation has occurred at the same checkpoint three times in a week, for instance, before it becomes a formal repeat violation on the regulator’s record.
The UAE’s seasonal Occupational Heat Stress Prevention Policy, now in application, illustrates the same pattern from a different angle. It is a regulation built on a fixed schedule rather than live conditions: work pauses for the same window each day, regardless of what the actual heat index, humidity, or individual worker state looks like at that moment. That is not a criticism of the policy, which has worked.
But an agentic system is not bound by a clock; it can correlate a thermal sensor reading with a camera feed showing a worker slowing down or pausing mid-task, and respond to that specific worker’s condition in real time, inside or outside the mandated window.
That is precisely the operating model agentic AI is built for, not a tool retrofitted to meet the requirements, but one whose basic mode of operation already matches what the regulation is asking for.
Stakes behind the shift
The scale of the underlying problem helps explain the urgency. Construction-sector studies across the GCC have found fatal occupational accident rates running higher than in North America or Western Europe, with falls, struck-by incidents and heat-related illness consistently among the leading causes.
The ILO separately estimates that 2.41 billion workers globally are now exposed to excessive heat each year, a number that climate trends suggest will keep rising, and one with direct relevance to a region where outdoor construction and infrastructure work make up a substantial share of the labour force.
But none of this means agentic AI is a simple plug-in for EHS departments, or a replacement for the people who run them. Giving a system the authority to pause work or escalate an alert raises legitimate questions about accountability, false positives, and how human oversight is preserved even as more of the initial response is automated.
An agent can detect that a worker has entered a restricted zone and halt the activity; it cannot weigh the judgment calls that follow, whether a near-miss reflects a training gap or a one-off lapse, how to handle a contractor dispute over a violation, or when a documented procedure needs to flex to an unusual site condition.
Those remain, and are likely to remain, decisions for a human safety officer.
Road ahead for agent AI adoption
What an agentic system changes is not who is accountable for safety, but how much of an EHS officer’s day is spent watching for problems versus deciding what to do about them.
The agent’s role is closer to that of a tireless first responder than a decision-maker; it notices, it flags, it sometimes intervenes within narrowly defined limits, and then it hands the harder questions back to a person.
What seems clear is that the underlying conditions, regulatory direction, data availability, and the sheer scale of occupational risk in the region are aligning in a way that makes EHS a logical, rather than incidental, candidate for agentic AI adoption.
The EHS departments that have spent years collecting safety data through cameras, sensors and inspection logs can now sit on exactly the kind of continuous, multi-source information that agentic systems are designed to reason over.
The question for EHS teams in the UAE is less whether to engage with this shift, and more how to do so in a way that keeps the human safety officer firmly in charge of the decisions that matter most.
Gary Ng is the CEO and Co-Founder of viAct, one of Asia’s top Sustainability-focused AI company that provides “Scenario-based Vision Intelligence” solutions for risk prone workplaces.
Related Posts:
Discover more from TechChannel News
Subscribe to get the latest posts sent to your email.
