The incident report arrived on Tariq’s desk on a Wednesday afternoon: a scaffolding near-miss at Site 2, the gas processing facility 40 kilometers outside Ruwais. A pipe fitter named Rajesh had flagged that three clamps on a scaffolding section were corroded. He’d told his supervisor. The supervisor had written it on a form. The form had traveled from the site office to the HSE coordinator by pickup truck. The HSE coordinator had entered it into the tracking spreadsheet. That spreadsheet was emailed to Tariq.
The incident happened on Sunday. Tariq saw it on Wednesday. Four days.
Nobody was hurt. Rajesh had the experience to spot the corrosion and the judgment to stop work. But four days between a field safety observation and the HSE director seeing it meant four days where that scaffolding was potentially still in use, four days where the corrective action hadn’t been initiated, and four days of non-compliance with ADNOC’s 24-hour critical observation reporting window.
“The scaffolding had been taken down by Tuesday,” Tariq said. “The site supervisor handled it. But I didn’t know that until Wednesday, and more importantly, ADNOC’s system didn’t know it either.”
Three sites, five languages, and a paper trail that ran on pickup trucks
Tariq is the HSE Director at a mid-size oil and gas operator based in Abu Dhabi’s ICAD industrial zone, with three field operations in the Ruwais industrial complex: a gas processing facility, a pipeline maintenance yard, and a storage terminal. Combined workforce across the three sites: about 1,100 people, roughly 70% of whom are contract workers.
The workforce speaks five languages on any given day. Hindi, Urdu, and Filipino are the most common among the field crews. Arabic is the language of supervision and management. English is the lingua franca of HSE documentation, because ADNOC’s reporting systems, the international standards (ISO 45001, OHSAS frameworks), and the company’s own management system all operate in English.
This language gap was at the root of the reporting problem. A pipe fitter from Kerala notices a corroded clamp. He tells his supervisor in Hindi. The supervisor, who speaks Hindi and Arabic but writes limited English, fills out a paper incident form in English, sometimes with help from a colleague. The form is handwritten. Handwriting quality varies. Technical terminology is sometimes approximated. A “corroded clamp” might be written as “rust on holder” or “bad fixing piece.”
The form travels to the site HSE coordinator, who interprets the handwriting, re-enters the data into an Excel tracker, classifies the incident by type and severity, and emails it to Tariq’s central HSE team in ICAD. Tariq’s team reviews the classification, determines if ADNOC notification is required, prepares the formal report in ADNOC’s required format, and submits it through ADNOC’s HSE reporting portal.
This process, from field observation to ADNOC submission, averaged 5 business days. For critical incidents (LTIs, spills, fires), the team could compress it to 24-48 hours by having the site coordinator call Tariq directly. But for near-misses, unsafe conditions, and observations, the standard five days applied, and sometimes longer during peak maintenance periods when the coordinators were stretched thin.
ADNOC’s HSE requirements are not suggestions. As Abu Dhabi’s national oil company and the ultimate concession holder, ADNOC sets the compliance framework for every operator in the emirate. Their Contractor HSE Management System (CHSEMS) specifies reporting timelines, incident classification criteria, investigation requirements, and corrective action tracking. Non-compliance affects not just your audit score but your eligibility for future contract renewals.
Tariq’s company scored 81% on their last ADNOC HSE audit. Passing, but not comfortably. The audit findings pointed to three areas: delayed incident reporting, incomplete corrective action tracking, and inconsistent investigation documentation. In other words, the information was there, but it moved too slowly and lost fidelity along the way.
The ADNOC digitization push and what it meant for mid-size operators
ADNOC’s In-Country Value (ICV) program and its broader digital transformation initiatives had been pushing operators toward technology adoption for several years. The introduction of ADNOC’s unified HSE reporting portal in 2024, which required digital submission of all incidents, inspections, and audit findings, effectively set a deadline: go digital or fall behind.
For large operators with in-house IT departments and enterprise HSE software, this was manageable. For mid-size operators like Tariq’s company, with 1,100 workers and an IT team of four people who also managed ERP, email, and the physical access control systems, the gap between ADNOC’s expectations and operational reality was significant.
Tariq evaluated three commercial HSE software platforms. The licensing costs ranged from AED 180,000 to AED 450,000 per year. Implementation timelines: 4-8 months. All three required the field workforce to use tablets or smartphones to submit reports through a structured application, which assumed reliable connectivity in the Ruwais industrial area (often spotty), devices that could survive 48-degree heat and industrial conditions, and that workers with varying levels of English literacy could navigate a form-based interface.
“I couldn’t justify asking a welder from Peshawar who’s been doing this job for 15 years to learn a new app on a tablet,” Tariq said. “He knows safety. He knows his trade. What he doesn’t know is how to fill out a digital form in English with dropdown menus that use terminology from ISO 45001.”
The alternative that emerged was an AI agent that could accept reports the way field workers actually communicate: voice messages on WhatsApp, in their own language, using their own terminology for safety concepts.
Building a reporting system for the field, not the office
The AI agent was designed around one principle: the person reporting an incident or observation should be able to do so in under two minutes, in whatever language they speak, without touching a form.
Voice-first reporting. A field worker opens WhatsApp and sends a voice note to the HSE reporting number. “Main scaffolding pe teen clamps ka rust hai, Gate 4 ke paas” (Hindi: “There’s rust on three clamps on the main scaffolding, near Gate 4”). The agent transcribes the voice note, translates it to English, and generates a structured incident report: observation type (unsafe condition), category (structural integrity), location (Site 2, Gate 4), equipment (scaffolding), hazard (corroded clamps), recommended severity (moderate).
The worker receives a WhatsApp confirmation in Hindi within 90 seconds: “Your safety observation has been recorded. Reference number HSE-2026-0847. A supervisor will follow up within 4 hours.”
Multilingual transcription and translation. The agent handles Hindi, Urdu, Filipino (Tagalog), Arabic, and English. Each language required its own calibration. Industrial terminology in Hindi doesn’t always have direct English equivalents; the team built a specialized glossary of 340 terms mapping common field expressions in each language to standardized HSE terminology. “Patti” (Hindi for a type of metal strip) maps to “steel strap.” “Buhay na kuryente” (Filipino for live electricity) maps to “energized electrical equipment.”
Photo and video documentation. Workers can attach photos to their voice report. The agent processes the image, tags it to the incident report, and performs basic hazard recognition: a photo of a spill gets tagged as “chemical/liquid spill,” a photo of exposed wiring gets tagged as “electrical hazard.” This is supplementary, not definitive, but it adds a visual record to every report automatically.
Automatic classification and routing. Based on the report content, the agent classifies the incident by ADNOC’s category system, assigns a severity level, and routes it to the appropriate person. Near-misses and observations go to the site HSE coordinator for review and closure. Potentially significant incidents trigger immediate notifications to Tariq and the site supervisor via Microsoft Teams. Any incident classified as potentially reportable under ADNOC’s framework gets flagged for priority processing.
ADNOC portal submission. For incidents that meet ADNOC’s reporting thresholds, the agent prepares the submission in ADNOC’s required format and presents it to Tariq’s team for review. One click to approve and submit. The agent tracks the submission status and any follow-up requests from ADNOC.
Corrective action tracking via Jira. Every incident that requires corrective action automatically generates a Jira ticket. The ticket includes the incident details, assigned responsible person, deadline (based on severity), and required evidence of closure. The agent sends WhatsApp reminders to the responsible person at 50%, 75%, and 90% of the deadline. Overdue actions trigger escalation notifications to Tariq.
The first field test and the supervisor rebellion
The pilot launched at Site 2, the gas processing facility, with 380 workers. The first two weeks revealed problems that no amount of office-based testing had predicted.
The supervisor problem. Site supervisors had been the gatekeepers of safety reporting. Every observation went through them first, which meant they could filter, contextualize, and sometimes quietly resolve issues before they reached the HSE department. The AI agent bypassed this layer: workers reported directly to the system.
Three supervisors pushed back hard. Their concern wasn’t about losing control, or at least, that’s not how they framed it. They argued that field workers lacked context for proper classification. “A new helper sees a small oil stain near a valve and thinks it’s a spill,” one supervisor told Tariq. “I can look at it and know it’s condensation from a temperature change. If every observation goes straight to the system, you’ll drown in false alarms.”
He had a point. During the first week, 34% of reports submitted directly by workers were reclassified by the HSE coordinator after review. Not because the workers were wrong about what they saw, but because the context changed the severity and required response.
Tariq’s compromise: workers could report directly, but supervisors would receive a parallel notification and had a four-hour window to add context before the report was finalized. If the supervisor added context that changed the classification, the HSE coordinator reviewed both the original report and the supervisor’s notes. If the supervisor didn’t respond within four hours, the original classification stood.
This satisfied most supervisors. It preserved their role in contextualizing observations without creating a bottleneck that defeated the purpose of the system.
The connectivity challenge. The Ruwais industrial area is not downtown Abu Dhabi. Cellular coverage is adequate in most areas but drops to intermittent near the storage terminal and certain sections of the pipeline yard. Workers would record voice notes and hit send, but the messages would sit in the WhatsApp outbox until they moved to an area with better signal.
The team addressed this with a practical solution: WhatsApp messages are queued on the device and sent when connectivity resumes. The agent timestamps based on when the message was sent by the worker, not when it was received, preserving the accuracy of the reporting timeline. For areas with persistent connectivity issues, the company installed three Wi-Fi repeaters at locations where workers congregate during breaks.
The heat factor. Between June and September, outdoor temperatures in Ruwais routinely exceed 48 degrees Celsius. Smartphones in direct sunlight overheat and shut down. Workers in PPE have limited dexterity with touchscreens. Voice reporting was designed for exactly this scenario, but in practice, workers often removed gloves to operate their phones, which is itself a safety issue in certain work areas.
The company purchased 50 ruggedized phone cases with screen protectors that work with gloved hands, and distributed them to workers in the highest-risk areas. Not a technology solution. A practical one.
The maintenance scheduling breakthrough nobody expected
The HSE reporting agent was the primary objective. But the maintenance scheduling improvement turned out to be the use case that convinced the operations manager to expand the budget.
The company’s maintenance program runs on a combination of manufacturer-specified intervals, operating-hour-based triggers, and condition-based monitoring. Before the AI agent, maintenance scheduling was managed in Excel by two planners who tracked about 2,200 equipment assets across three sites. Scheduling accuracy, defined as the percentage of maintenance tasks completed within 10% of their scheduled date, was 74%.
The 26% gap had multiple causes. Sometimes parts weren’t available. Sometimes the crew was reassigned to a higher-priority job. Sometimes the planner simply missed an item in the spreadsheet. But the most common cause, responsible for about 40% of the gap, was that condition data from the field didn’t reach the planners in time to adjust the schedule.
The HSE agent was already receiving field observations about equipment condition: corroded fittings, leaking seals, abnormal vibrations, discolored pipe sections. The team realized that this stream of field observations, once translated, classified, and structured, was essentially a real-time condition monitoring feed that could inform maintenance scheduling.
The agent was extended to tag equipment-related observations and push them to the maintenance planning system. A field worker reporting “vibration on pump P-204” generates both an HSE observation and a maintenance flag. The maintenance planner sees the flag alongside the scheduled maintenance data and can pull forward the inspection if warranted.
Maintenance scheduling accuracy went from 74% to 93% within four months. The improvement was primarily driven by better condition data from the field, not by any change in the planning methodology itself.
The numbers after six months
Six months after the full rollout across all three sites, the data presented to ADNOC during the company’s annual performance review:
HSE report processing: 5 days to 8 hours. The average time from field observation to completed, classified, and filed HSE report dropped from 5 business days to 8 hours. For critical incidents, the time from occurrence to ADNOC notification dropped from 24-48 hours to under 3 hours.
ADNOC compliance score: 81% to 96%. The next ADNOC audit, conducted five months after the full rollout, scored the company at 96%. The improvement was driven primarily by three factors: timely reporting (no more 5-day delays), complete documentation (every incident has a structured report, photos, and an audit trail), and corrective action closure rates (92% of corrective actions closed on time, up from 68%).
Maintenance scheduling accuracy: 74% to 93%. As described above. The operations manager described this as “the best unintended consequence of any HSE project I’ve seen.”
Reporting volume. Total safety observations submitted per month increased from 180 to 470. This was expected: when reporting takes two minutes via voice note instead of 15 minutes filling a paper form, people report more. The increase in reporting volume is, in HSE terms, a positive indicator. More observations means more hazards identified before they become incidents.
Worker participation. During the paper-based era, about 22% of the workforce had submitted at least one safety observation in the previous 12 months. After six months with the AI agent, that figure was 61%. The biggest increase was among the Hindi and Urdu-speaking workers, who had the lowest participation rates under the English-language paper system.
Investigation quality. HSE investigations for significant incidents now include transcribed witness statements (captured via voice note in the witness’s own language), timestamped photo evidence, and a complete chain from initial report to investigation findings to corrective actions. The investigation files that took Tariq’s team 3-4 days to compile now take about 4 hours.
What works well, and what doesn’t yet
The system performs well for the high-volume, structured work: observations, near-misses, routine inspections, and corrective action tracking. The five-language voice reporting removed the single biggest barrier to field-level participation.
Where it falls short:
Complex investigation narratives. Root cause analysis for significant incidents requires interviewing multiple witnesses, reconstructing event sequences, and analyzing contributing factors across organizational and technical dimensions. The agent can collect inputs but cannot perform the analysis. It assists by organizing data; the thinking is human.
Regulatory interpretation. When ADNOC updates its HSE requirements, or when OSHAD updates its system framework, someone has to interpret the changes and update the agent’s classification rules. This is a monthly maintenance task that requires Tariq or his senior HSE engineer.
Contractor management nuances. With 70% contract workers, corrective action ownership involves a complex web between the operator, main contractors, and sub-contractors. The agent routes based on rules, but edge cases require human judgment about who actually owns the corrective action.
Offline scenarios. Workers in truly disconnected environments (inside vessels during turnaround maintenance) cannot send messages until they exit. The team is exploring a lightweight offline-capable app for these scenarios, though Tariq is reluctant to add another system.
What ADNOC’s audit team actually said
Tariq doesn’t usually share audit feedback, but he made an exception for one comment from the lead auditor during the closing meeting: “This is the first time we’ve audited a contractor of your size and found the documentation to be audit-ready rather than audit-assembled.”
The distinction matters. “Audit-ready” means the documentation exists in real time, as a natural byproduct of operations. “Audit-assembled” means someone spent weeks before the audit pulling together evidence from various sources, filling gaps, and formatting everything to look complete. Tariq’s team had spent three weeks preparing for the previous audit. This time, preparation took four days, mostly for the presentation deck.
“The agent didn’t make us a better HSE operation overnight,” Tariq said. “What it did was make the HSE operation we already had visible. The field workers were always seeing these hazards. The supervisors were always resolving them. The problem was that none of it was documented in a way that proved it was happening.”
He paused. “ADNOC doesn’t just want you to be safe. They want you to prove you’re safe. Those are two different problems, and for years, we were solving the first one and failing at the second.”
The company is now exploring expansion of the system to include permit-to-work automation, a process that currently involves 11 steps and three levels of approval for hot work, confined space entry, and working at height. Tariq estimates that automating the routine permits (which account for about 60% of the volume) could free up two full-time equivalents on the field supervision team.
Rajesh, the pipe fitter who reported the corroded scaffolding clamps, submitted 14 safety observations in the six months after the system launched. He’d submitted one in the previous two years. When asked why the change, he said, through a colleague translating from Hindi: “Before, I would tell my supervisor and he would handle it. Now I can say it myself and know it was heard.”