AI aids process hazard checks, but humans stay in charge

Process manufacturers are using artificial intelligence in Process Hazard Analysis to help review safety studies and operating records. Industry specialists say the technology needs strict limits and human oversight.

In process safety work, artificial intelligence is being used for tasks that involve checking large volumes of information against defined rules or known data sets. In Process Hazard Analysis, that includes reviewing operating, maintenance, alarm and incident records, comparing lengthy studies for consistency, and spotting repeated issues engineers might otherwise miss.

The approach is drawing attention because PHA exercises can run to hundreds or thousands of pages. Engineers and reviewers often spend substantial time finding inconsistent terminology, missing equipment tags and incomplete descriptions before they can focus on whether hazard scenarios and safeguards are technically sound.

Defined uses

One area seen as more suitable for AI is repeatable checking against a fixed reference point. Systems can compare a completed study with a required list of deviations, identify equipment tags shown on a piping and instrumentation diagram but missing from a report, and flag cases where consequence descriptions appear inconsistent with assigned risk rankings.

AI can also identify safeguard credits that appear to exceed limits set in a company's Layer of Protection Analysis procedures. In these cases, the system is not being asked to reach new engineering conclusions, but to test existing material against defined requirements.

Concerns rise when AI is asked to go beyond that role. Deciding whether every hazard has been identified, or predicting how a deviation could spread through an interconnected process, requires technical judgment and an understanding of plant-specific conditions that may not be fully represented in the available data.

That distinction matters in major hazard industries because serious incidents can emerge from rare combinations of equipment failure, abnormal operation, impaired safeguards and human action. Historical records may not contain those combinations in a form an AI system can interpret reliably.

Data limits

The quality and scope of the data available to an AI tool are central to whether its output is useful. Organisations considering AI-assisted tasks in PHA need to establish whether the data reflects the equipment, operating modes and failure conditions under review, and whether the system produces consistent results when the same analysis is repeated.

"Where the data is sufficient, AI can augment human judgment much better than a completely manual process," said Usman Khan, P.E., SIS Engineering Authority, INEOS Olefins & Polymers USA.

AI can organise maintenance histories, failure records, safeguard demands and overdue work orders to test whether assumptions used in a hazard review match actual operating experience. Supporters argue this gives human teams another line of inquiry while leaving engineers to make the final technical assessment.

Khan warned that data gaps sharply limit what the systems can do.

"AI is subject to the fidelity of the data it is trained on," Khan said. "If it is not in the data set, AI does not know what to do with it."

That problem becomes more acute when process relationships are missing or poorly described. A system may identify a broad connection between high liquid level and high pressure, for example, but fail to recognise a more complex chain in which a vessel overflows into a knockout pot, sends liquid to a downstream compressor and creates a wider hazard.

The same issue applies to safeguards. A listed safeguard may appear in the records, but AI may not be able to determine whether it is independent, whether it can respond within the available process safety time, or whether the same initiating event could disable it.

Validity gap

The risk for operators is that AI-generated output can look complete and well structured even when its technical basis is weak. That can make fabricated or invalid conclusions harder to detect during a fast-moving review process.

A study published in Safety Science evaluated four large language models that generated 120 HAZOP worksheets from the same piping and instrumentation diagram. While BERT-based similarity scores were above 86% when compared with an expert worksheet, only 19% to 37% of the generated scenarios were judged valid, and the models tended to favour procedural safeguards over engineered controls.

"A PHA is not just a document being created," Khan said. "It is a consensus of a team of experts about how deviations progress into process hazards."

Governance first

Process safety specialists say companies should define approved uses for AI before introducing it into hazard reviews. That includes setting out which tasks the technology may perform, what information it can access and which decisions it cannot make.

Under that model, process safety leaders would approve the application and validation rules, information technology and cybersecurity teams would manage the tool and data environment, and the PHA team would retain authority over technical conclusions. Companies are also being urged to record the tool and version used, the information supplied, the resulting output, the human review carried out and the final treatment of any findings.

Such records create an audit trail and help teams assess whether later changes to the tool, the data or the assumptions could alter the analysis. Data access remains another issue because secure internal systems protect sensitive operating and incident information but can also restrict broader learning across the sector.

Khan said any conflict between an AI conclusion and the judgment of the PHA team should trigger an independent review of the data and assumptions behind both positions.

"Ultimately, responsibility remains with the plant manager or asset manager," Khan said. "AI should be considered part of the team, not a substitute for the team and not the overruling voice on the team."