Process manufacturers are increasingly looking to artificial intelligence (AI) as a tool to strengthen safety analysis. The appeal is clear. AI can rapidly search through years of operating, maintenance, alarm and incident data, compare lengthy studies for consistency, and flag recurring problems that might otherwise be missed. In Process Hazard Analysis (PHA), those capabilities could improve preparation and give experienced reviewers more time to focus on hazards that require engineering judgment.

However, AI also introduces a different kind of risk. It can produce a polished answer that appears technically sound but is based on fabricated information or faulty technical connections that do not hold up under engineering review. That becomes especially concerning in PHA because major accidents can result from rare combinations of equipment failure, abnormal operation, impaired safeguards and human action. Those conditions may never have occurred together within the historical data available to the AI system.

Those limitations do not diminish AI’s potential value but instead help define the conditions under which it should be used. Organizations need to determine what AI can be trusted to do, where its role ends and who is responsible for checking its work. Clear guardrails should be established before the technology is introduced into the analysis.

The boundaries will vary by application, as large language models, pattern-recognition systems and AI tools connected to validated engineering models do not have the same capabilities or failure modes.

Start with defined, repeatable checks

AI is already finding its way into process safety studies. The challenge now is learning how to incorporate it in ways that improve the quality of the analysis without allowing the technology to exceed its appropriate role.

Some of the clearest applications involve comparing available information against a defined requirement, rule or body of data. AI can compare a completed study against the required list of deviations, flag equipment tags that appear on a P&ID but are missing from the report and identify potential inconsistencies between consequence descriptions and their assigned risk rankings. It can also flag safeguard credits that appear to exceed the limits established in a company’s Layer of Protection Analysis (LOPA) procedures.

In each case, AI has a defined reference point against which to check the work. The risk increases when AI is asked to extrapolate beyond the information it has been given, such as deciding whether every hazard has been identified or predicting how a deviation could propagate through an interconnected process.

These checks become particularly useful when a PHA produces hundreds or even thousands of pages. Human reviewers can spend significant time correcting tag numbers, inconsistent terminology and incomplete descriptions. Removing those distractions allows experienced engineers to spend more time evaluating whether the hazard scenarios and credited safeguards are technically valid.

Use operating data to test assumptions

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

Before approving an AI-assisted task, the organization should determine whether the available data represents the equipment, operating modes and failure conditions being evaluated and whether the system produces consistent results when the analysis is repeated.

AI can organize maintenance histories, failure records, safeguard demands and overdue work to test whether the assumptions used in a PHA reflect actual operating experience. When enough relevant, high-quality data exists, AI can identify patterns and flag assumptions that the team may need to challenge.

Human teams bring process knowledge, experience and the ability to imagine undocumented scenarios, but they can also be affected by incomplete recall, inconsistent preparation and groupthink. AI can provide another line of inquiry, provided its findings remain subject to engineering review.

Know where the data ends

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

Whether the information comes from training data, company records or documents supplied for a specific analysis, AI cannot reliably account for process relationships that are missing or poorly represented.

Consider a high liquid level. AI may recognize the relationship between high level and high pressure, but it may not understand that the vessel could overflow into a knockout pot, send liquid to a downstream compressor and create a larger hazard unless those connections are represented in the available information.

The same limitation applies to safeguards. AI may identify a listed safeguard without understanding whether it is independent, can respond within the available process safety time or could be disabled by the same failure as the initiating event.

This gap can be difficult to recognize because AI-generated work may look complete. A 2026 Safety Science study evaluated four large language models that generated 120 HAZOP worksheets from the same piping and instrumentation diagram. Although BERT-based similarity scores exceeded 86% when the results were compared with an expert worksheet, only 19% to 37% of the generated scenarios were judged valid. The models also favored procedural safeguards over engineered controls.

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

Put governance in place from the start

Organizations should define how AI will be used before introducing it into a PHA, then establish clear constraints for the people working with the tool.

That plan should identify the tasks AI is approved to perform, the information it may access and the decisions it is prohibited from making.  Under a practical governance structure, process safety leaders would approve the application and its validation requirements, information technology and cybersecurity teams would control the tools and data environment, and the PHA team would retain authority over technical conclusions.

The organization should also document the AI tool and version used, the information supplied, the output produced, the human review performed and the final disposition of its findings. This creates an audit trail and allows reviewers to determine whether a later change in the tool, data or assumptions could affect the analysis.

Data access presents another challenge. Secure, isolated systems protect confidential operating and incident information, but they can limit how lessons learned at one company improve AI applications elsewhere. Public incident reports may not include enough detail about operating conditions, equipment configurations and safeguard status to support reliable technical analysis.

Khan believes standardized industry data sharing could improve AI’s ability to recognize recurring patterns and test assumptions across companies. Any shared system would still need to protect proprietary information, account for differences among facilities and prevent incomplete or biased records from becoming accepted as reliable evidence.

When an AI conclusion conflicts with the PHA team, Khan recommends an independent review of the data and assumptions behind both positions. Neither the AI output nor the team’s initial judgment should prevail without technical support.

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