Industries / Industrial equipment and services / Industrial services / Process safety consulting
Hazard studies
Hazard studies are structured team reviews of a design or a running plant that ask how each part could go wrong and what would stop it.
4
Milestones achieved, from 1963 to 2001.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1963
A phenol plant questioned
ICI used "critical examination", a method for questioning management decisions. Ken Gee, a production manager at ICI, applied it to the design of a new phenol plant. For four months he spent three days a week going through every part of the design, and found hazards and operating problems that would otherwise not have been foreseen. Trevor Kletz called this the first recognisable hazard and operability study (HAZOP), a team review of how each part of a plant could depart from its intended operation. A history poster from TU Delft's safety science section dates ICI's development of HAZOP to 1963.
Credit
Ken Gee, ICI
1964
A score for fire risk
Dow Chemical developed its Fire and Explosion Index in 1964 to rate the fire and explosion potential of single process units. The index multiplies a material factor, set by the chemicals present, by a unit hazard factor built from penalties for general and special process hazards. In a presentation hosted by Purdue University's Process Safety and Assurance Center, SABIC said it still uses the score to screen process options and to choose which hazard-study method each part of a plant gets.
Credit
Dow Chemical Company
1998
An old plant never studied
On 25 September 1998 a heat exchanger at Esso's gas plant at Longford, Victoria, Australia, had gone super-cold and brittle, and it fractured when warm oil flow was restarted. The escaping gas ignited, killing two men, injuring eight and cutting Victoria's gas supply for two weeks. The site's two other gas plants had been through a HAZOP. Exxon had a policy from the early 1990s of studying all existing plants, but Esso had deferred the HAZOP of Gas Plant 1 indefinitely on grounds of cost. Andrew Hopkins, who analysed the Royal Commission evidence, writes that the study would have identified the loss of warm oil flow and the embrittlement that followed.
LessonA hazard study put off to save money leaves the hazards it would have found in place.
Credit
the Longford Royal Commission (findings); Andrew Hopkins, Australian National University (analysis)
2001
An international HAZOP guide
The International Electrotechnical Commission (IEC) published IEC 61882, an application guide for HAZOP studies, in 2001. It sets out how to define and prepare a study, how to run its sessions with guide words (short prompts the team applies to each part of the design), and how to record and follow up the results. The second edition, of 10 March 2016, aligned its terms with ISO 31000, the international risk management standard, and added an improved case study of a procedural HAZOP, one that examines a written operating procedure instead of a drawing.
Credit
IEC technical committee 56 (Dependability)
Watch for
Worksheets drawn from one drawing
Four large language models (GPT-4o, GPT-4o mini, Llama and Gemini) each wrote a HAZOP worksheet from the same piping and instrumentation diagram (P&ID), the drawing of a plant's pipes, vessels and instruments. Scored against an expert's worksheet, all four reached F1 similarity scores above 86%. Only 0.19 to 0.37 of the scenarios they wrote were judged semantically valid, and their safeguards leaned heavily on procedures.
Achieved whena group outside the authors repeats the test on several drawings with independent expert scoring, and most of the generated scenarios are judged valid.
Claimed by
Junseo Lee, Sunhwa Park, Sehyeon Oh and Byungchol Ma, in Safety Science, February 2026
Meeting transcripts into worksheets
Kenexis, a US process safety consultancy and software maker, says that using a large language model to process the transcript of a HAZOP meeting cuts the time a HAZOP study needs by over 50%, while keeping and sometimes improving the quality of the record. The figure comes from its own comparative case study.
Achieved whena client or an independent reviewer publishes timed results from real studies run both ways.
Claimed by
Kenexis, December 2025
Open
01
Scenarios real studies missed
A public test set of 25 accidents in which the investigators found that a hazard study had missed or misjudged the scenario that happened. Each case holds only what a study team had before the accident: a redrawn P&ID, the operating conditions and the materials. The missed scenario is kept sealed as the answer. AI hazard-study tools, and one human team for comparison, are scored on whether they raise it.
The briefCredit · unclaimed02
An open test for scribes
Ten recorded mock HAZOP sessions on published teaching designs, with practising engineers taking part, each released with its audio, its transcript and the worksheet the team approved. A scoring script checks any transcript-to-worksheet tool for rows it gets right, rows it misses and rows it invents, and measures the hours a person then needs to finish the worksheet.
The briefCredit · unclaimed
Open 01 · the brief
Scenarios real studies missed
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of 25 accidents in which the investigators found that a hazard study had missed or misjudged the scenario that happened. Each case holds only what a study team had before the accident: a redrawn P&ID, the operating conditions and the materials. The missed scenario is kept sealed as the answer. AI hazard-study tools, and one human team for comparison, are scored on whether they raise it.
Why it is still open
Published tests of AI-written HAZOPs score tools against one expert worksheet, which measures agreement with a team, not whether a tool catches what teams miss. The misses are already documented: Mark Kaszniak's 2010 paper in Process Safety Progress draws its examples of hazards that analyses failed to find from CSB investigations, and Longford's missing study is described in detail. On 19 November 2026, GitHub and Hugging Face searches turned up HAZOP tools, among them TU Dortmund's preHAZOP, but no test set built from missed scenarios.
Built with what exists
CSB investigation reports; pyDEXPI, an open-source Python library for DEXPI drawing files, to redraw each P&ID; current models as the tools under test, such as Claude Fable 5.2 (Note: "Three working devices") and the open-weight Mistral Large 4 (Note: "A trillion open parameters").
First step
Redraw the warm-oil and heat-exchanger section of Longford's Gas Plant 1 from published accounts, and seal the loss of warm oil flow as its answer.
Done looks like
25 cases published under an open licence, at least three AI tools and one human team scored blind, two process safety engineers agreeing on at least 90% of the scores, and the share of missed scenarios each tool raises published.
Open 02 · the brief
An open test for scribes
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Ten recorded mock HAZOP sessions on published teaching designs, with practising engineers taking part, each released with its audio, its transcript and the worksheet the team approved. A scoring script checks any transcript-to-worksheet tool for rows it gets right, rows it misses and rows it invents, and measures the hours a person then needs to finish the worksheet.
Why it is still open
Kenexis's figure of more than 50% time saved comes from its own case study. On 19 November 2026, GitHub found no repository of HAZOP transcripts, and the only HAZOP dataset on Hugging Face was a list of guide words.
Built with what exists
ElevenLabs Scribe v3 for transcription (Note: "Fewer word errors"); a current large language model to draft worksheet rows; the CSV export of a worksheet tool such as Kenexis Open-PHA; Audacity for recording.
First step
Record one two-hour mock session on a published teaching design with four volunteer engineers, and have them approve the worksheet.
Done looks like
Ten sessions of at least two hours each, released under an open licence with every participant's written consent, and at least three scribe pipelines scored, with the share of approved rows each reproduced, the share of rows it invented and the finishing time published.