AI HAZOP: what AI can do in a HAZOP study – and what stays with the team

A HAZOP study is labour-intensive: a multidisciplinary team works through the P&ID node by node and guide word by guide word. Much of the effort goes into preparation and documentation – and that is where AI can help.

The gain is less about “automating” HAZOP than about shifting the work: the team starts from a reviewable draft instead of a blank HAZOP worksheet and can spend session time on the difficult engineering questions.

• Know what is not in the documents: operating experience, undocumented modifications and known weak spots are known only to the team.
• Avoid plausible but wrong statements: language models can phrase technically incorrect scenarios convincingly. Every statement has to be checked.
• Prove completeness: AI can work through many combinations, but it cannot prove that no relevant scenario is missing.
• Judge risk and tolerability: how severe a consequence is and which residual risk is acceptable is decided by the operator according to its own procedures.
• Take responsibility: accountability for the study and the agreed actions remains with the operator and its team.

General-purpose chat tools can list typical deviations for a pump or a vessel, and that can be a useful prompt for discussion. They are not a HAZOP study: they do not know your P&ID or your operating modes, they do not work node by node against a design intent, they do not record review decisions, and they may produce confident but wrong scenarios.

There is also a data question: P&IDs and process descriptions are confidential. They should only be processed in an environment whose hosting, subprocessors and use of data for model training are governed by contract.

Attempts to automate HAZOP go back decades, from rule-based expert systems to knowledge graphs and, since 2023, large language models. Several studies have tested how well language models generate HAZOP scenarios.

• Lee et al. (Safety Science, 2026) had four multimodal language models generate HAZOP worksheets for the same P&ID and compared them with an expert reference. The text was highly similar to the reference, yet only 19–37 % of the generated scenarios were technically valid, and the suggested safeguards leaned heavily towards procedural measures.
• Elhosary and Moselhi (Process Safety and Environmental Protection, 2026) grounded the models in 6,120 past HAZOP records and 1,140 incident reports. That raised the share of valid scenarios by roughly 10–18 percentage points – to 48 % at best.
• Dokas (Safety Science, 2026) found that the quality of LLM hazard analyses varies between repeated runs and that causal reasoning remains weak. There is no established benchmark for AI quality in HAZOP yet.
• A review of 18 HAZOP software tools (IChemE Hazards 34, 2024) names hallucination and reduced accuracy as the key challenges.

The studies consistently describe AI as support for an experienced team. None of them shows a language model producing a complete, valid HAZOP without expert review.

Three requirements follow for AI HAZOP software: plant context instead of free-text prompts, visible assumptions, and mandatory expert review of every suggestion.

1. Capture plant context: plant profile, process description, substance data and the P&ID are the basis.
2. Read the P&ID: image analysis recognises components and tags in scanned drawings, image PDFs or photos; machine-readable DEXPI XML is imported directly.
3. Prepare nodes and drafts: for each node, HAIZOP drafts deviations, causes, consequences without safeguards, existing safeguards and recommendations.
4. Review gate: the team accepts, corrects or rejects every suggestion. Nothing reaches the report without expert approval.
5. Report and audit trail: export as a HAZOP report; the audit trail records who reviewed, changed and approved what.

The core platform runs in a documented EU cloud environment. Customer data is not used to train AI models unless this is expressly agreed separately. Details on hosting and contract documents are on the security page; the feature set is on the platform page.

• Review gate: no suggestion reaches the report without expert approval.
• Traceability: assumptions, context and uncertainties are visible with each suggestion.
• Risk without and with safeguards: the software keeps unmitigated risk separate from residual risk.
• Your own risk matrix: your company’s levels and tolerability limits can be configured.
• P&ID formats: scanned or legacy drawings are usable, not only new CAD data.
• Audit trail and versions: changes are recorded with person, time, old and new value.
• Data governance: hosting, subprocessors and training-data use are governed by contract.
• Connected methods: LOPA, SIL assessment and management of change can build on the same study data.
• Language: interface, suggestions and reports in English and German.

IEC 61882 describes HAZOP as a method – team, nodes, guide words, assessment and documentation. AI changes the preparation, not the method: what matters is that the team carries out, reviews and owns the study.

Transparency and human oversight are core principles of the EU Artificial Intelligence Act. HAIZOP implements them with labelled AI suggestions, a mandatory review gate and an audit trail. How the regulation applies to a specific use in your company is a legal assessment for the operator. See also our AI transparency notes.