AI-Assisted HAZOP Preparation: Keep the Engineering Evidence and Review Responsibility Clear
Define source documents, review responsibility and service-continuity limits before using AI to support HAZOP preparation.
An AI tool can help organise a large set of engineering documents, but a convenient summary is not the same as a verified study input. A missing drawing revision, an unsupported assumption or an unavailable service can affect how useful the preparation is. For a HAZOP team, the practical question is where AI assistance fits within an engineering process whose responsibility remains with competent people. This article addresses preparation and information management, not automated hazard acceptance or a replacement for a facilitated study.
Joint guidance published through ASD’s Australian Cyber Security Centre on 23 October 2026 discusses AI token consumption and operational resilience, including limits and continuity planning. That is the recent context for this discussion. The guidance is not a HAZOP standard, and this article does not claim that an AI product meets a particular engineering assurance requirement. The review framework below is our practical analysis for teams considering support tools within their authorised workflow.
Define a bounded preparation task
Start with a task that can be described and checked. Examples might include organising an approved document list, identifying references in a supplied report or drafting an index for human review. Do not begin with a broad instruction to assess whether the design is safe. A bounded task has a defined input, expected output and acceptance check, making it easier to detect omission or unsupported interpretation.
Record what the tool is not being asked to decide. It should not silently determine the design basis, accept a risk or close an engineering action. Those decisions require the relevant project process and competent authority. A clear task boundary also helps reviewers understand how much reliance to place on the output. The fact that a tool can produce a polished technical paragraph does not establish that the paragraph is correct.
Control the source documents
Use documents the project is authorised to provide to the chosen service. Consider confidentiality and information-handling requirements before uploading drawings, operating records or other project material. Do not assume that an external AI tool is an approved repository simply because staff can access it. Follow the organisation’s applicable arrangements and keep sensitive details out of informal prompts and shared demonstrations.
Prepare a source list identifying titles, revisions and dates. Keep that list with the output so a reviewer can understand the evidence boundary. If a drawing is superseded during preparation, assess whether the output needs to be regenerated or corrected. A summary based on an old revision should not appear current merely because it was produced recently. Generation time and source currency are different facts.
Preserve the difference between extraction and interpretation
Extracting an equipment tag or document reference is different from interpreting its safety significance. Identify which kind of work each output section contains. A tool may combine the two without making the transition obvious, particularly when asked to produce a concise narrative. Reviewers should be able to trace factual statements to the supplied sources and recognise where engineering judgement has been added.
Require unsupported statements to remain unresolved rather than accepting a plausible completion. If the source does not identify an operating condition or safeguard function, the output should not invent one. A useful preparation note can flag that information as missing. The team can then assign the question to the appropriate discipline instead of carrying an unverified statement into the workshop as though it were part of the design basis.
Make review evidence specific
Choose acceptance checks that address the intended use. For a document index, verify completeness against the approved source list. For an extracted reference table, check selected entries and any high-consequence item against the original. A review that only asks whether the summary reads well is not enough for a technical preparation task. The check should test accuracy, omissions and scope.
Record the reviewer and the result without overstating the assurance provided. A sample check is a sample check, not proof that every statement has been verified. Where the project requires a more comprehensive review, follow that requirement. Keep corrections visible so the team learns which tasks or source formats produce unreliable output. Do not erase the distinction between an initial generated draft and the accepted preparation record.
Plan for service limits before a workshop
The October guidance highlights that AI services may become unavailable or degraded when usage limits are reached. A workshop should not depend on a live external service to retrieve essential study inputs. Decide what information must be available in the approved project environment before the session starts. Confirm that participants can use the drawings and records needed for the actual study without relying on a successful AI request.
Treat the fallback as an ordinary preparation requirement. Keep the accepted outputs and source references in a usable format under the project’s document controls. If the tool is unavailable, the facilitator should know which tasks can proceed and which information is genuinely missing. A continuity plan should not rely on bypassing quotas, changing access or purchasing emergency capacity without the required authorisation.
Keep workshop decisions with the study team
During the workshop, distinguish information retrieved from a source from a suggestion generated by a tool. The team should examine credible causes, consequences and safeguards using the applicable study method and project inputs. Do not use a generated list as evidence that every relevant scenario has been considered. The completeness and quality of the study remain matters for the competent team and its agreed process.
Avoid assigning protection credit or quantitative risk reduction from an AI response alone. The supporting assessment, criteria and evidence need to be established through the appropriate engineering work. This article does not prescribe a risk threshold, safety integrity level or relief-system design. Its purpose is to keep the provenance and responsibility of preparation clear so the workshop is not built on hidden assumptions.
Retain an auditable preparation record
Keep the approved source list, task brief, generated draft, review corrections and accepted output where the project’s process requires them. Record relevant limitations without copying confidential material into an unrestricted log. The objective is traceability: another engineer should be able to understand what the tool did and what a person checked. A final document that hides the preparation route can create an unjustified impression of assurance.
If a later design change affects the preparation, record how the output was reassessed. Do not treat the accepted summary as a permanent substitute for the underlying source. It may be a useful navigation aid while still needing revision when the design changes. The same discipline applies when a different model or software version is used for a repeat task.
Start with a small, reviewable trial
Select a non-consequential preparation task with appropriate authorisation and compare its output with an established manual result. Assess accuracy and review effort as well as speed. A tool that saves drafting time but creates a difficult verification burden may not be useful for that task. Do not translate a successful trial into a claim that the technology can perform an entire safety study.
The practical benefit of AI assistance is conditional on a clear boundary, reliable source control and competent review. It may help organise information while leaving engineering responsibility unchanged. Evaluate that benefit through the actual preparation task and its checks, without promising a percentage productivity gain or implying that a generated output is a completed HAZOP.
Putting the review into practice
Use HAZOP support to discuss the preparation boundary, HAZID support for earlier hazard-identification needs and process engineering for the underlying process inputs. These services do not convert generated text into verified engineering evidence.
Source: https://www.cyber.gov.au/publication/managing-ai-token-consumption-for-operational-resilience. Source date: 23 October 2026. Reviewed 4 November 2026. This article uses an earlier verified relevant development; no meaningful same-day announcement is claimed. The practical framework is original analysis, not a source quotation.