Replacement-in-Kind: From Ambiguity to Repeatability
One of the most compelling demonstrations was the application of AI agents to Replacement-in-Kind (RIK) versus Management of Change (MOC) determinations.
- Past challenges: Human engineers often struggled with repeatability, especially when dealing with nuanced component substitutions.
- AI progress: By 2025, large language models (LLMs) achieved near-perfect repeatability for well-understood RIK cases. Multiple models—ChatGPT, Gemini, Claude, Grok—delivered consistent outcomes, with some leaning more conservative than others (a plus in safety contexts).
- Practical workflow: AI agents can now lead junior engineers step-by-step, asking clarification questions (e.g., about system pressure protection, sizing, and documentation) and guiding them to a justified conclusion.
- Key outcome: RIK determination has become a “well understood” AI use case, supporting reliability while reducing the risk of oversight.
P&ID Data Extraction: Turning Drawings into Usable Knowledge
Another breakthrough is the ability of AI to extract features directly from Piping & Instrumentation Diagrams (P&IDs).
- 2024 limitations: Early attempts produced inconsistent results, requiring heavy human verification.
- 2025 improvements: AI agents can now reliably recognize and categorize equipment, nodes, and control features within P&IDs.
- Impact: This accelerates downstream tasks like risk assessment, maintenance planning, and system audits—areas traditionally slowed by manual interpretation.
HAZOP Node Analysis: From Checklists to Intelligent Insights
Dr. Hoff also presented advances in AI-assisted HAZOP analysis, a domain long considered too complex for automation.
- Initial capability: AI could draft basic hazard identification checklists.
- Current capability: Agents can now complete full HAZOPs on agitator nodes, identifying design requirements, critical parameters, and potential hazards. Beyond flagging issues, AI generated complete design recommendations, including baffles, power requirements, efficiency calculations, and mixing performance.
- Benchmark result: For complex engineering problems, AI achieves correct outcomes about 50% of the time—comparable to human effort, but at a fraction of the cost and cycle time.
The Bigger Picture
Dr. Hoff’s work underscores that AI in process safety is no longer experimental—it is reliable, repeatable, and applicable to real-world scenarios. While large corporations have yet to prioritize process safety applications in their AI programs, the benchmarks demonstrated at Calgary prove that targeted use cases such as RIK determination, P&ID feature extraction, and HAZOP analysis are already delivering measurable value.As distilled AI models become deployable on local infrastructure, the barriers to adoption will shrink further. In Dr. Hoff’s words, the question is no longer if AI can support process safety, but how quickly organizations will embrace it.
