Revolutionizing Offshore Safety: Learning from the Past, Protecting the Future, AI-Driven Permit to Work System with Integrated Machine Learning and Large Language Models

Abstract The Permit to Work (PTW) process is essential for safety in Upstream offshore activities. However, manual reviews can be inconsistent and time-consuming, posing safety risks and inefficiencies. To tackle these issues, AI for PTW was tested. The AI/ML technology acts as a cold eye reviewer, identifying gaps, providing historical learning from past incidents, and highlighting missing details in permits, demonstrating its potential for broader deployment. This paper presents a two-layer AI framework aimed at improving the quality of Permit to Work (PTW) submissions, particularly in job hazard analysis (JHA) and work method statements. The process involves: Data Ingestion and Indexing: Data is ingested and indexed in an Elasticsearch database. Machine Learning Analysis: ML models analyze PTW documents to find missing details, inconsistencies, or risk control gaps. Large Language Model (LLM) Integration: An LLM, accessed via API, interprets contextual prompts and provides real-time semantic suggestions. Intelligent Recommendations: The system typically delivers recommendations to the dashboard within minutes, subject to workload and data availability. Historical Data Reference: A second AI layer references historical permits and incident databases to offer additional suggestions based on past scenarios. Secure Deployment: The platform is deployed in a major cloud environment and is designed to support scalability and applicable compliance requirements.. This multi-stage AI review process transforms traditional PTW evaluation into a proactive, data-informed safety assurance method for offshore operations. User feedback and observations on the AI permit system highlight both positive and negative aspects. Positive observations include successful logins, helpful dashboard features, valuable AI-driven recommendations for JHA, and efficient permit management during Turnaround (TA). Negative observations include redundant task descriptions, data input issues, AI limitations in identifying specific hazards, unrelated learning summaries, and search functionality limitations. Users also expressed the need for a wider database source for extracting RA steps and linking related learning to other databases for comprehensive analysis. The feedback provides valuable insights into the strengths and weaknesses of the current permit system. Users reported productivity benefits and more consistent risk management during high-activity periods. Addressing the identified issues and leveraging the positive aspects can significantly enhance the system's effectiveness and user satisfaction. This paper will show a scalable dual-layer AI framework that integrates machine learning with a large language model to enhance PTW risk assessments. AI PTW provides context-aware suggestions and references to past incidents and delivers timely, intelligent feedback and can be adapted across different assets and work environments. Practicing engineers will benefit from a practical, modular solution that improves permit quality and embeds continuous learning into safety workflows. Refer to Figure 3 Intelligence in Action - AI Architecture Supporting Smarter PTW Reviews