Key takeaways
- Start with planning evidence, not autonomous control or safety decisions.
- Use stable equipment, document and work-order identifiers to join the data.
- Treat risk scores as review priorities, not facts about future completion.
- Keep management of change, permit, isolation and startup authority outside the model.
Where AI can help a turnaround
Turnarounds combine maintenance scope, inspection findings, materials, contractors, permits, isolations, schedules and changing field conditions. The useful role for AI is to reduce the effort required to assemble and compare this evidence.
High-value starting points include duplicate-scope detection, work-pack completeness checks, retrieval of prior job history, extraction of prerequisites, comparison of vendor recommendations and daily summaries of emerging constraints. These tasks support planners and reviewers without transferring plant authority to a model.
Build the equipment and document spine
AI cannot reliably connect a work order, drawing, inspection finding and spare part if each system uses different names for the same asset. Create a mapping around stable equipment tags, functional locations, document numbers, work-order IDs and schedule activity IDs.
For every source, record ownership, update frequency and authoritative status. A planning spreadsheet may be current but unofficial. A document system may be authoritative but updated later. The assistant should expose that distinction.
Use AI for work-pack quality
A work-pack assistant can extract the job objective, equipment, task sequence, labour assumptions, materials, tools, drawings, permits, isolations, inspection points and acceptance evidence. It can compare the draft against a plant-approved checklist and flag missing or conflicting items.
It should never invent a missing prerequisite. A useful result says what is absent, cites the relevant source and routes the pack to the responsible planner, engineer or safety role.
Risk signals need careful language
Schedule and completion-risk models can help prioritize review when they use comparable historical data. Signals may include late materials, incomplete engineering, unresolved scope, repeated rework, constrained resources and dependencies with little float.
The output is a ranking or signal, not a prediction that an activity will fail. Teams should test whether the signal identifies useful review candidates earlier than the existing process and whether it creates bias against unusual but well-controlled work.
Keep OT and safety boundaries explicit
NIST OT guidance emphasizes the performance, reliability and safety requirements of systems that affect the physical environment. A turnaround assistant should normally sit in an enterprise or controlled data zone, consume approved copies or interfaces and remain read-only during the first pilot.
Do not connect a general-purpose language agent directly to a distributed control system, safety instrumented system or programmable controller. Permit approval, energy isolation, management of change, process safety review and startup authorization remain under established plant procedures and qualified roles.
Pilot one bounded workflow
- Select one unit, discipline or repeat job class.
- Define the authoritative sources and map the identifiers.
- Establish baseline planning time, rework and missing prerequisites.
- Create a representative test set containing clean, incomplete, conflicting and obsolete records.
- Run the assistant read-only and require source citations for every flag.
- Review false alarms, missed gaps and user overrides with planning and safety owners.
Measures that matter
Useful measures include time to approved work pack, number of late missing prerequisites, duplicate or contradictory scope found before execution, planner review time and percentage of flags accepted with evidence. Keep safety, quality and startup measures as protected constraints, not as trade-offs for speed.
The goal is a more reviewable turnaround plan. AI should make the evidence easier to see while the accountable team decides what work is safe, ready and authorized.
