AI for Refinery Turnarounds: Turn Maintenance Records into Reviewable Work Packs

Turnarounds are evidence and coordination problems

A refinery turnaround brings together inspection findings, maintenance history, equipment identity, isolation requirements, materials, access, specialist skills, permits, procedures and the planned sequence of work. The challenge is not only to create a schedule. It is to decide whether each proposed work item is understood, supported by evidence and ready for review.

AI can help organise that evidence, identify missing links and draft a work-pack outline for planners and discipline leads. It should not be treated as an autonomous maintenance planner, a process-safety authority or a substitute for the approved work-management and field-control process.

The useful starting question is: what must be true before this work item is ready for an accountable person to review?

Define a work-pack readiness decision

Make the first workflow narrow and explicit. For each candidate work item, ask:

• Is the asset identity unambiguous across the relevant systems?
• What triggered the work: inspection finding, condition signal, overdue task, modification, incident learning or planned scope?
• Which source records support the need and current condition?
• What isolation, access, tools, materials, skills and permits may be required?
• Which approved procedure, drawing, specification or acceptance criterion applies?
• What is still missing, stale, conflicting or outside the workflow’s scope?
• Who reviews the draft and who approves the work for execution?

The output should be a reviewable work-pack draft with evidence links and open questions. It should not be a silent write into the official maintenance or permit system.

Connect the records before adding a model

The model is not the first integration problem. The first problem is identity and meaning.

Map how the same asset, tag, system, location, job, inspection, material and document are represented in source systems. Record the source, revision, timestamp, status and access boundary. If two records appear to describe the same equipment but use different identifiers, route the conflict for resolution.

Useful inputs may include:

• equipment register and hierarchy;
• maintenance history and failure descriptions;
• inspection reports, thickness readings or integrity findings;
• operating context and recent deviations;
• drawings, procedures and isolation information;
• material, tool and specialist-skill requirements;
• prior work-pack templates and completion notes;
• planned windows, access constraints and handover conditions.

Do not imply that all available data is approved evidence. A work order comment, a draft drawing and an approved procedure may have different authority. Make that difference visible to the reviewer.

Use AI to structure the work pack

A bounded assistant can perform practical tasks without owning the decision:

• extract candidate tasks from inspection and maintenance records;
• group related evidence by asset and system;
• identify missing fields in a work-pack template;
• compare a draft against an approved procedure or prior revision;
• surface conflicting dates, identifiers, quantities or statuses;
• draft a task summary with source links and unresolved questions;
• suggest a review queue based on explicit, owner-approved rules;
• prepare a handover checklist for planning, execution and close-out.

Each output should show the source record and the reason it was included. If the system infers a relationship, label it as an inference. If it cannot resolve a conflict, state the conflict. A concise evidence gap is safer than a plausible paragraph that hides uncertainty.

Separate priority from readiness

A high-consequence or time-sensitive item may deserve early attention, but that does not mean its work pack is ready for release. Keep at least two dimensions separate:

• Operational priority: - why the item may matter to the planned window, asset integrity or process objective.
• Work-pack readiness: - whether the evidence, scope, constraints, review decisions and execution information are sufficient for review.

A priority queue can contain items that still need inspection, isolation planning, materials confirmation or engineering clarification. The AI workflow should make that distinction visible instead of turning incomplete information into a false ranking.

If a team uses a scoring rule, document its inputs, thresholds, owner, exclusions, review frequency and override path. Do not present a score as a process-safety conclusion or a maintenance decision.

Handle process-safety boundaries carefully

Process safety is concerned with the integrity of operating systems and processes that handle hazardous substances. A work-pack assistant may help organise records related to those systems, but it cannot establish that a barrier is healthy, that an isolation is adequate or that a change is safe without the approved engineering and operating process.

Keep the AI workflow outside direct control of:

• process control commands;
• safety instrumented functions and emergency shutdown logic;
• isolation or de-isolation authority;
• permit authorisation and field release;
• hazard-study decisions;
• management-of-change authorisation;
• acceptance of a safety-critical repair.

The system may prepare evidence for the people responsible for these decisions. It should not create the impression that a generated work pack is an authorised action.

Make missing evidence a first-class outcome

Use explicit states rather than forcing every work item into ready or not ready:

1. Evidence linked: - required sources are identified and the work-pack draft is ready for review.
2. Evidence gap: - a required record, measurement, procedure, identity or constraint is missing.
3. Conflict: - sources disagree and an owner must resolve the discrepancy.
4. Out of scope: - the work item requires a decision or control outside the workflow.
5. Unavailable: - a source, interface or service is not accessible.
6. Reviewed: - an accountable person has accepted, revised or rejected the draft with a recorded reason.

Define the response for each state. An evidence gap may create an inspection task. A conflict may go to the asset owner. An out-of-scope item may follow the existing engineering or process-safety route. An unavailable source may require the manual process.

Evaluate against real turnaround variation

A demonstration using clean records will not reveal whether the workflow is useful during a complex planning window. Build an evaluation set with controlled cases:

• complete work histories and clear asset identity;
• duplicate or conflicting identifiers;
• stale procedures and superseded drawings;
• incomplete inspection findings;
• different equipment classes and maintenance disciplines;
• material, access or specialist-skill constraints;
• tasks that share an isolation or work area;
• urgent work mixed with planned scope;
• records that contain ambiguous language or abbreviations;
• requests that should be rejected or escalated.

Measure evidence linkage, missing-field detection, conflict routing, source-revision accuracy, reviewer corrections, safe abstention, tool-call boundaries and time to produce a reviewable draft. Do not reduce the assessment to a generic model score.

Keep development examples separate from acceptance cases. After a model, retrieval source, prompt, template, interface or integration changes, rerun the relevant evaluation. A change that improves extraction can still alter source selection or escalation behaviour.

Keep the planner and field teams in control

The planner decides whether the draft is useful for the planning process. The maintenance or engineering owner decides whether the scope and evidence are technically acceptable. The process-safety and operations roles retain their existing authority. The field team follows the approved work, permit and isolation process.

Design the interface around review:

• show the source excerpt and revision;
• identify inferred fields;
• separate facts from suggested wording;
• show missing and conflicting evidence;
• make the reviewer’s correction easy to record;
• prevent a draft from becoming an official action without the required review decision;
• preserve the relationship between the draft, decision and final record.

Reviewers should not be asked to approve an opaque summary when the decision depends on a traceable evidence package.

Track changes and operational drift

Turnaround conditions change. New inspections arrive, access shifts, equipment status changes, material availability moves, procedures are revised and the planned window may be re-sequenced. The workflow should expose what changed and why a previously reviewed draft may no longer be current.

Monitor:

• source freshness and revision state;
• identity-matching exceptions;
• evidence-gap and conflict rates;
• reviewer overrides and reasons;
• changes in work-pack templates or rules;
• interface failures and delayed source updates;
• draft-to-approved conversion states;
• tasks returned from the field for clarification;
• configuration and retrieval changes.

Do not retrain or change a threshold simply to make exceptions disappear. An increase in conflicts may reveal a source-system or planning problem that should be investigated.

Plan delivery from India to downstream operations

An India-based delivery team may support refinery planning workflows for international clients, but the engagement must define the data, access and deployment boundary. Document where maintenance records are processed, who may access them, how remote support is authorised and logged, which systems remain client-controlled and how the client can continue work if support access is unavailable.

The delivery arrangement should distinguish source records, client-specific mappings, evaluation cases, configuration, deployment files, documentation, training and third-party dependencies. Cross-border delivery, data transfer, cybersecurity, safety, tax, employment, licensing, contracts and sector obligations require client-specific review by qualified counsel and responsible professionals in the relevant jurisdictions.

This article is an implementation guide, not process-safety, legal, tax, cybersecurity or regulatory advice. A general workflow pattern cannot establish that a particular refinery, unit, work pack or jurisdiction is safe or compliant.

A practical implementation sequence

Use a staged path:

1. Select: - choose one turnaround work-pack family and one accountable planning owner.
2. Map: - document asset identity, source systems, evidence authority, constraints and decision boundaries.
3. Prepare: - define the work-pack template, missing-data states, manual fallback and evaluation cases.
4. Prototype: - generate drafts and evidence links without changing the official maintenance record.
5. Review: - compare drafts with planners, maintenance, engineering, operations and process-safety roles.
6. Shadow: - run alongside the current process and record corrections, conflicts and abstentions.
7. Release narrowly: - support one approved workflow with access controls, audit trail and escalation.
8. Transfer: - provide source maps, configuration, tests, runbooks, training and support procedures.
9. Expand carefully: - reassess each new unit, discipline, source, decision or authority boundary.

The workflow is ready to progress when a planner can see why a work item was included, an owner can resolve the open questions, the official process remains authoritative and the team can operate manually when the AI layer is unavailable.

The practical takeaway

AI can make downstream turnaround preparation more structured by connecting maintenance evidence, exposing gaps and drafting reviewable work packs. It should not hide uncertainty, bypass process-safety authority or turn an incomplete record into an automatic schedule or release.

Start with one work-pack family. Make asset identity and source status visible. Separate priority from readiness. Route conflicts and missing evidence to named owners. Keep decision rights and field controls outside the AI layer. Evaluate the complete workflow, then expand only when the evidence supports the next boundary.