Maintenance work order planning AI agent: a blueprint for process plant maintenance
How a governed AI agent workforce takes the gathering out of maintenance work order planning: it assembles work packages from the CMMS, stores and permit-to-work systems, proposes schedule slots, and leaves every release, permit and schedule decision with people.
- Problem
- Planners spend the week gathering scope, parts, permits and crew data, so packages reach the field incomplete.
- Agent design
- Preparation, sequencing, closeout and backlog pattern agents that prepare and propose, under the planning lead.
- Systems
- CMMS or ERP maintenance module, stores, permit-to-work, production plan and scheduling tool.
- Hard limits
- No agent releases an order, raises or overrides a permit, sets isolations or defers a safety-critical task.
- Value levers
- Planner hours released, fewer jobs stopped by missing parts or scope, less idle contractor time.
- First release
- Shadow mode on one unit for preventive and standard corrective work, with planners reviewing every package.
Where maintenance work-order planning breaks today
In a refinery, chemical plant or terminal, a planner turns an approved notification into a job a crew can execute: reading the request, walking down the equipment, finding a task list or a similar past job, building operations and craft hours, checking every part against stock and open purchase orders, listing the permits and isolations, and confirming the production window with operations. Most of the week goes on gathering from the CMMS, stores, the permit office and the shift log, and little on the judgment only a planner can make.
When the gathering is rushed, the gaps show up at the job site: the gasket kit is in another warehouse, the scaffold was never requested, or operations has not released the line. The job breaks the schedule, wrench time drops, and the order returns to an aging backlog. Thin closeout notes mean the next planner starts from scratch.
| Work order type | What makes it hard to plan | What the agents do | Agents may finish alone? |
|---|---|---|---|
| Preventive maintenance (time or meter based) | Task list exists, but parts, permits and crew are rarely pre-checked | Copies the task list, checks components against stock and lead time, lists permit types, proposes a slot before the due date | Package yes, planner approves |
| Repeat corrective job with a standard job plan | The last good package sits in an old order | Reuses operations and materials from the closest past orders on the same equipment class and highlights differences | Package yes, planner approves |
| Non-routine corrective from a notification | Vague request; scope depends on the walk-down | Drafts scope from similar jobs and lists the questions the walk-down must answer | Planner completes scope |
| Work needing a production window | Equipment comes out of service only in an agreed slot | Matches the job to planned outage and rate-change windows and offers options | Operations picks window |
| High-hazard work (hot work, confined space, line break) | Permits and isolations drive duration and sequence | Lists permit types and isolation history as reference; the area authority builds the isolation plan | Package yes, permit never |
| Safety-critical element task or deferral request | Integrity commitment with a fixed due date | Builds the package; any deferral request goes to the technical authority with due date and risk record | Deferral never |
| Emergency or break-in work | Failure in service; work starts before planning | Stays out of the live response; backfills history and parts used at closeout | No |
| Work linked to a management of change | Scope alters design, materials or operating limits | Flags the MOC link and stops; engineering owns the scope | Never |
The maintenance work order planning agent workforce: four roles that prepare and propose
Agents prepare and propose; planners approve. Each agent runs under its own identity with CMMS authorizations that stop at draft, so none can release work, complete an order or touch a permit, and every proposal records its inputs. The fourth agent, added to the three on the live page, reports where packages and schedules break.
| Agent | Job | Reads | Can write | Never does |
|---|---|---|---|---|
| Preparation agent | Assembles each work package and flags gaps to their owner | Notification, equipment history, task lists, bills of material, stock and open POs, permit history | Draft operations, components and documents on unreleased orders; gap notices to owners | Release an order, raise a purchase requisition, or set isolation points |
| Sequencing agent | Proposes the weekly schedule from the ready backlog | Ready backlog, priorities, crew calendars, production plan, permit office capacity | Proposed schedule and change log in the scheduling tool | Commit a schedule, touch the frozen week, or slot an incomplete package |
| Closeout agent | Chases completion data and drafts job history after the work is done | Time confirmations, goods issues, technician notes | Draft history and failure code suggestions; thin-record flags | Close an order or overwrite codes a technician entered |
| Backlog pattern agent | Reports backlog health and schedule breakers by unit, craft and cause | Audit log, order history, schedule compliance data | Weekly backlog and schedule-breaker report | Change a priority or act on any open order |
Architecture: planning agents beside the CMMS, outside release authority
The agents read the CMMS, stores and permit systems through governed interfaces and write only drafts: operations and components on unreleased orders, and proposed slots in the scheduling tool. The control plane holds the site's planning rules, the planner approval step, the log of inputs and schedule changes, and a pause switch that leaves orders where they stand and returns planners to today's manual process.
How one maintenance work order moves through the design
Two questions route every order. Is it break-in work, a safety-critical deferral or an MOC-linked job? Those leave the agents' path at once. If not, is the package complete, with parts available in time and permit types identified? Only complete packages reach the planner as ready; gaps go to their owner and the agent re-checks.
CMMS, stores and permit-to-work integration for planning agents
On every platform the agents read through published interfaces, write drafts the planner can see and change, and leave status changes to people. Where the CMMS is SAP, each agent interaction is mapped against SAP's API policy and indirect access licensing during design, before any build starts.
| Platform | Where planning work lives | How agents read | How agents write | Design note |
|---|---|---|---|---|
| SAP S/4HANA Plant Maintenance | Notifications, maintenance orders, task lists and maintenance plans | Published OData APIs for maintenance orders, notifications and related master data | Operations, components and long text on orders still in created status | No release or completion authorization for the agent identity |
| IBM Maximo Application Suite | Work orders, job plans, planned labor and materials | Maximo REST interfaces for work orders, assets and inventory | Planned labor, materials and attachments while the order awaits approval | Status security keeps the agent identity from moving any order to approved |
| Permit-to-work system | Permit types, isolation certificates and history by equipment | Vendor API or a read-only export, depending on the product | None; agents never create, change or close a permit | Permit office capacity per shift is a sequencing input |
| Scheduling tool and production plan | Weekly schedule, crew calendars, outage and rate-change windows | Scheduling tool API, production plan extract | Proposed schedule as a draft scenario for the scheduler | Frozen week locked to agents; turnarounds in Primavera P6 out of scope at first |
Agents can run on ERP vendor tooling, a hyperscaler AI platform or an existing automation stack, chosen by where identity and logging are strongest.
Human oversight in maintenance planning and scheduling
Oversight follows the hazard: the nearer a decision sits to live equipment or safety-critical status, the further it sits from the agents. Each site documents and tests what each agent may do alone, what needs sign-off, where logs live, and how the workforce is paused.
| Decision | Who decides | Agent's role | Control |
|---|---|---|---|
| Gap notices to stores, engineering or operations | Agent | Sends and re-checks the notice | Every notice logged against the order |
| Work package approval | Planner | Assembles the package with its inputs listed | Planner signs every package before it joins the ready backlog |
| Weekly schedule | Scheduler with operations | Proposes slots and shows the trade-offs | Agent cannot commit; every change logged |
| Order release and technical completion | Planner or supervisor | None | Agent identities hold no release or completion authorization |
| Permits and isolations | Area authority | Lists permit types and isolation history as reference | No write access to the permit-to-work system |
| Safety-critical deferral and priority changes | Technical authority | Attaches due dates and risk records | Deferral follows the site's existing risk-assessed process |
| Planning and sequencing rule changes | Maintenance manager | Proposes changes from the backlog report | Versioned rules; each proposal records the version it used |
Illustrative economics of work-order planning agents
The model uses example inputs for a large process site with 24,000 planned work orders a year; replace them with your own. It keeps two views apart: planner and in-house craft hours become cash only if redeployed or used to replace overtime and contract planners, while avoided contractor idle time and parts expediting are cash either way.
Fixed assumptions: planner hours with agents include signing every package; value from earlier repairs and avoided failures is left out. Example inputs only. They describe no client and are not a forecast.
Most planning value is people's time, so the cash view is thin. Decide before the build where released planner hours go: more walk-downs and better scope on non-routine work, absorbing backlog without contract planners, or planning further ahead so parts arrive before the window. If none applies, the case rests on contractor idle time and expediting, and the scope should start smaller.
Rollout path for planning agents, one unit at a time
The path follows the live page: confirm CMMS data readiness, encode rules with the site's planners, release at one unit with full review, then expand unit by unit. The maintenance manager signs each gate.
- Weeks 0 to 5
Baseline
Measure task list and bill of material coverage, history quality and backlog age, then write the planning rules with the unit's planners and scheduler.Gate: maintenance manager signs planning rules v1 - Weeks 5 to 11
Shadow mode
Agents build packages and proposals beside planners, who plan as normal. Each agent package is compared with the planner's version.Gate: 90% of PM and standard packages accepted without material rework - Weeks 11 to 18
Controlled release
Agent packages enter the planner queue and proposals go to the scheduler, with every package reviewed and the pause switch tested.Gate: zero release or permit control breaches; schedule compliance at or above baseline - Week 18 onward
Scale
Add non-routine corrective work, then further units with their own rules and baselines.Gate: per unit, per order type
How maintenance planning performance is measured
The live page's measures plus two control measures, each baselined from the unit's own CMMS data. No achieved values are claimed.
| Metric | Definition | Why it matters |
|---|---|---|
| Schedule compliance | Share of scheduled hours or orders completed in the scheduled week | Tests whether packages are good enough to execute |
| Planner hours per work order | Planning hours, normalized for order mix | Drives the capacity case |
| Jobs delayed by incomplete packages | Share of scheduled orders that slip for missing parts, scope, permits or scaffold | Separates planning breakers from operational ones |
| Package rework rate | Share of agent packages a planner sends back or changes materially | Control metric: decides whether an order type or unit scales |
| Backlog aging | Ready and total backlog in crew-weeks, and orders older than 90 days | Shows whether the backlog is shrinking or only moving |
| Wrench-time proxy | Hours confirmed on orders as a share of craft hours available | A stand-in for wrench time without a field study |
| Closeout record quality | Share of completed orders with failure codes and reusable history | Sets how good the next package can be |
Questions maintenance leaders ask about planning agents
Can the agent release work orders or issue permits?
No. The agents draft packages and propose slots. Release, completion, permits and isolations stay with planners, supervisors and the area authority, and agent identities hold no authorization for them.
Does our CMMS data need to be clean before we start?
It needs to be measured. The baseline checks task list coverage, bill of material accuracy and history quality for one unit. Agents start with well-covered order types, and the gaps become a data work list.
Does this cover turnaround planning?
Not in the first release. Turnarounds run on their own scope control and scheduling, often in Primavera P6. Routine and corrective work comes first; turnaround packaging is assessed once unit-level rules are proven.
What about SAP licensing for agent access?
Each agent interaction with SAP is mapped against SAP's API policy and indirect access terms during design, and the cost enters the business case before build.
What happens to planners' jobs?
Planners keep every approval. Gathering moves to the agents, and the hours go to walk-downs, non-routine scope and planning further ahead, as the site decides before the build.
Working with WGA on maintenance work-order planning
The engagement starts with a readiness check on the unit's CMMS data and planning process, and a baseline of schedule compliance, planner hours, schedule breakers and backlog age. WGA writes the packaging and sequencing rules with the site's planners and scheduler, agrees targets with the maintenance manager before launch, and builds within the site's identity, permit and SAP licensing controls.
Senior WGA consultants do the work with fees fixed against the outcome and support each unit's release until results are documented. The planning lead and IT team inherit the rules, runbooks and pause procedure to extend the agents to the next unit without us.
Discuss maintenance work-order planning in your operation
Talk to our AI practice