<!-- Source: https://www.nitronedge.com/use-cases/ai-agent-maintenance-work-order-planning/ -->
# AI agent for maintenance work-order planning

Turn equipment condition alerts into draft maintenance work orders with parts, permits and crew suggestions for planners to approve.

## Key facts
- **The job:** Turn equipment condition alerts into draft maintenance work orders with parts, permits and crew suggestions for planners to approve.
- **Who it is for:** Maintenance planning and reliability teams at upstream, midstream and downstream sites.
- **Typical timeline:** 10 to 12 weeks to a supervised pilot
- **Accelerator:** SUSE Edge &amp; Private AI Accelerator
- **Technology options:** SUSE, RealWear, Tributech, NVIDIA

Maintenance planners turn alarms and inspection findings into work orders, then chase parts, permits and crews. Much of their day goes on looking things up across separate systems. An AI agent can assemble that information and draft the work order, leaving the planner to check and approve it.

## The job, in one sentence

When a condition alert or inspection finding comes in, the agent looks up the asset's history and manuals, drafts a work order with likely tasks, parts and permits, checks stock and crew availability, and passes it to a planner for approval.

## How the agent works, step by step

1. **Receive.** It picks up condition alerts from the monitoring system and findings from inspections.
2. **Look up.** It retrieves the asset's maintenance history, manufacturer guidance and similar past work orders.
3. **Draft.** It proposes the job steps, estimated hours, required parts and the permits that apply, such as hot work or confined space.
4. **Check.** It checks parts stock and lead times in the ERP, and crew and contractor availability.
5. **Prioritise.** It suggests a priority using asset criticality and the site's risk matrix.
6. **Hand over.** The planner reviews, edits and approves. Only then is the work order released.
7. **Support the field.** Approved work orders can be shown hands-free on RealWear headsets, with the relevant manual pages.

## Systems the agent touches

| System | What the agent does | Access |
|---|---|---|
| Condition monitoring and historian | Reads alerts and sensor trends | Read only |
| Enterprise asset management (for example SAP PM or Maximo) | Reads history; drafts work orders | Draft only, planner releases |
| Document library | Reads manuals and procedures | Read only |
| ERP inventory | Checks stock and lead times | Read only |
| Permit-to-work system | Identifies permits needed | Read only |
| Field devices (RealWear) | Shows approved instructions | Display only |

## Controls we build in

- The agent never releases a work order or issues a permit.
- Safety-critical steps come from approved procedures, not generated text.
- Sensor data integrity checked at source where Tributech is used.
- Edge or on-premises deployment where sites have limited connectivity or strict OT rules.
- Separation between OT networks and the agent, following IEC 62443 practice.

## Typical timeline

| Phase | Weeks | What happens |
|---|---|---|
| Discovery | 1 to 3 | Choose one unit or asset class; map alerts, history and planning steps |
| Build | 4 to 7 | Connect read-only data, index manuals, configure drafting |
| Shadow run | 8 to 9 | Planners compare agent drafts with their own |
| Supervised pilot | 10 to 12 | Planners approve agent drafts for the chosen unit |

Timelines are typical for one unit or asset class and depend on OT security approvals.

## What we measure

Baselines agreed with the reliability lead: planner hours per work order, time from alert to released work order, work orders delayed by missing parts or permits, rework after field feedback, and planned versus reactive maintenance ratio.

## Where it fits

Sites with limited connectivity often run this on an edge platform, which is what our SUSE Edge and Private AI accelerator provides. Hands-free field support links to our RealWear work.

## FAQ

### Does the agent act on its own?
No. In the pilot the agent gathers information, drafts and proposes. People approve every action that changes a record, moves money or affects a customer.

### How long does a pilot take?
Typically 10 to 12 weeks to a supervised pilot, depending on data access, integration and internal approvals.

### Where does the agent run?
On in-country cloud regions or on premises, chosen to meet your data protection and sector rules.
