A solar fault is detected. A complete CMMS work order is generated. Thirty seconds. No human in the loop. This is how AI agents generate CMMS tickets for utility-scale solar plants, and why those tickets can be trusted.
I know how that sounds, so let me walk through exactly what happens, step by step, because the difference between this and the “automation” features you have seen before is architectural, not incremental.
Why CMMS automation has failed before
Most CMMS vendors offer alarm-to-ticket bridges. When a SCADA alarm fires, a work order skeleton appears with an alarm code, timestamp, and device identifier. That is better than nothing. (For how an analytics layer should sit on SCADA instead, see our guide to an AI layer over solar SCADA.)
But it creates a specific problem: it generates work orders from every alarm, including the false positives, the sensor noise, and the inverter communication blip that resolved itself four minutes later.
Sites using these bridges accumulate large numbers of false work orders. The CMMS fills with noise. Schedulers learn to ignore anything below a certain severity threshold. Real faults get buried. You move the signal-to-noise problem from the alarm queue to the work order queue.
The work orders that are created are also skeletal: an alarm code and a timestamp. No imagery, no fault classification, no GPS coordinates, no estimated repair time. The technician receiving that work order still calls the control room for context before going anywhere.
That is why manual CMMS entry persisted even after “automation” was available: the automated output was too incomplete to be trusted.
The ClearSpot pipeline
ClearSpot’s AI agents for solar O&M run a different pipeline: detect, rank, confirm with a drone, then write the ticket. The walkthrough below is an illustrative example of a single string fault; the figures in it describe that example, not a site average.
Second 0 to 4: anomaly detection
The agent detects a production anomaly on String C-14, 11 percent below expected output, with the rate of change steepening over 72 hours and no matching weather event in the SCADA met data. It tags the anomaly as a candidate for investigation.
Second 4 to 8: revenue impact calculation
The agent calculates affected capacity multiplied by current irradiance multiplied by current energy price. This anomaly is worth €42 per hour at current generation conditions. It ranks in the top five active issues on the site by revenue impact.
Second 8 to 18: autonomous drone dispatch
The agent sends a flight request to the autonomous drone assigned to the site. The drone is airborne within the next available window, immediately if during operating hours, queued if not. It flies directly to the GPS coordinates of String C-14, already known from the site’s digital twin map.
For more on how this works physically, see how autonomous drone solar inspections close the loop between an electrical anomaly and visual evidence.
Drone in flight: thermal and RGB capture
The drone captures thermal and RGB imagery. The thermal camera uses 640×512 resolution and 0.05°C sensitivity. The inspection mission photographs String C-14 and the eight adjacent modules. Flight time for a targeted single-string inspection is about four minutes from launch to landing.
The camera has to be radiometric, meaning it records a temperature value for every pixel. A non-radiometric camera shows a hot spot but not how hot it is, and the work order depends on that number. Capture follows IEC TS 62446-3, the IEC technical specification for outdoor infrared thermography of PV modules, which sets minimum irradiance, viewing angle and image resolution so that temperature differences between modules are comparable.
Post-flight: AI classification
The agent receives the imagery and classifies the issue as a confirmed thermal anomaly, with a delta T of +14°C on modules 3 and 4 of String C-14. The pattern is consistent with bypass diode failure on two modules. Confidence: 94 percent.
The temperature delta (ΔT) is measured against healthy neighbouring modules under the same irradiance, not against an absolute temperature. That is what lets the agent separate a real defect from a module that is simply warm on a sunny afternoon, and it is the same comparison a thermographer would make by hand.
Work order generation
About 30 seconds after imagery receipt, the CMMS work order is created automatically:
- Fault type: Bypass diode failure, confirmed, 94 percent confidence.
- Location: String C-14, modules 3 and 4, with GPS coordinates attached.
- Evidence: thermal image with highlighted anomaly, plus RGB confirmation image.
- Revenue impact: €42 per hour, or €1,008 per day at current generation.
- Recommended action: replace bypass diode on modules C-14-03 and C-14-04.
- Parts required: 2× bypass diode assembly, pulled from the site CMMS library.
- Estimated repair time: 2.5 hours.
- Priority: P1, revenue-ranked.
The technician receiving this work order needs no phone call, no context briefing, and no diagnostic visit. They arrive with the right parts and know exactly where to go and what they will find.
The integrations
ClearSpot’s inverter-agnostic monitoring architecture integrates with major CMMS platforms and can adapt the work order format to each system’s native structure. The thermal images are attached as standard attachments in the format your CMMS already accepts. The supported CMMS, SCADA and drone connections are listed on the integrations overview.
If your site uses a procurement system for parts ordering, ClearSpot’s multi-agent O&M system can trigger a parts request at the same time as the work order, so by the time the work order is reviewed, the required parts are already on order.
For reference on how best-in-class CMMS integration and O&M workflows are defined across the industry, the IEA PVPS Guidelines for Operation and Maintenance of Photovoltaic Power Plants provide a useful benchmark.
How CMMS tickets are prioritised
A ticket queue sorted by alarm time sends crews to whatever fired first. The agent ranks every confirmed fault on three axes before the ticket reaches a scheduler:
- Revenue at risk: affected capacity, current irradiance and the energy price or PPA rate, so a fault on a large string at midday outranks a small one at dusk.
- Safety and escalation risk: connection overheating, arcing signatures or damaged cabling are raised above pure production losses whatever their revenue impact.
- Warranty and contract exposure: faults covered by module or inverter warranties carry the evidence pack a claim needs, and faults that affect availability guarantees are flagged for the asset manager.
The ranking is recalculated as conditions change, so a P2 ticket can move up when irradiance rises or when a second fault appears on the same inverter.
How to estimate the savings on your site
The business case is easiest to build from your own CMMS history rather than from vendor averages. Pull twelve months of work orders and calculate:
- Administrative time: the number of work orders multiplied by the average time a coordinator spends creating and enriching each one, multiplied by a fully loaded hourly rate.
- False work orders: the share of tickets closed as “no fault found”, “cleared on its own” or duplicate. Each one cost triage time and sometimes a truck roll.
- Diagnostic visits: site visits whose only purpose was to find out what was wrong before a second visit fixed it.
- Lost generation while waiting: the time between first alarm and repair, multiplied by the affected capacity and the energy price.
Drone confirmation before ticket creation targets the second and third items directly; complete tickets target the first; faster, revenue-ranked dispatch targets the fourth. Across its deployments ClearSpot reports a 30–50% reduction in O&M cost and a 65% reduction in downtime, with 99% fault-detection accuracy. The approach is set out in more depth in our guide to agentic AI for solar asset management.
Rolling out AI-generated CMMS tickets
- Map your CMMS fields. Agree which fields the agent fills (fault type, asset ID, location, evidence, parts, priority) and how they map to your CMMS’s native work order structure.
- Start in draft mode. Tickets are created as drafts for a coordinator to approve, so the team can check classification and priority against its own judgment.
- Load the parts library. Recommended parts are only as good as the asset register and spares list the agent can read.
- Switch on auto-release by fault class. Once a fault class has a reliable track record, its tickets go straight to the scheduler; others stay in draft.
Closing thought
What does your current mean time from fault detection to a technician having a complete, actionable work order in hand look like, and what would it change if that number were 30 seconds?
Frequently Asked Questions
The agent detects a production anomaly, ranks it by revenue impact, dispatches a drone to confirm it with thermal and RGB imagery, classifies the fault and then writes a complete work order into the CMMS, typically about 30 seconds after the imagery arrives.
It does not create a ticket from every alarm. It waits for anomaly detection, revenue ranking and drone confirmation, and only then generates a complete work order.
They flood the CMMS with low-value tickets, increase administrative labour and teach schedulers to distrust the system, so real faults get buried.
Fault type and confidence, location with GPS coordinates, thermal and RGB evidence, revenue impact, recommended action, required parts and estimated repair time.
Yes. The work order is mapped to the native structure of your CMMS, and thermal images are attached in the format your system already accepts. Tickets can start as drafts for approval.