[Draft article - review and edit before publishing.]
We build AI inspection software, so you might expect us to claim it can do everything. It can't - and being clear about the line is exactly what makes AI trustworthy in a safety-critical discipline. Here's an honest account of what machine intelligence genuinely does well in Ex inspection work, and what should always stay with the human holding the camera.
What AI does brilliantly
- Reading and transcribing. Nameplates, serial numbers, Ex markings, temperature classes - extracted from a photo in seconds, without transposition errors, at the four-hundredth item of the day as accurately as the first.
- Connecting records. Linking the equipment in front of you to its Ex certificate, its inspection history and the applicable checklist items - the retrieval work that used to eat office afternoons.
- Never forgetting the checklist. The system knows what needs inspecting, in what order, to what standard - so nothing is skipped because it was raining, late, or the last unit of a long shift.
- Flagging the obvious-in-hindsight. Corrosion, missing bolts, damaged glands - pattern recognition surfaces candidates for the inspector's attention in real time, before they become incidents.
What stays with the inspector
- Judgement. Is that corrosion cosmetic or structural? Does this installation deviation actually create an ignition risk? These calls carry professional accountability, and accountability doesn't delegate to software.
- Context. The inspector knows this pump vibrates, that this area floods in monsoon season, that maintenance touched this panel last month. AI augments that knowledge; it doesn't possess it.
- The final signature. Every AI-generated entry in Mantl-Core is reviewed and validated by the inspector. The dossier is theirs - the AI just built the scaffolding.
The goal was never to remove the inspector. It was to remove everything that stops the inspector inspecting.
Why this division of labour matters
Inspections fail in two ways: things not looked at, and things not recorded properly. Machines are superb at guaranteeing coverage and record quality; humans are superb at evaluating what they see. Software that respects that split makes inspections faster and more reliable. Software that pretends to replace judgement makes them dangerous.
The takeaway
Ask any AI inspection vendor where their line sits. Ours is explicit: AI handles capture, transcription, retrieval, sequencing and compilation; your inspector handles judgement, context and sign-off. If you'd like to see that philosophy running on real equipment, book a live demo.