An AI maintenance assistant is an AI model that explains an alert, searches the manuals and the plant's history, and suggests the fix to the maintenance engineer. Maintenance software has long been good at raising alerts. The difficulty was that it raised too many, and the engineer had to find the cause alone, by reading old repair notes and thick manuals.
Imagine a maintenance engineer at a refinery in Panipat. An alert says: "Feed pump vibration high." The assistant reads the pump's last six months of readings. It finds three earlier repair records for the same pump. It searches the maker's manual, which lists misalignment and bearing wear as the usual causes of high vibration. Then it writes a short note: the pattern matches bearing wear, as in the repair done in March, and the likely fix is a new bearing. It also checks that the spare bearing is in the stores.
The engineer can also ask the assistant questions in ordinary words, for example "what did we do the last time this pump shook?" The answer comes with a link to the repair record or the page of the manual it used, so the engineer can check it instead of trusting it.
The assistant only drafts the note. The engineer reads it, agrees or disagrees, and approves the work order. This is the arrangement called human in the loop. A wrong guess costs nothing, because no repair starts until a person approves.
The assistant is a part of asset performance management (APM), and it uses the alerts that APM raises. It works well only when the manuals and repair notes are in a form that can be searched. In aircraft making, the equivalent manuals are called aircraft technical publications. Where a model goes further and acts on machines by itself, it becomes physical AI, which needs much stricter checks. Together, these tools support predictive maintenance.