A buyer's guide to AI assistants for maintenance: the five questions to ask

1. Ask it something, then ask where that came from

A useful answer in maintenance is one a technician can check before acting on it. So the first test is not whether the answer is right; it is whether the answer says which page of which manual, which earlier intervention on which unit, or which approved procedure it drew on — and whether tapping that source opens the page, not a search box.

Ask, too, whether every answer is cited the same way, or whether some come with sources and others with a confident paragraph. The honest tools separate the two visibly. An answer with no source should look different from an answer with one, because the technician should treat it differently.

2. Ask it something it cannot know

Pick a question whose answer is not in any document you have given it: a modification made last year, a part number that changed, a fault the plant has never had. Then watch what happens. The three possible behaviours are: it invents an answer with the same confidence as before; it says it does not know and stops; or it says the record holds nothing on this, and offers what a competent maintenance engineer would think, labelled as exactly that.

The first is disqualifying. The second is safe and not very useful at 03:00. The third is what you want — provided the label is unmissable. An assistant that answers from general trade knowledge and says so is more useful than one that only speaks when it has a citation, and far more useful than one that hides the difference.

Honesty about uncertainty is the contract. Citation is what it looks like when a source can carry the answer.

3. Ask where the knowledge comes from — and what you have to write first

This is the question that decides your first three months. Some assistants answer only from content that was created inside their own platform: instructions, guides, forms your team wrote and approved. Until that content exists, the assistant has nothing to say. Ask how many weeks of content creation stand between signing and the first useful answer.

The alternative is an assistant that starts from what you already have: the manufacturer’s manuals as they are, the procedures in whatever folder they ended up in, the closed work orders exported from the CMMS, and the threads in which incidents are actually solved from now on. Ask specifically whether the manufacturer’s manual is a source. In maintenance it is the one document that is always there, and it is surprising how many tools cannot read it.

4. Ask where an answer applies

A fix is not just a fact; it has a scope. It may be true for one unit that was modified, for every unit of a model, for a family of equipment from the same manufacturer, or for one configuration of one serial range. Ask the assistant a question about a machine you run several of, and see whether it knows the others exist — whether an intervention in Berlin is offered when you ask about the same machine in Munich, and whether it tells you it is the same model.

A tool that files knowledge under a plant, an area or a line will answer within that folder. A tool that attaches knowledge to the machine and to its model can carry a lesson to the next line, the next site and the unit that will be installed next year. That is the difference between search and memory, and it is visible in one question.

5. Ask who approved it

Ask to see a procedure the assistant cites, and then ask who published it, when, which version it is, and what replaced the previous one. In a regulated plant this is an audit question; in every plant it is a trust question. A technician follows a procedure because someone they know stood behind it, not because a model generated it fluently.

The corollary is to ask what the assistant is allowed to do on its own. Drafting is fine — a closure, a guide, a summary. Publishing, and anything that touches your CMMS or another system, should require a person, and the log should show who. Ask to see the log.

Two questions for IT, before the demo ends

What does it read, and what does it never do. The first answer should be a list you control: your own record, the documents you allowed, nothing from the open web, nothing from another customer. The second should include: it does not train on your data, it does not let a provider retain it, and it does not act on your systems without approval. If your company has standardised on one assistant already, ask whether this one can be a knowledge source for it rather than a second chat window — and whether that is available today or a direction.

Where Norlan is on this

Cited answers, the visible separation between an answer from the record and one from the trade, the manufacturer’s manual as a source and scoping by unit, line or model are in the product. Opening the manual on the cited page, translation, and serving as a knowledge source for an assistant you already run are product direction, labelled on the assistant page. Bring these five questions to the demo; we would rather answer them on your machine than here.