Can an AI agent answer a customer's technical question about a part?

The useful line runs between retrieval and judgement, and it is sharper than it sounds. Retrieval is telling a caller what the drawing says the wall thickness is, what the specification lists as the material, what the last inspection report recorded, what the standard requires for that finish. Judgement is telling them the part will hold at that pressure, that the alternative alloy is equivalent for their application, that the tolerance is achievable on the shop's machines, or that the finish will pass their customer's incoming inspection. The first is a lookup. The second is engineering advice, and it is advice the buyer will act on.

Reliance is the mechanism that makes this expensive. A buyer who is told a substitution is equivalent orders on that basis, and when it is not, the conversation is about what the shop said and not about what the shop's system said. Sales engineering has always carried this exposure and shops have always managed it by routing certain questions to people who can carry them. Putting an agent in the same seat does not change the exposure; it changes how many of those questions get answered before anybody notices they were asked.

The failure that actually occurs is subtler than a wrong answer and harder to catch. Asked something outside the material it holds, a retrieval system will answer from the nearest thing it has: a similar part number, a superseded revision, a different customer's specification for a similar component. The answer is specific, formatted like the correct ones, and wrong in a way nobody detects until the parts arrive. This is the reason the design question is not how accurate the system is but what it does when it has nothing.

So the construction that works is a narrow one. Give it the controlled documents, require every answer to cite the document and revision it came from, and have it decline — visibly, to the customer, with a route to a person — when the answer is not in the documents it holds. A decline that hands the caller to an engineer within the hour is a good customer experience and a shop that answers slowly and correctly beats one that answers instantly and occasionally wrongly, particularly in a business where a wrong answer arrives back as scrap.

Revision control is the load-bearing detail and it is where most implementations quietly fail. A part has revisions, customers hold different revisions, and the answer to a dimensional question is different depending on which the caller has. An agent that answers from the latest revision to a customer working from an older one is confidently wrong at scale, and the fix is not better retrieval — it is asking which revision the caller is looking at before answering at all.

There is a real prize on the other side of these constraints, and it is not deflecting calls. It is that a shop with its drawings, specifications, past quotes and inspection history in one retrievable place can answer routine questions in minutes rather than when the estimator gets back, and can tell a caller within the same conversation whether the shop has made something like this before. That is a genuine competitive difference in a business where quoting speed decides which shops get shortlisted, and it does not require the agent to make a single engineering judgement.