How should a manufacturer use AI on RFQs and quoting?

Enquiries arrive in a form that makes them expensive to handle: a drawing, a spreadsheet, an email describing a part, sometimes a photograph, frequently incomplete. Turning that into the fields an estimator needs — material, quantity, tolerances, finish, delivery expectation, whether it is a repeat — is the work that consumes the time before any judgement is applied. It is also an extraction task with a checkable output, which makes it the right first application.

Triage at receipt is the second and is worth more than faster pricing. Most shops quote work they should have declined, and the cost of that is not the quoting time but the capacity consumed if it is won. A triage step that compares the requirement against a stated capability envelope — sizes, materials, processes, tolerances, minimum quantities — and routes the ones outside it removes work rather than accelerating it. Stating that envelope explicitly is a prerequisite and most shops have never written it down.

Pricing is where this should stop, and the reason is structural rather than technical. A price is a function of what the job costs and what you want the work to be worth, and the second depends on the current backlog, the machine that is idle, the material you already hold, the customer's history and what else is competing for the same slot. None of that is in the enquiry, and a system quoting from part characteristics alone is pricing a part rather than a job.

Where cost estimation is genuinely useful is as a starting figure from historical jobs rather than as a price. This part resembles these previous parts, which took this long and cost this much. That is a retrieval task over your own records, it is checkable because the comparable jobs are named, and it puts an estimator in front of evidence rather than a blank sheet. The distinction from an automated price is that a person is still deciding.

Drawings and specifications carry a constraint that does not apply to most business documents. Customer technical data frequently arrives under confidentiality terms, and some of it is subject to export control obligations that restrict who may access it and where. Sending a drawing to an outside service is a disclosure, and whether it is permitted depends on the agreement with that customer and on the classification of the data, not on the tool. This is a question to answer before connecting anything to the RFQ inbox.

The measurement worth building alongside is the one nobody has: how many enquiries arrived, how many were quoted, how many were won, and how long each stage took. Shops that add extraction get this as a by-product, and it usually reveals that the constraint is not quoting speed at all but the proportion of enquiries that were never going to convert.