AI for manufacturing quoting, what works and what does not yet

Anyone who has sat through 4 vendor calls this year has heard the same demonstration. A clean model goes in, a price comes out, everybody nods. The useful conversation about AI for manufacturing quoting starts after that, with the parts nobody demonstrates.

What follows splits the process into stages and says plainly which are solved, which are partly solved and which are still a person. The middle is genuinely good. Both ends are not, and the claims that fail are always about the ends.

What manufacturing quoting has actually solved

Reading a messy package is solved well enough to rely on, with a human checking the flagged fields. Six files, a scan, a workbook and a quality spec can be sorted, read and turned into a structured takeoff faster than an estimator can open them, provided the system carries confidence and shows its working.

The mechanics of that are covered in what it takes to read a drawing. The short version is that it works when regions are detected and cropped first, and it fails when a whole sheet gets handed to a model.

Matching new work against past jobs is the second solved thing and the most undersold. Given a part and a record of closed jobs, finding the 5 that resemble it, ranking them by how close the match is and surfacing what each actually cost is a retrieval problem, and retrieval is reliable.

Drafting supplier requests is the third. A part with 2 outside operations needs 2 requests written, sent, chased and reconciled. That is clerical work with a clear format, and automating it removes days of elapsed time rather than minutes of effort.

Retrieval is reliable. Judgement is not.

What is partly solved and needs watching

Building the routing from a part is partly solved and the failure is quiet. A system can propose operations from geometry and from similar jobs, and it will be right most of the time. When it is wrong it is usually missing an operation rather than adding a wrong one.

A missing deburr or a missing wash does not look like an error. It looks like a slightly cheaper quote, which is exactly the direction that costs money, and it is why the proposed routing needs a human read rather than an approval click.

Cycle time estimation is the second. From a similar job it is good. On a part with no close relative it becomes a model estimate, which is the weakest form of the number, and the system should say which of the 2 it just did.

Outside processing prices are third. A system can suggest a figure from what that supplier charged last time on similar work, which is useful for a provisional number and is not a quote. Whether provisional lines are clearly marked as provisional is a real product question.

What is still a person, and will be for a while

Deciding whether to take the work is judgement about your business and it is not a data problem. Capacity, strategic value of a customer, whether you want to be in this part family, whether this program will fund the machine you need next year. None of that is in the package.

Negotiating is the second. What a customer will accept, when to hold a price, what the relationship is worth against this order. Anyone selling software for this is selling something else.

The third is everything that was never written down. Your best estimator knows this customer’s cleanliness spec doubles the wash, that the plater comes in low on this finish, that the last part in this family scrapped 4 percent in month one. If that knowledge lives only in a person, no system has access to it.

The fix for the third is the one worth acting on. Knowledge that stays in a head cannot be used by software or by the person who replaces them, and writing it into the record is valuable regardless of what you buy.

Undocumented knowledge is unusable knowledge.

There is a fourth that gets promised and should not be. Predicting whether you will win a specific quote at a specific price is a claim about a buyer’s internal decision, made with 3 data points and no visibility of the competing bids. Treat any confident number here as marketing.

Where the time actually goes, against where the tools aim

Most of the elapsed week is waiting rather than working, which is the observation that reorders the whole priority list. A quote taking 9 days usually contains 4 hours of effort and 8 days of queue, handoff and waiting on an outside supplier.

Automating the 4 hours is worth having. Removing the queue is worth more, and it comes from different capabilities, mainly from doing the work in one sitting rather than across 3 desks and from getting supplier requests out on day 1 instead of day 4.

That reframes what to buy. A tool that makes estimating 30 percent faster and leaves the handoffs alone changes 4 hours into 2.8 and leaves the 8 days. A tool that collapses the sequence changes the date the quote arrives, which is the thing the customer actually experiences.

The distinction that matters most

A model guessing what a part should cost is a different and weaker thing than a system pricing from what your jobs did cost. Both can be described as AI and the second is the one that survives contact with a customer.

A generative estimate is a plausible number produced from patterns across work that is not yours. It sounds confident, it cannot be audited, and when procurement asks why the hone line is what it is, there is no answer behind it except that the model said so.

A retrieval system answers with a job number, a date and a closeness. That is checkable, it improves as your record grows, and it is defensible in the room where the price gets argued. The longer case for that sets out why the distinction matters commercially rather than just technically.

The test is simple. Ask any tool where a specific line came from. If the answer names a source you can open, it is retrieval. If the answer is a description of a model, you are being asked to trust a guess, and a should-cost engine at least admits that is what it is doing.

What to ask a vendor who says AI

Ask which stage they mean, because the word covers all 7 and the honest answer is 3 of them. A vendor claiming end to end automation is either describing a demo path or has not met a real inbox.

Ask what happens when the system is unsure. A tool with no concept of uncertainty will be wrong silently, and silent errors in quoting are the expensive kind because they win work at the wrong price.

Ask what improves as you use it. If the answer is that the vendor’s model gets better for all customers, your data is training something your competitors also use. If the answer is that your record gets deeper and your matches get closer, the benefit stays with you.

Ask what it will not do.

Then ask what it does not do. A vendor who can name 3 things their product will not solve is describing a product. A vendor who cannot is describing a category, and the gap between those 2 is where the 18 month disappointment lives.