Two different things are sold as AI in manufacturing quoting, and they deserve different levels of trust. The first reads what buyers send (an RFQ e-mail, a drawing PDF) and turns it into structured data a quoting system can price. That is real, it is shipping, and it saves the most tedious hour of an estimator’s day. The second produces the price itself from a model that has never seen your costs. That one needs much more care.
What is real: reading the request
Most of the time in a manual quote goes into reading, not pricing. An estimator opens the e-mail, finds the quantity in the third paragraph, opens the drawing, reads the material from the title block, finds the tightest tolerance and the thread callouts, notes the finish, and types it all into whatever system prices the part. It is skilled work, but most of it is transcription.
That is where software has improved fastest.
- Drawing analysis. AMFG launched Sentinel AI in July 2025, which reads a 2D drawing PDF in about three seconds and extracts the title block, materials and tolerances (Metal AM). Its help pages describe extracting threads, surface roughness and secondary services too, with each item shown against its callout on the PDF so a person can check and adjust it, and they advise high-resolution vector PDFs with legible title blocks (AMFG help).
- RFQ e-mail intake. DigiFabster announced an AI Quote Agent that connects to a shop’s RFQ inbox, extracts the key details, uploads the CAD and drawings, suggests materials and machines from the shop’s capabilities and produces a quote “ready for review and approval” (3D Printing Industry).
Notice what both have in common. The AI extracts and suggests; a person reviews; the price comes from the shop’s own pricing. That is the right shape. Extraction is a task where an error is visible (the drawing says 6061, the form says 7075) and cheap to correct before anything is made.
What is real, at scale: learning from millions of prices
Marketplaces price with machine learning, and it works for them. Xometry says every job on its partners’ job board arrives already priced by its instant quoting engine, which it says draws on large amounts of market data and machine learning (Xometry), and in early 2026 it introduced a machining lead-time model trained on a dataset four times larger than its predecessor’s (VoxelMatters).
The condition is the volume. A marketplace sees a very large number of parts and the prices at which shops accepted them. It is learning a market price, which is what a marketplace needs: a number at which some shop will take the job.
A single shop is in a different position. It has hundreds or a few thousand past jobs, spread over processes, materials and quantities. And it does not need the market price; it needs its own cost, plus the margin it chooses. A model that learns “what parts like this usually sell for” answers a different question from “what will this part cost me to make”.
What is not real: a price with nothing behind it
A general language model asked “how much to 3D print this bracket in PA12” will give a confident, plausible number. It has not measured the part, does not know your machine, your powder price, your machine-hour rate or how full your next build is. The number has the shape of a price and none of the content.
The danger is not that it is always wrong. It is that you cannot tell when it is. A price from a cost model can be checked line by line: this much material at this price, this much machine time at this rate. A price from a model that cannot show its working can only be trusted or not.
| Task | Good fit for AI today | Why |
|---|---|---|
| Reading an RFQ e-mail into fields | Yes | Errors are visible and corrected before pricing |
| Reading a drawing’s title block and tolerances | Yes, with review | Saves transcription; each item can be checked against the PDF |
| Suggesting process and material from a description | Yes, as a suggestion | A person or a rule confirms |
| Recognising features in a 3D model | Yes, increasingly | Measurable against the geometry |
| Pricing from a marketplace’s history | For a marketplace | Needs a very large volume of priced parts |
| Pricing a shop’s job without its cost data | No | Nothing to check the number against |
| Deciding margin for you | No | Margin is a business decision, not a prediction |
What a shop should want instead
The useful division of labour is plain:
- Software reads the request. E-mail, drawing, files and notes become structured fields, each shown against its source for a person to confirm.
- A cost model prices it. Your machines, materials, rates and rules, applied to the measured geometry, with every line of the price visible to you.
- Your own jobs correct the model. Actual times from finished work refine the estimates. That is calibration, and it is ordinary statistics on your own data, not a black box.
- A person decides the exceptions. Unusual parts, large orders and new customers go to review.
Quotometry works at steps 2 to 4: it measures each part once, prices it from the shop’s own configuration with a full breakdown, calibrates against jobs the shop has run, and holds odd parts for manual review. It does not read RFQ e-mails or drawings today; that is the step bureau platforms have moved on first.
Questions to ask any vendor selling AI quoting
- What does the AI do: extract, suggest or price? Get a straight answer.
- If it prices, from what data? Yours, a marketplace’s, or a general model’s?
- Can you see why a price is what it is? Line by line, not as a confidence score.
- What happens when it is wrong? Is the extracted data shown for review, or passed straight through?
- Who decides the margin? It should be you.
AI that removes the typing from quoting is worth having now. AI that removes the cost model is not, at least not for a shop whose prices have to cover its own costs. If speed is what you are after, much of it comes from answering routine parts at once; the cost of a slow quote shows why that matters more than it seems.
Questions people ask
Can AI price a manufactured part?
AI can estimate a price from many past prices, which works for a marketplace that has priced a very large number of parts. It cannot know your machine-hour rate, your material prices or your margin unless you give them to it, and a general language model asked for a price will produce a plausible number with nothing behind it. For a shop, the dependable price comes from its own cost model.
Where does AI help a quoting team most today?
In intake: reading an RFQ e-mail, pulling out the quantities, materials and deadlines, attaching the files, and reading a 2D drawing's title block, tolerances and notes. That is slow, error-prone typing for a person and well suited to software that extracts and then shows its work for a person to check.
Quotometry prices every part from your own machines and rates, on your website, in seconds.