Owners and estimators ask us the same questions about AI quoting, usually in the same order. This page answers them in that order, in plain words. Where we cite a number, we link the page we read it on. Where we give an opinion, we say so.
If you want the bigger picture first, read The Quotes You Never Sent, which explains why cheaper AI changes which RFQs are worth answering.
What is AI quoting?
AI quoting is software that does the first pass on a quote. It reads the RFQ email, the drawing, and any spec sheets, finds your past jobs on similar parts, and drafts a price using your own material costs, machine rates, outside-process costs, and margins. An estimator reviews the draft, adjusts it, and sends it.
The AI here is a model, the kind of AI behind ChatGPT and Claude, connected to your quoting data. In 2026 these models can read PDFs, images, and long email threads in one go. Anthropic's Opus 5.5, for example, takes in about a million tokens at once, where a token is a chunk of text, often part of a word. Anthropic says a million tokens is about 555,000 words.
How is it different from an instant-quote website?
Instant-quote marketplaces and in-house quoting systems solve different problems. A marketplace such as Xometry or Protolabs prices the part with its own cost model and makes the part through its own network or factories. Protolabs says in its annual report that its software turns a 3D CAD file into a quote without most of the skilled quoting labor.
An AI quoting system built for your company prices with your costs, your machines, your outside vendors, and your judgment about each customer, and your estimator signs the result. Your buyers get a fast quote from you, not from a marketplace.
What does the AI actually do with an RFQ?
It drafts each line of the quote and shows where every number came from. Below is a made-up example for a repeat customer asking for 50 machined aluminum brackets at a new revision. The format is the point: every value has a source, and anything the AI is unsure about is flagged for the estimator.
| Quote line | AI draft | Where the AI got it | What the estimator checks |
|---|---|---|---|
| Material | 6061-T6 bar, 52 pieces including 2 for setup | Drawing title block, material cost table | Current bar price |
| Run time | 0.42 hours per part | Actual time on the same part at the previous revision | Whether the new revision adds work |
| Setup | 1.5 hours | Same past job, actual setup time | Nothing, if the fixture is unchanged |
| Anodize (outside) | $3.10 per part, flagged | Last purchase order to the anodizer, 8 months old | Calls the vendor for a current price |
| Margin | 28% | Sales manager's floor for repeat customers | The relationship and the competition |
| Note to estimator | "Revision C adds a pocket not in revision B. Run time may be low." | Comparison of the two drawings | Adjusts run time |
The estimator's job changes from typing the quote to checking it. In this example, the AI flagged two lines for a closer look (the anodize price and the run time) and said why.
Can it read our drawings?
Current models accept PDFs and images as input (see Anthropic's PDF support), so they can read title blocks, notes, dimensions, and material callouts on a drawing, and compare two revisions to point out what changed.
They still make mistakes on dense drawings, small text, and tolerance callouts, and they do not replace a person who understands GD&T. That is why the systems we build flag anything critical (tight tolerances, special processes, unfamiliar materials) for the estimator instead of guessing. Drawings covered by ITAR or other export rules need a separate decision about where the model runs, which we cover below.
What data does it need from us?
Four things, roughly in order of importance:
- Cost tables: material prices, machine and labor rates, outside-process costs, and margin rules.
- Past quotes, including the ones you lost, with the price you sent.
- Job actuals: what each job really took in time and material.
- The rules of thumb your best estimator uses and has never written down.
The data does not need to be perfect. It needs to be clean enough for the parts you start with. Our article on why your parts list matters more than your AI model explains how to tell.
How accurate is it, and how do we know?
You measure it on your own past jobs before it touches a live RFQ. Pick 50 to 200 past quotes your estimators consider correct, hide the answers, and let the AI draft each one. Then compare line by line. That collection of past jobs with known answers is a test set.
Before testing starts, write down the bar the AI must clear. For example: 90% of quote lines within 5% of the estimator's number, and no missed special processes. If it misses the bar, the misses tell you what to fix, usually a cost table, a missing rule, or a part family that is not ready. After it passes, run it in shadow mode for two weeks: it drafts live RFQs on the side while estimators quote as usual, and the estimating lead compares the two.
Who signs the quote?
A named estimator signs every quote. In the systems we build, the AI can draft but has no permission to send email to a buyer, so a person has to review and press send. Every draft and every change the estimator makes is logged, which also shows where the AI needs work.
Will it replace our estimators?
No. It changes their day. Estimators spend less time typing line items and looking up old jobs, and more time on the calls only they can make: what a customer will pay, which job to chase, whether a tolerance is realistic on your machines. Our article Verify, Don't Retype covers how that change feels to the team.
The people are also hard to replace. Almost 1 in 3 US cost estimators is 55 or older, per the Bureau of Labor Statistics, and BLS expects all of the roughly 17,600 estimator openings a year to come from people retiring or moving on. In our view, the practical effect of AI is capacity. An estimator who reviews drafts can answer more RFQs in a day, including the small ones that used to get declined.
Is our pricing data safe? Will it train someone else's AI?
Anthropic and OpenAI say no for their business products and APIs. Anthropic states that by default it will not use inputs or outputs from its commercial products, including its API (the connection that software, such as a quoting system, uses to reach the model), to train its models. OpenAI states that data sent to its API is not used to train its models unless you opt in.
Beyond the lab's terms, safety comes from how the system is set up. The AI gets its own login with read access to the quoting inbox, the cost tables, and job history, and nothing else. It cannot see payroll, HR files, or email outside the quoting inbox, and every action it takes is logged.
What about ITAR or controlled drawings?
Controlled drawings need a deliberate decision about where the model runs. One option is a cloud model under a contract and hosting setup your compliance lead approves. Another is an open-weight model, one you can download and run on your own hardware, such as Meta's Muse Glimmer or Alibaba's Qwen3.8-27B, which are licensed for that use per our guide to the AI labs. You can also send controlled work through a separate quoting path that never touches a cloud model.
How fast can we get quotes out?
Same day is a realistic target for routine work. Vendors that sell quoting software say manual quoting takes 15 to 20 minutes per machined part and their software gets it to about 90 seconds, and that a circuit board bill of materials drops from 2 to 4 hours to 5 to 10 minutes. Treat those as vendor figures. Machine shops that rank well in Modern Machine Shop's Top Shops survey quote in about one day and turn more quotes into orders. The limit becomes how fast your estimator reviews, which is why the review screen matters as much as the AI.
Which RFQs should we start with?
We recommend starting where the AI has the most history and the least risk: repeat parts, part families you quote every week, and the small jobs you currently decline or let expire. Leave complex assemblies, new processes, and your largest customers for later, once the test set shows the AI is ready for them.
What does it cost?
Model fees are small. Opus 5.5 lists at $4 per million input tokens, so reading a 100,000-token drawing package (about 55,000 words) costs about $0.40. Most of the cost is building the system around your data, testing it, and training your team. Our engagements start at $10,000, and we price the rest after a diagnostic.
What to do this week
- Export your last 200 quotes with the price sent and whether you won.
- Ask your senior estimator to mark 50 they would call correct. That is the start of a test set.
- List where your cost data lives today (spreadsheets, ERP tables, someone's notebook) and who keeps it current.
- Write down the three questions your estimator always asks before pricing a part. Those become the first rules.
Questions people ask
What is AI quoting for manufacturers?
Software that reads an RFQ and its drawings, finds similar past jobs, and drafts a price from the manufacturer's own costs and rates. An estimator reviews, adjusts, and sends every quote.
How accurate is AI quoting?
It depends on your data and part mix, so measure it on your own past jobs. Test the AI on 50 to 200 past quotes with known answers and set a written bar, such as 90% of lines within 5% and no missed special processes, before it drafts live quotes.
Does AI quoting use our data to train public models?
Not by default with Anthropic's or OpenAI's business products. Both state that data sent through their APIs is not used to train their models unless the customer opts in or sends feedback.
Terms in this piece
- Model
- the AI itself, trained by a lab such as Anthropic, OpenAI, or Google. Products such as ChatGPT and Claude run on top of models.
- Token
- a chunk of text, often part of a word. Labs price their models per million tokens.
- Test set
- past jobs with known answers, set aside to judge the AI before it does live work.
- Shadow mode
- a period when the AI drafts live work on the side while people work as usual, so the two can be compared.
- Open-weight model
- an AI model you can download and run on your own hardware, so drawings never leave your building.
- First pass
- the AI's draft of a quote, which a person reviews before anything goes to a buyer.
