Where an Estimator's Day Actually Goes
Most of an estimator's time is not pricing. It is opening emails, finding the latest drawing revision, reading notes, hunting through the ERP for a similar part, checking what the last run cost, and asking a vendor for a plating price. Those lookups are repeatable and mostly retrieval work, which is exactly what modern language models do well with retrieval over your own records.
0005 inch bore on a thin wall will hold, remains with the human.
What AI Can Do Today
Practical tasks with current open-weight models and retrieval over your own data are narrow but valuable. They work best when each output shows its source so the estimator can verify in seconds.
- Triage RFQ emails: part count, quantity, due date, missing information
- Extract material spec, finish, tolerances and quality notes from a drawing
- Find similar parts and show last-run actual hours, scrap and price
- Draft the quote with line items for estimator edit
- Draft a reply asking the customer for missing data
Limits You Should Expect
Reading drawings is the weak link. Clean PDF drawings with text layers extract well; scanned prints, old tolerances in handwritten notes and complex GD&T frames need human checking. AI will not reliably predict machining time for a new geometry without your own history, and it will not know that your 5-axis machine is booked for three weeks.
Treat every extracted value as a suggestion with a citation, and keep a rule that the estimator confirms all material, tolerance and process callouts before a price goes out.
Try it here, free
CNC Quote Review Checklist Before Sending
Thirty-plus checks to run before a CNC quote goes out, from drawing and tolerance review to cost build, risk and terms.
Keeping Drawings Private
Defense and aerospace RFQs often contain export-controlled drawings. Sending them to a public AI service can create an unauthorized disclosure under ITAR or breach CUI handling requirements. A private deployment runs open-weight models such as Llama, Qwen or Mistral class models with vLLM or Ollama on GPUs inside your facility, with retrieval over your ERP and document store, user-level permissions, and logs.
Netray builds this, but your compliance officer remains responsible for approving the boundary and handling.
Is It Worth It for Your Shop?
It tends to pay off when you receive dozens of RFQs a week, your ERP holds years of job actuals, and the estimators are the bottleneck. It does not pay off when you quote a few jobs a week, when your history is sparse or inaccurate, or when the delay is waiting on vendors. Test it on ten of your old quotes first: compare the draft to what you quoted and to what the job actually cost.
If it does not save estimator time on that test, do not scale it.
Frequently asked questions
Can AI read machine shop drawings and extract requirements?
It can extract material, finish, quantity and many notes from clean drawings, especially PDFs with text. Scanned or low-quality prints, and complex GD&T, still need a human check. Use extraction as a draft with citations to the drawing location, and have the estimator confirm callouts before quoting.
Will AI replace machine shop estimators?
Unlikely in the near term. AI removes lookup and drafting work, but pricing decisions involve risk, capacity, customer history and judgment about tolerances. Shops use it to let experienced estimators handle more RFQs, not to remove them. Human approval stays on every quote.
Is it safe to use AI with ITAR or CUI drawings?
Only if the AI runs inside your controlled boundary, on infrastructure you manage with US-persons access and logging, and your compliance officer approves it. Public AI services generally should not receive controlled drawings. Private open-weight model deployments on your own hardware are the common approach.
We can pilot a private estimator assistant on your ERP history and a sample of past RFQs, then compare its drafts to what you really quoted. No cloud upload of your drawings is needed.