AI quoting for job shops: what quote histories teach it

How does AI quoting work in a job shop?#

AI quoting in a job shop works by matching a new request to patterns in past work. The tool reads part features from a model or drawing, such as material, size, tolerances, hole and pocket counts and finish requirements, compares them with parts the shop has quoted or made, and predicts setup time, run time, material and outside processing.

Many tools then apply the shop's own rates and rules to turn time into a price, and an estimator reviews the result. Some tools also show which past jobs drove the estimate, so the estimator can see why a number looks high or low.

The prediction is only as good as the history behind it. A model trained on quotes alone learns what the shop used to guess; a model trained on quotes linked to actual costs learns what parts really took.

The four inputs that teach a quoting model#

A quoting model needs four linked inputs, and most shops have only some of them in good shape. The table shows where each one usually lives and where it tends to break.

The weakest input is usually the outcome. Estimators remember which quotes they won, but the ERP often leaves a lost quote sitting open, and the reason the customer walked away lives in an email or a phone call. Even a rough reason code, such as price, lead time or no response, makes the history far more useful.

Why do actuals and win-loss records matter more than volume?#

Actual costs and win-loss records matter more than quote volume because they are the main feedback a quoting model gets. Without actuals, the model cannot tell a quote that was right from one that lost money. Without outcomes, it cannot tell a competitive price from one the customer ignored.

That is why a shop with fewer quotes but clean time tickets can train a better model than a shop with a huge, unlinked estimating archive. It is also a good reason to share job history, not just quotes, when you set up a tool.

Before buying a tool, run a simple check. Pick recent jobs and confirm you can trace each one from quote to job to time tickets to shipment. If the trail breaks, fix the habit before you feed the history in.

The customer-drawing caveat#

The customer-drawing caveat is that the most informative input, the part geometry, usually belongs to someone else. Customers send models and drawings for quoting and production, and NDAs and purchase order terms commonly limit their use to those purposes.

Using drawings inside your own quoting tool may fit within those limits, depending on the terms. Letting a vendor train a shared model on them, or licensing them to a third party, is a different use that generally needs the customer's permission. Drawings for defense programs can be ITAR technical data, and some commercial aerospace and industrial work falls under the EAR, so those files stay out of any outside tool unless your export compliance lead approves.

One workable approach keeps the drawings out and uses features your estimators already recorded, such as material, envelope, operation list and tolerance class, as the training input. Ask counsel whether those derived records raise the same concerns under your customer agreements.

What to ask a quoting vendor about your data#

Ask a quoting vendor how your data is used before you sign, because the answer decides whether your history improves only your tool or everyone's. These questions fit in one call.

• 1Does your shared model train on our quotes, jobs or drawings, and can we opt out?
• 2Is our history used only to tune a model for our shop?
• 3Will you test the tool on our past jobs, held back from tuning, and report its error against our actual hours by part family?
• 4How do you handle export-controlled parts and drawings?
• 5Where is our data stored, and who can access it?
• 6Can we export our history and the tool's estimates if we leave?
• 7Will you delete our data, including derived features, when the contract ends?

Illustrative: a CNC shop tests its own history#

Illustrative: a fictional CNC turning and milling shop that serves industrial equipment makers wants to try an AI quoting tool. Its ERP holds years of quotes, jobs and time tickets, but lost quotes were rarely closed out, and drawings sit in a shared folder sorted by customer.

The owner has an estimator trace recent jobs from quote to shipment and finds labor booked to the wrong operation on some second-op work. The shop fixes the booking habit, closes out old quotes as won or lost where emails show the answer, and keeps drawings out of the vendor upload. The vendor tunes a shop-specific model on quotes, actuals and recorded features only, with no shared training.

Estimators still review every quote. What changes is where they spend their time: on new part families and unusual tolerances rather than repeat work.

Your quote history has a second use#

A linked quote-to-actual history can also be licensed to AI developers who build quoting, planning and manufacturing agents, separate from any tool the shop buys. That use is a license, not a sale: the shop keeps ownership, sets scope and approves each step.

How SourceX handles quote histories#

SourceX handles a quote history as a supplier package in the SourceX five-step transaction. The fit check collects metadata only, such as the ERP in use, years of quotes and whether time tickets link to jobs. Anything designed by customers, and any work under export controls, is left out of manufacturing packages. Value is known only once a buyer engages, and the SourceX Enterprise Data Value Framework sets out what drives it.

If the shop proceeds, Rights checks customer and supplier terms, Preparation codes customer names and strips drawing references from notes, and the owner signs off on scope before Delivery. A SourceX Evidence Packet records the provenance, rights and permitted use behind the package, how personal details were handled and who authorized release. SourceX never hosts large histories; they remain on the shop's servers or travel on encrypted drives.