AI-Powered Quoting for
CNC Manufacturing

The client — a twelve-person shop running three-axis and five-axis mills, lathes and wire EDM — does good work. But the owner was the only person who could estimate jobs — reading drawings, mentally decomposing geometry into operations, checking material stock, guessing setup time, and typing a price into a spreadsheet attached to an email. Every quote took one to three days. During busy periods, RFQs sat unanswered for a week.

The math was brutal. Prospects who needed parts in two weeks would not wait three days for a quote. They would send the same drawing to five shops and go with whoever responded first. Alpine was losing deals not because their prices were wrong, but because their prices arrived too late. The owner knew this. He also knew he could not hire a second estimator at his scale — the overhead did not justify itself for a 12-person shop. He needed a system that could do 80% of the estimating work so he could focus on the 20% that required his judgment.

We scoped the project around that constraint: build a pipeline that ingests CAD files, extracts the geometry and features that drive cost, runs a trained model against historical job data, and presents a draft quote for human review. The owner stays in the loop on every quote — but instead of building each one from scratch, he reviews and adjusts a machine-generated estimate. The bottleneck shifts from creation to validation, and validation is fast.

The estimator bottleneck is a revenue ceiling, not a staffing problem

Alpine’s owner thought he needed a second estimator. He did not. He needed to stop being the bottleneck. The real problem was not that quotes took too long to build — it was that every quote required the same person. When that person was on the shop floor, on vacation or handling an urgent customer call, the quoting queue stopped. We reframed the project around throughput: the goal was not to replace the estimator but to give him a draft he could review in five minutes instead of building from scratch in two hours. The ~22% revenue growth came not from better prices but from responding faster and quoting more jobs.

Historical job data is the real asset — most shops do not know they have it

Alpine had three years of completed job records in their ERP — material consumed, machine hours logged, setup times, scrap rates, actual vs. quoted prices. Nobody had ever treated this data as a training set. We spent the first two weeks of the project cleaning and structuring it: normalizing material grades, mapping machine codes to cycle-time categories, and linking quoted features to actual production outcomes. That structured dataset became the foundation for the ML cost model. The model is only as good as the historical data, and most shops are sitting on gold they have never mined. The data cleaning phase is not overhead — it is the project.

CAD parsing is the hardest engineering problem in the pipeline

Extracting machinable features from a STEP file is not a solved problem in the general case. Holes, pockets and planar surfaces are straightforward. But compound curves, draft angles, thin walls and features that require five-axis access push the parser into territory where heuristics break down. We built a feature-recognition layer that handles the common 80% of Alpine’s part geometries reliably and routes the remaining 20% to the operator with a low-confidence flag. The temptation is to chase 100% automation. The correct approach is to handle the common cases well and make the edge cases visible, not hidden.

Confidence scores change how operators interact with AI estimates

Early prototypes showed a single price. Operators either trusted it or did not. Adding a confidence interval and a per-feature breakdown changed the dynamic entirely. When the model says a quote is $4,200 with 92% confidence, the machinist glances and approves. When it says $4,200 with 68% confidence and flags the surface-finish estimation as uncertain, the machinist digs into that specific feature and adjusts. The confidence score turned the dashboard from a take-it-or-leave-it system into a collaborative tool. Operator override rates dropped from 40% in the first month to under 15% by month four — not because we forced compliance, but because the model got better and the operators learned where to trust it.

The $62K deal: AI as a competitive advantage, not just efficiency

One RFQ came in for a batch of aerospace brackets. The customer had spec’d three-axis machining. The AI cost model flagged that a five-axis approach would reduce cycle time by 35% because it eliminated two setups and improved tool access to undercut features. The operator confirmed the recommendation, and Alpine quoted the job at a price the customer could not match elsewhere — because competing shops were quoting the slower three-axis approach. Alpine won the $62K contract. This was not a quoting-speed win. It was an engineering-insight win that only happened because the AI could evaluate alternative manufacturing strategies faster than a human estimator scanning a drawing.

Related use cases: Fleet Driver Management & Compliance Platform • AI Data Automation — Vendor Onboarding & Retail Pricing