Why Quoting Becomes the Bottleneck in Job-Shop Manufacturing
In make-to-order manufacturing such as precision machining, mold making, and contract parts production, the quote is the sales process. Yet in many shops, quoting lives in the head of one or two people, often a plant manager or production planner with fifteen years on the floor.
Dependence on a few experts: The judgment "two MCT setups, outsourced heat treatment, four hours" is rarely written down. When that person is on vacation, quoting stops.Slow replies mean lost orders: Suppose 120 RFQs arrive each month and each takes 40 minutes. That is 80 hours a month on quoting alone. On top of other duties, replies slip by three to five days, and customers place the order with whoever answered first.Errors surface too late: The gap between quoted and actual cost only shows up at job close-out. Even when loss-making jobs repeat, it is hard to tell which operation caused them.What AI Quoting Does and Does Not Do
The first design decision is where the system's responsibility ends.
What the AI does: It estimates machining time, routing (lathe → MCT → grinding, and so on), stock weight and material cost, and whether an operation will be outsourced. In short, it estimates cost.What it does not do: It does not set the selling price. Margins, customer-tier discounts, and rush surcharges are commercial policy, kept in a separate rules table. Policy changes then need no retraining, and pricing stays easy to explain to customers and colleagues.Keep an approval step: The system drafts, and the estimator reviews, adjusts, and approves. Those adjustments become the next round of training data.Building the Input Data
Features from drawings and 3D models
Overall dimensions and stock weight (bounding-box stock size)Tightest tolerance, number of precision tolerances, surface finish calloutsGeometric complexity: pockets, holes, and taps, number of machined faces, undercutsMaterial grade and post-processing such as heat treatment, plating, or anodizingWith STEP files, geometric features can be extracted automatically, but many RFQs arrive as 2D drawings only. A realistic start is a form where the estimator fills in 10 to 15 fields.
Linking past quotes to actual job results
The training target is not the quoted price but the labor hours, scrap, and outsourcing cost the job actually incurred. Quotes usually sit in spreadsheets while job results live in the MES or daily work logs, with different part codes on each side. Building a mapping table from quote number to sales order to work order is typically the most time-consuming part of the whole project.
Model Design
Per-operation prediction vs. direct total-cost prediction
A single model that predicts total cost is simple to build but tells you nothing when it is wrong. Predicting each operation's time and summing them needs more data, but it pinpoints errors, for example "grinding time was underestimated." In practice we recommend the per-operation structure.
Similar-part retrieval plus regression
Use the feature vector to retrieve the five most similar past parts with their actual results, and show them next to the regression model's prediction (gradient boosting or similar). Estimators trust a result far faster when they can see the past jobs behind it, not just a single number.
Using prediction intervals to flag uncertain quotes
Compute 10th and 90th percentile bounds with quantile regression. If the interval is wider than ±30% of the median, route the quote to manual review. New materials and unfamiliar geometries get caught automatically.
Operation and Validation
Feedback loop: For every won job, record the error between estimated and actual cost, and track mean absolute percentage error (MAPE) per operation each month.When to retrain: If a new machine or a growing share of new materials pushes one operation's error up for two to three consecutive months, treat it as a retraining signal.ERP and MES integration: Material prices should come from ERP purchase records and operation times from MES actuals, so the loop runs without manual work.Rollout Phases and How to Judge the Payoff
Phase 1 — Data cleanup: Map the last two to three years of quotes to job results. The gate to the next phase is whether several hundred records can actually be linked.Phase 2 — Similar-part recommendations: Measure whether retrieval alone, with no prediction model, shortens quote preparation time.Phase 3 — Per-operation prediction and approval workflow: Track quote turnaround in days, how much estimators adjust the draft, and the error between quoted and actual cost.POLYGLOTSOFT supports every phase, from cleaning up shop-floor quote and job data to building AI cost estimation models, ERP and MES integration, and approval workflows. If quoting is stuck on a single person in your shop, we would be glad to talk through how a system can share that load.
