RFQ Automation for Machine Shops: What to Automate First

For RFQ automation in a machine shop, start with the steps that are lookup and arithmetic: filing the inquiry, reading the drawing's written fields, finding similar past quotes, and applying price tables and customer rules. Keep the machine-time estimate and the final price with a person. This article is for owners, sales staff and estimators in shops that quote made-to-drawing work, and it explains the order with numbers from a quoting backtest we ran for one such shop. The backtest is an internal test on historical quotes, not a live result.

Zhang Chenxi (Nature), who builds agent systems for manufacturers and distributors. 10 October 2026. Drafted with AI assistance.

The question is which parts of a quote a program can carry today, which parts it cannot, and what to record so the second answer improves over time.

From inquiry to quote: eight steps and who should do each

A quote passes through the same steps in most shops, whether the estimator works in a spreadsheet or a quoting tool. The right-hand column is our recommendation, based on the work described below.

The sections below take these in order of how safe they are to automate.

Steps 1 to 3: filing, reading and finding are software work

Intake sounds trivial and carries more risk than it looks. In our drawing-reading test, eight drawings went through a minimal file conversion that silently dropped every dimension annotation and block note, about a quarter of the visible text. Nothing downstream could recover text that never arrived. If you automate one thing about intake, make it check that what the program received is what the customer sent.

Reading the written fields of a drawing is the part that language models do well. In a test on 32 drawings, each read by two model readers, every material reading was right: 38 where the drawing named a material, and 26 where the drawing left it out and the reader left the field blank. Finish readings were right in 60 of 64 under a semantic reading of the field; the four misses put a deburring note or "not stated" into the finish field. In a later run with a text-first pipeline, five key fields were read correctly in 70 of 76 cases; the six misses were all blanks, and there were no wrong values. Picking the overall size out of a sheet full of dimension numbers was far weaker, with 10 of 31 correct. So let software draft the fields, and have a person confirm size and material.

The material field needs that confirmation for a second reason. Of the seven drawings we could match to quotes, only one carried the same material grade as the quote. The others differed, or named only a general class. A drawing and the shop's price table do not always describe the same thing, so a correct reading still does not make a quote ready.

Finding similar past quotes is the next step to give software. Exact repeats are rarer than they look: most lines that appear more than once in the shop's files are revisions of the same quote made within a few days, and once those are merged, only about one line in twenty is a part quoted again on a later order. The value of a lookup is as a reference for similar parts. Using only fields from the quote sheets, nearest-neighbour lookup had a 22.1% median deviation on cost subtotal against 25.0% for our drawing-based pipeline on the same 51 lines. The gap was inside the noise, so a lookup that puts the closest earlier quotes and their dates next to a new inquiry is worth building before any drawing analysis.

Steps 4 and 6: arithmetic and customer rules are cheap to automate once the operations are known

Material cost is weight times the day's price. Finishing and secondary operations come from the shop's own price table. Customer rate, margin, tax and rounding come from rules per customer. Each is a formula or a lookup, and the check is simple: given the cost lines, does the program reproduce the price the shop sent?

On 28 historical quote lines, it did so within ±3% on 27 of them once customer-specific rates and set-quantity pricing were in the rules table, and on 23 with default rules alone. The customer rules were read from those same lines, so this is in-sample: it shows the program reproduces the structure of the shop's price formula, not that those customers will be priced the same way next time. The shop's quote sheets already hold these formulas, so most of the work is moving them into a form a program can run and keep current.

Once the list of operations is known, this arithmetic is cheap to automate; deciding which operations a part needs is harder. When a drawing does not say whether an operation is needed, the program can fill in what is most common for that kind of part and mark it, so the estimator can delete it, because a missing line makes the price too low and that is the costlier error. The default has its own failure, described next.

The cost-line backtest tells a more modest story. After we added the secondary-operation rules, the median deviation of the cost subtotal against the shop's own past quotes fell from 33.2% to 22.5%. That test had 14 quote lines, most from a single order, and the rules were tuned on those same lines, so the figure is in-sample. On a wider set of 51 lines matched to drawings, outside that tuning, the median deviation was 25.0%. Both are internal backtests on historical quotes, not live results, and they measure agreement with the shop's earlier numbers, which are themselves a judgement. On those 51 lines the material column was close, partly because unit prices were read back from the same lines. The secondary-operation columns were off by about half at the median and ran about 16% high overall; the largest single cause was parts classified as the wrong kind picking up operations they never needed. The subtotal benefits from machine minutes that ran high offsetting rates that ran low. The tests show the shop's formulas can be reproduced; choosing which formulas apply to a part still needs a person's check, and nothing here shows what the program would have saved or earned.

Steps 5 and 7: machine time and the final price stay with a person

Whatever method we tried, the machining-charge column was the weak one, around 50% median deviation across the geometry pipeline, the table lookup and the tree model. Two reasons stand out.

First, the minutes on a quote are a pricing number. The time an operator copies from the machine screen is pure run time, and the minutes on the quote sheet also carry setup, first-off and the shop's habits. Two same-size parts appeared in the history with machining charges about 2.6 times apart and no minutes entered on either. A program cannot learn a rule from a column where the rule is not written down.

Second, the final number reflects the relationship. Across consecutive versions of the same quote, the margin changed in 108 version pairs drawn from 39 quotes, and 99 of those changes were the same move, a five-point cut. That looks like a decision made in negotiation. A program can show the earlier price, the margin used and the room left, and the estimator decides.

So we designed the system as an assistant. It shows a number with the basis for each line, asks at most three questions, and the shop decides the price. For machine time, the useful output is a range with the reference quotes next to it, since a confident single number would hide a judgement the estimator needs to make.

Step 8: record the outcome, or nothing improves

In the shop's records, nothing tied a quote to what the job later cost or whether the customer accepted it. Weight, material used, machine time and outside-processing cost had no consolidated record either, and no single number ran from quote to work order to production log. So our backtests used the shop's own quote sheets as the answer key, and those sheets are themselves open to question.

The plan we wrote for the shop's next stretch of orders adds four small habits. Each draft carries a part number that the estimator copies to the work order, unknown weights are marked "to fill", the production log gets a machine-number column, and once a week someone spends ten minutes asking, for quotes sent two weeks earlier, whether they were won and at what price. The first three take seconds each, the fourth ten minutes a week. Together they give later work something to measure against: actual cost and actual outcome, instead of agreement with old quotes.

What to do first in your own shop

If you are choosing where to begin RFQ automation, work down the table. Get inquiries filed and matched to customers and parts, then put a similar-quote lookup in front of the estimator, then move the price tables and customer rules into a form a program can run. Start recording outcomes in the same weeks, because that data takes months to accumulate and nothing replaces it. Treat machine time and the final price as the estimator's, and judge any tool by whether it shows its basis for each line and flags what it assumed.

If you run a machine shop, or a distribution business with a similar quote-and-follow-up workflow, and want to see what this would look like on your own inquiries, send us one recent RFQ and how it was quoted. We reply with a written assessment of which steps an agent could carry and which should stay with your team. Details are on our page for AI agent systems for manufacturers and distributors.

If you want a system like this built around your own workflow, see AI agent systems for manufacturers and distributors or write to hi@towow.ai.