CNC Quoting Software: What AI Actually Does for Job Shops

For every job you win, you quote about two. In CNCCookbook's survey of shop owners, the average quote-to-win rate came in around 51%, with the best shops near 70%. Roughly half the effort spent at the quoting desk never turns into an invoice.

Treat that number as a sense of scale rather than an industry law. It comes from a single survey whose sample size and year are not stated. But you already see something like it in your own books: quoting is a production cost you cannot bill.

This guide covers three things. Why CNC quoting takes so long, how to build your machine hour rate from real components, and where AI actually enters the calculation. Like our other manufacturing guides, it is part of our industry-by-industry AI map.

Why does CNC quoting take so long?

Because the accurate method is expensive. The only way to know a part's real cycle time is to build the toolpath in CAM. That means hours of engineering time for a job you may not win. Most shops fall back on spreadsheets and experience instead.

The same survey found the most common quoting method by a wide margin is a spreadsheet. Right behind it sits the method everyone in the trade knows but nobody writes into a proposal: the eyeball guesstimate. Looking at the part and calling a number.

Reading that as sloppiness would be unfair. Guessing is an economically rational defense. A shop that expects to lose half its quotes cannot afford to run trial machining on every part, because that donates engineering hours to jobs it will never see.

The real problem in quoting is not accuracy. It is the cost of accuracy. The precise method demands too much work for jobs you will not win.

That gap is exactly where AI enters. The goal is not to replace the estimator's judgment. It is to stop that judgment from starting from zero every single time.

How do you calculate your machine hour rate?

Your hourly rate is the sum of five components: machine depreciation, energy draw, maintenance and tooling wear, operator cost, and that machine's share of shop overhead. Setup time sits outside this figure as a separate line. None of that is controversial in the trade.

What is contested is what those components cost in your shop. Published averages for machining labor rates are unreliable; the figures you find online are individual shops advertising their own pricing, not sector benchmarks. Energy tariffs and currency movements shift the number more than once a year. So what follows is a skeleton, not a price list.

Work out annual cost per machine, then divide by annual running hours:

• Depreciation: Machine price divided by economic life. Assume seven years and the annual figure is one seventh.
• Energy: Average kW draw times annual running hours times your industrial tariff. Do not forget the compressor and lighting.
• Maintenance and tooling: Last year's actual invoice total. Look it up, do not estimate.
• Operator: Gross wage plus employment costs. If one operator runs two machines, allocate half.
• Overhead share: Rent, accounting, admin staff. Allocate by machine count or floor area.

Say those five come to 1,000 units per year for one machine, and that machine actually cuts metal for 1,800 hours. Your hourly cost is 0.55 units. The critical figure there is 1,800. Plenty of shops run the math on 2,400 hours, actually achieve 1,500, and quietly underprice every quote without noticing.

Handle quantity separately. Setup cost is fixed, so unit cost falls as the batch grows. The difference between a one-off price and a fifty-piece price usually comes from how that setup is spread, not from cycle time.

Where does AI actually fit into the calculation?

In three places: estimating cycle time and cost from part geometry, turning incoming requests into structured data, and making it visible why a price came out high. The first has the strongest research behind it.

There is a substantial peer-reviewed literature on estimating cost from 3D CAD models. Three-dimensional convolutional networks clearly outperform two-dimensional approaches at this task, and architectures that model the part as a feature graph produce comparable results. Pricing from a STEP file is not a vendor fantasy; it is a studied problem.

The more interesting finding for a shop floor is this: some of these models do more than output a number. They can visualize which machining feature drives the cost. Instead of "this part is expensive," the system can point at a deep pocket or a tight tolerance. At the quoting desk that is far more useful than a black-box price. Telling a customer "make this pocket 2 mm shallower and the price drops" is a sentence that wins work.

There is also research on estimating cost from 2D engineering drawings, which matters because plenty of RFQs still arrive as PDFs and DWGs rather than solid models. That work is less mature than the 3D side.

Parsing the incoming request is the easiest win technically and the fastest to pay off. Having a language model pull quantity, material, tolerance and lead time out of a customer email and populate your quote form is a workflow you can build today. The model is not predicting anything here, just reading and organizing, so the error risk is correspondingly low.

A scenario: the shop with no quote archive

Picture a six-machine tool and die shop. Requests arrive by messaging app and email, calculations live in spreadsheets scattered across folders, and nobody records actual cycle times. This is the most common setup we run into. Drop automated quoting software into that shop today and there is no data to train anything on.

So the first step is not buying software. It is pulling the last twelve months of quotes into one table: part, material, quantity, price quoted, won or lost, and if won, how many hours it actually took. That last column is usually the hard one because nobody tracked it. You reconstruct it approximately from machine counters, operator timesheets and dispatch records.

Two things happen once that table exists. First, the owner sees for the first time, in numbers, which job types they systematically underprice. The result is usually surprising: it is not the lost jobs that hurt, it is the cheap ones you won. Another pattern shows up repeatedly too: a meaningful share of quotes never resolve at all, with the customer neither accepting nor declining. Unless that silent group is measured separately, your win-rate math stays misleading. Add a third status to the table: won, lost, no response.

Second, you now have data a model can learn from. The approach we follow in reducing scrap rate with production data runs on the same logic: measurement discipline first, algorithm second. Shops that skip that order end up where we described in our predictive maintenance starter guide. Without data discipline, AI in a machine shop means querying an empty memory.

What does automated quoting software cost?

The honest answer is that we do not know, because nobody publishes it. None of the known products in this space list prices on their websites or in software directories. They all run on a request-a-demo model, which means pricing is set in a sales conversation.

That model implies negotiation based on shop size. Do not walk into the demo unprepared. Ask these:

• What does it read besides STEP? If most of our incoming files are PDFs, how much use is this product to us?
• Is the estimate trained on our historical jobs or on industry averages? If the latter, it knows nothing about our machine park.
• Can it show why a price came out the way it did? If not, you cannot defend the quote.
• Can we export our data? When the contract ends, does the quote archive stay with us?
• Where does it pull material prices from, and what happens when they move?

Be skeptical of the "90% reduction in quoting time" figures that show up in sales decks. Those come from marketing material, not independent measurement. Building your own baseline is not hard: log your quote count and time spent per quote for one month and you have a comparison point.

Where it does not work

Automated pricing performs well on parts similar to ones it has seen many times. On complex die and mold work, unusual geometry and true one-offs, the estimate drifts. That is not a hidden defect in the products; it is inherent to the method. A model cannot know what it has never seen.

There are other things the model cannot see. Your machine utilization, how urgent the job is, your history with the customer, whether you need cash this month. All of these move the price and none of them appear in a CAD file. Position automated quoting as a first draft that lands on the estimator's desk, not as the final word.

Material is a separate risk. Metal prices move with commodity markets, and if you buy imported stock, with currency as well. In that environment, issuing a quote valid for thirty days means quietly absorbing the risk yourself. Set your validity period against how volatile your material actually is, and print it on the quote.

Three more mistakes come up repeatedly in shops moving to automated quoting:

• Burying margin inside cost. Margin belongs on its own line so you know what you are conceding during negotiation. Buried, you will discount below cost without seeing it.
• Not dividing setup across quantity. If a one-off and a fifty-piece run go through the same logic, one of them is wrong.
• Not archiving lost quotes. A shop that records only the jobs it won can never learn where its pricing sits too high. Loss data teaches as much as win data.

Frequently asked questions

How is a CNC part price calculated?

Material cost, cycle time multiplied by machine hour rate, setup time, tooling wear and margin. The formula is simple; estimating cycle time accurately is the hard part. That is why two shops can quote wildly different prices on the same drawing.

Can you generate a price automatically from a STEP file?

Yes, and there is serious research plus commercial products behind it. But accuracy depends on feeding the system your own historical jobs. Out of the box, a tool pricing against industry averages knows nothing about your machine park or your operators' pace.

What if the customer sends a sample instead of a drawing?

Then automated pricing has no input. The part has to be measured and modeled first. Scanning and reverse engineering is a separate cost line and belongs in the quote. In practice that work is often done for free, and written off entirely when the quote is lost.

Is this too early an investment for a small shop?

The software might be early; the data discipline never is. Even at fifteen quotes a month, keeping past quotes in one table improves pricing quality within a few months. Make the software decision after that table fills up.

What should you actually do?

• Measure real running hours. Without knowing how many hours per year each machine actually cuts, your hourly rate is fiction.
• Pull twelve months of quotes into one table. Won, lost, actual hours. Three columns is enough to start.
• Start recording actual cycle times today. No estimating system can be built without that column.
• Automate email parsing first. Lowest risk, fastest return.
• Go into demos with a question list. If a vendor cannot answer the five questions above clearly, do not proceed.

AI in quoting offers no magic that removes judgment from the process. What it does is more modest and more valuable: it remembers what the shop did before. If your memory is currently scattered across spreadsheets in different folders, that is the first thing to fix. If you want a hand putting that structure in place, the YZ Uzman team is a phone call away.