What AI means for the modern precision machine shop

How many manufacturers use AI?

In 2025, 20 per cent of EU firms with ten or more staff used AI, up from 13.5 per cent in 2024 (Eurostat, 2025). Manufacturing as a whole came in at just over 17 per cent. Basic metals and fabricated metal products, the group that includes subcontract machining, reached just over 14 per cent, against 8.6 per cent a year earlier.

Statistics Netherlands (CBS) counts a different population: firms with two or more people working, the smallest included. On its provisional figures, 16 per cent of Dutch manufacturers used AI in 2025, against 10 per cent in 2024 and 7 per cent in 2023. Because the populations differ, the two series cannot be compared directly.

They do point the same way: about one manufacturer in six now uses AI, and the share has risen two years running. Where those firms put it to work is the next question.

Where is AI used now, and where not yet?

Of the Dutch firms that use AI, 35 per cent apply it to marketing or sales and 18 per cent to production or service processes (CBS, 2025, provisional). Across the EU, small firms use AI in production less often than large ones: of small firms using AI, 19 per cent do so in production processes, against 33 per cent of large ones (Eurostat, 2025).

Much of the AI help on offer for machining sits in CAM: suggested toolpaths, tools and cutting data. That work starts once the order is in. Before then every request needs a cost price, including the many you will not win, and that cost price is made in the estimating office.

How drawing, estimate, CAM and NC hang together when software keeps one record per part is covered in software-driven machining, from revision to quote.

What changes in the estimating office?

A request passes through intake, drawing, estimate and quote, and at each step software can put down a first version. It pairs model and drawing for each part, takes material, tolerances, roughness, finishing and threads off the drawing as proposals, and builds the estimate lines for material, machining, setup and programming.

Keep three things apart here: features, requirements and operations. On the model, feature recognition might count 32 holes, 8 of them threaded. The requirement sits on the drawing: an M6 thread, a fit, a position tolerance. The operation is what the manufacturing engineer makes of it: drilling, tapping or reaming, and in which setup. Each of the three calls for a different judgement.

The cost engineer and the manufacturing engineer then make the decisions that cost money: how many setups, which machine, what is subcontracted and what margin goes on top. Part of the time goes on the retyping and looking up around those decisions. Quoting faster with AI shows where that time sits, and can AI read an engineering drawing what software does with one.

Is it real AI or a formula?

Both, in different places. Recognising and proposing is where AI is used: feature recognition on the 3D model, drawing reading to take the requirements off a PDF drawing, suggesting a material from what the title block says. That is work with uncertainty, so each result is a proposal for a person to confirm and correct.

The cost price itself is arithmetic: time times rate times quantity, plus material and subcontracted work. It should not be a prediction. You should be able to check every estimate line, with its inputs beside it. An estimate is fine if labelled as one; a figure without provenance is not.

A general chat assistant can describe a drawing and name a price, but none of your machines, hourly rates, purchase prices or subcontractors stand behind that number. Estimating software should work from your shop's own data.

What holds small manufacturers back?

According to Eurostat (2025), EU firms that considered AI but do not use it point mainly to a lack of relevant expertise (71 per cent), unclear legal consequences (53 per cent) and concerns about data protection and privacy (49 per cent). Dutch firms in the same position most often cite lack of experience, followed by privacy and the legal consequences (CBS, 2025).

For small firms, especially in manufacturing, the OECD (2025) adds two hurdles: finding AI software suited to their work, and a lack of quality data and of digital readiness. In a survey of SME manufacturers in North Brabant, 83 per cent lacked a clear view of which parts of their production process lend themselves to automation with AI (FME and partners, 2026).

For a twelve-person subcontract shop, that means starting from a question in your own work rather than an AI project, such as how much time goes on retyping per request, or where drawing requirements get missed. What to ask a vendor about your drawings and your customers' is in what to ask an AI vendor about customer drawings.

Do you need years of data first?

In a white paper on one-off and small-batch production, Fraunhofer IPT wrote in 2021 that manufacturers seldom have the relevant data in sufficient quantity and quality, nor the systems to process them well. In a report on robotics, TNO (2026) notes that a data-driven AI model can be hard to prepare for every situation because there is much variation and uncertainty. The same report describes Dutch manufacturing as high-mix, low-volume.

Software that derives a price from your archive of past quotes needs a clean archive first, with actual hours attached; many small firms lack one, and the OECD notes that SME company data are often not ready for AI use. Even a clean archive records what happened, not why: the ERP holds the hours an order took, not the extra setup a tight fit called for. A price built on it repeats the old quote, along with any margin that was already too thin.

An estimate is better built from what is already on the desk: the model's geometry, the drawing's requirements, and your own machines, hourly rates and purchase prices. The result is a pre-calculation that shows its inputs line by line, even for a first part.

How far to trust such an estimate is the subject of how accurate an AI estimate is, line by line.

What stays with your skilled people?

The choices that make or break a price: how the part is held, what risk you take on a thin wall or a tight fit, what margin you dare to ask. Software can propose a build-up; the judgement on it stays with the cost engineer and the manufacturing engineer. That is what human-in-the-loop means here: no price leaves the shop until a skilled person has signed it off.

What does change: when an experienced cost engineer corrects a proposal, that correction with reason can be recorded. The line then shows that setup time went up, and why. Shop knowledge becomes something a colleague can read back. Keeping a cost engineer's know-how goes into how.

When you choose software for this work, look at what sits under the AI: does it calculate with your machines and rates, and can you check every line? The questions to ask are in estimating software for machining.

Where is this heading?

The next steps lie after the quote: cost prices calibrated against actual hours, times from the CAM project and the NC program instead of an estimate, and a quote that reaches your ERP without retyping.

Calibration will show where estimate and shop floor part ways, per operation and per machine, and the judgement on that stays with the people who do the work. Our ambition goes further: the knowledge of your whole shop in one connected model, from the first drawing to the part that ships.