How to Verify AI CNC Accuracy Claims: A Buyer's 5-Point Checklist

Why accuracy numbers are harder to compare than they look

Two quotes can both be honest and still be incomparable, because they are not reporting the same statistic. ISO 230-2 positioning accuracy A is a bidirectional range built from mean deviations plus a multiple of the standard deviation across all targets and cycles. Some builders quote a plus-or-minus half-range instead. Other national codes use different target counts, cycle counts and evaluation rules: VDI/DGQ 3441 is common in German and EU acceptance, ASME B89.1.12 in North America, and JIS B 6336 / JIS B 6338 in Japan.

This matters more than any single number. A machine reported at "5 µm" under one code and another reported at "±3 µm" under a different code may be the same machine. The only figures you can legitimately compare are figures reported to the same standard, the same edition, the same environment and the same compensation state. This is a measurement-literacy point, not a criticism of any country's industry — builders everywhere quote the code that suits their home market.

There is a second, subtler trap. Repeatability is often more important to you than accuracy. A systematic positioning error can be compensated in the controller; a wide random scatter cannot. In batch production, a machine that always lands 4 µm off target is frequently a better asset than one that lands anywhere between zero and six.

Point 1 — Demand a documented ISO 230-2 / GB/T 17421.2-2023 report, not a catalog number

The international reference for axis positioning is ISO 230-2:2014. The current Chinese national counterpart is GB/T 17421.2-2023, published 6 August 2023 and in force since 1 March 2024, which is an identical (IDT) adoption of ISO 230-2:2014 and fully replaced the 2016 edition — that edition was formally withdrawn on 1 March 2024.

Source: National public service platform for standards information (std.samr.gov.cn / openstd.samr.gov.cn) and the National Digital Library of Standards; GB/T 17421.2-2023, status "current", issued by SAC/TC 22.

A report you can actually use will state all of the following:

That last item is the one buyers forget. A machine measured with full volumetric compensation enabled is a different machine from one measured bare, and you are entitled to know which one you are buying — and, more importantly, what happens when the compensation table drifts out of date. Ask who re-measures it, how often, and at whose cost.

Point 2 — Make a five-axis claim cut the S-shaped test piece

Static tests measure the machine. A cut part measures the process. For five-axis machines the most informative single workpiece is the S-shaped test piece: a thin-walled twisted ruled surface on a rectangular base. It was standardised in China as GB/T 39967-2021, published 9 March 2021 and in force since 1 October 2021, and it was incorporated into ISO 10791-7:2020 as the M5 free-form surface test.

The reason it is hard is precisely the reason it is diagnostic. Machining the S-shape forces continuous open-angle to closed-angle transitions, repeated rotary reversals, and thin-wall cutting where chatter and deflection cannot hide. A rotated cube and a polished demo video prove none of that.

Sources: GB/T 39967-2021 standard record (national standards platform; drafting bodies include Chengdu Aircraft Industrial, the National Machine Tool Quality Supervision and Test Centre, Tsinghua University, Kede Numerical Control and Qinchuan Machine Tool); ISO 10791-7:2020 test-piece catalogue.

Better still: run your own part family. Specify acceptance as cycle time, first-pass yield and Cpk on the characteristics that actually matter to you, over a defined sample size and run duration. A standard test piece lets you compare two machines; a run-off on your parts tells you whether you should buy either.

Point 3 — Judge thermal compensation by decay rate, not by the "AI" label

Almost every machine now advertises thermal compensation, and a growing number call it AI. The label is not the evidence. What you want is a drift curve: how far does positioning move, and how fast, over 8 hours, 24 hours, and a multi-day continuous run — and under what ambient conditions.

At CCMT2026 (the 14th China CNC Machine Tool Fair, Shanghai, 21–25 April 2026), the review of vertical machining centres published by the organiser, the China Machine Tool and Tool Builders Association (CMTBA), recorded several current approaches. Reported domestic examples include a machine holding whole-machine temperature spread within 0.5 °C over 97 hours of continuous cutting, and a linear-motor three-axis model with positioning accuracy of 0.003 mm. International exhibitors showed their own routes to the same problem: one Japanese builder's 3D calibration and thermal-affinity concept with real-time sensor-driven displacement compensation, and another's spindle-jacket and ballscrew-nut cooling to suppress thermal growth at source.

Source: CCMT2026 exhibit review — "Overview of Vertical Machining Centre Exhibits," published on ccmtshow.com (April 2026), organised by CMTBA. The organiser notes the views are the author's own. All figures above are vendor or exhibition cases, not universal specifications.

Two more contract-grade details: ISO 230-3:2020 covers the determination of thermal effects and gives you a neutral vocabulary for the test, and ISO 230-4:2022 covers circular (ballbar) testing, which is a fast, cheap way to detect servo mismatch and backlash long after acceptance.

Point 4 — Demand longitudinal accuracy-retention evidence

This is the point most buyers skip, and it is the one that costs the most money. Delivery-day accuracy is a snapshot. What you own is a machine that has to hold that accuracy for a decade.

The CMTBA's CCMT2026 review of design maturity and manufacturing readiness puts the issue in unusually plain language: for a long time, domestic machines showed acceptable accuracy at factory inspection but degraded over extended service — meaning accuracy remained a "medical-checkup indicator" rather than an "in-service indicator" the customer could feel. The review's own framing is that accuracy retention is designed in, not compensated in, and it cites a manufacturer that put a three-year-old, fully loaded production machine on the stand and showed it had held its accuracy — a more persuasive demonstration than any fresh test report.

Source: CCMT2026 exhibit review — "Forward Design and Manufacturing Maturity: The Dual Advance of Domestic Machine Tools," ccmtshow.com (April 2026), organised by CMTBA. Manufacturing readiness level (MRL) dimensions cited: reliability, accuracy retention, process stability, whole-lifecycle user satisfaction.

What to ask for, in writing:

• Re-test data at defined intervals — same ISO 230-2 protocol, same machine, at 12 and 24 months, or whatever the supplier's own service regime produces.
• A named re-calibration interval and price — who measures, who pays, and what happens to the compensation table.
• Reference customers at comparable duty cycles — ideally in your region, ideally in your industry.
• The failure taxonomy from service logs — not a single MTBF number. Widely circulated MTBF figures comparing domestic and imported machine tools are republished by secondary sources without a traceable primary methodology, so we do not publish them here and recommend you treat any single figure as directional until a service log supports it.

Point 5 — Get controller origin, software ownership and service radius in writing

Accuracy is produced by a chain — structure, drive, feedback, controller, and the compensation model sitting on top of it — and it is maintained by people. Ask who owns each link.

• Controller origin. Which company's control is it, which country, and is the supplier an integrator or the owner? This determines how fast bugs get fixed and whether firmware updates are a phone call or a negotiation.
• Who owns the compensation model. If the thermal and volumetric models are third-party black boxes, ask what happens when the machine behaves unexpectedly.
• Data access. Open interfaces such as OPC UA or MTConnect let you monitor drift yourself rather than taking someone's word for it. Confirm what data leaves your network and who can see it.
• Service radius. Distance, not price, sets lifecycle cost. A regional hub with parts on the shelf beats a cheaper machine whose nearest engineer is a flight and a visa away.

The honest backdrop

None of the above is an argument against Chinese machines — we sell them, and the direction of travel is genuinely positive. The same CMTBA review that flags the retention gap also records domestic five-axis models at 2 µm positioning and 1.8 µm repeatability, linear-motor three-axis models at 0.8 µm repeatability, and a gantry machine holding 5 µm positioning over a 3,000 × 6,500 mm envelope. Those are real numbers attached to real machines.

The honest structure of the argument is therefore this: short-term accuracy has converged materially; long-term retention is where differentiation now lives; and neither fact tells you anything about the specific machine in front of you. Verify the machine, on your parts, over a run length that resembles your production. A national trend is not a substitute for your own acceptance protocol.

The one-page checklist to put in front of a supplier

Data credibility rating

Honest limitations

• Exhibition and vendor figures are not specifications. Every performance number in this article sourced from CCMT2026 describes one machine at one event. Ask for the equivalent data measured on a production machine in a production environment.
• Standards move. ISO and GB editions are revised; the Chinese positioning standard changed edition as recently as 2024. Always cite the year in your contract.
• This is not a legal document. It is a buyer's framework. Have your own quality engineer or a third-party metrology laboratory write the acceptance protocol for your parts.
• We are a supplier with an interest. We benefit when you buy machines and service. That is exactly why you should insist the items on the checklist are written into the contract rather than promised — a supplier confident in their data will put it in writing.

Guidance

• Convert every adjective into a document. "AI" becomes a drift curve. "Sub-micron" becomes a test report. "Supported" becomes a response time. Each conversion removes a future argument.
• Compare like with like, or do not compare at all. Put both quotes onto the same standard, edition, environment and compensation state before the numbers mean anything.
• Buy retention, not delivery. A machine that holds its accuracy for ten years is worth more than one that measures better on day one. Weight points three and four accordingly.
• Make SAT the payment gate. It is the single strongest clause available to you and it costs nothing to ask for.
• Prefer the supplier who publishes provenance. Where two machines look similar on paper, the builder that states controller origin, test protocol and service terms is the one you can plan around.

Frequently asked questions

Related reading