JAPAN VIDEO REVIEW · FULL AUTOMATION
Does NC Programming Eat Half of Manufacturing Cost? The Spec-Sheet Honesty of Japan's ARUMCODE Official Video
01What this official video is about
ARUM is an AI startup in Osaka; we told this company's full business story in the real case study of ARUM, a Kanazawa town factory in Japan. This official product video, uploaded in 2021 — "ARUMCODE1 - NC プログラミング完全自動化 AI ソフトウェア" — has by now racked up nearly 16,000 views, and is the most direct first-hand material for understanding its product claims[1].
The video's core claim in one sentence: NC programming accounts for 50% of manufacturing cost in high-mix low-volume metal machining, and ARUMCODE1 uses AI to fully automate that segment, sharply cutting the prime cost of manufacturing[1].
02How to read "programming is 50% of cost"
The 50% is ARUM's marketing figure — no need to take it at face value, but the structure it points to deserves serious thought. High-mix low-volume job machining has a brutal arithmetic: the smaller the batch, the higher the "prep cost" allocated to each part. Make a thousand, and a few hours of programming spread across them is negligible; make one or two, and programming can cost more than the cutting itself. Manufacturing-systems theory treats this kind of "fixed prep work independent of batch size" as a determinant of the profitability of small-batch production[4].
So the correct reading of this figure isn't "every shop's programming is half its cost," but rather: the smaller your batches and the more varied your mix, the greater the financial leverage of programming automation. This also explains why ARUM's target customers are high-mix low-volume town factories — almost structurally identical to the situation of Taiwan's high-volume job shops (for the same structure on the quoting side, see the AI quoting video review).
03The most instructive part: writing out the boundaries
Most AI product videos give you only the prettiest demo; this video's description box, however, attaches a spec table that writes "what it can and can't do" out crystal clear[1]:
| Item | ARUMCODE1 base spec (official description box) |
|---|---|
| Supported machines | Vertical 3-axis machining centers (indexed 5-axis as future support) |
| Supported controllers | FANUC, Mitsubishi Electric (Siemens, Heidenhain as future support) |
| Registration limits | 50 machines, 1,000 tools |
| Machining target | 6-face-machined block stock (6F material), recommended max 200×200×50mm |
| Clamping condition | At least 5mm of vise grip |
This table reveals the engineering nature of "full automation": pin down the input space first, and only then can the rules run all the way through. Block stock, vise, 3-axis, specified controllers — each constraint eliminates ambiguity, giving every step of automation a definite answer. Academically, this is precisely the known boundary of feature-recognition and process-planning automation: within a well-defined feature space, recognition can reach very high accuracy[3], but once the planning solution space opens up (irregular parts, complex clamping, multiple operations), automation has to fall back to human-in-the-loop[2].
04The real meaning of "full automation" and lessons for Taiwan shops
Reading the video's claims and specs together, the conclusion is clear: "full automation" is a scope word, not a capability word. Within the world of 200×200×50 block stock it can be fully automated; step outside that world, and you're back to a division of labor between humans and AI. This isn't a shortcoming — fully automating the high-frequency, rule-based class of parts has enormous value in itself; the danger is mistaking "the full automation in the demo video" for "full automation of any part whatsoever."
For Taiwan shops, this video offers two evaluation moves you can use directly:
- Ask for the boundary table: when evaluating any AI programming/quoting tool, first ask for a spec table like this one — applicable geometry, material form, clamping, controllers, size limits. If they can't produce one, discount even the prettiest demo.
- Compare against your own part spectrum: classify the past six months of orders by geometric complexity, and work out the proportion that falls within the "fully automatable" scope — that's the ceiling on the automation dividend you can capture.
BestAI CAM's design philosophy is complementary to this: we don't claim "full automation," but rather full-pipeline coverage with a human at the key points — AI reads and models from the 2D drawing (the most common input on the Taiwan floor), generates code per the in-house tool crib and machine limits, enforces 3D cutting simulation, and runs an independent AI dimension cross-check; tolerances, datums, special methods and the go-to-machine decision are gatekept by professionals (see the complete guide to CNC automated programming). In-scope parts move fast, and out-of-scope parts still have an AI-accelerated prep chain available — a route that differs from "fully automated within a delineated scope" but is just as honest.
05FAQ
Does NC programming really account for 50% of manufacturing cost?
That's ARUM's claim in its official video, and it applies to the high-mix low-volume scenario it targets, not as a universal rule. The right way to read it: the smaller the batch, the higher the share of fixed prep costs like programming. Work it out with your own shop's data — multiply last month's programming hours per order by your labor cost, divide by the order value, and you'll know your real share.
How is ARUMCODE's "full automation" different from general AI CAM?
It trades strict input constraints (vertical 3-axis, 6F block stock, 200×200×50mm, at least 5mm of vise grip, specified controllers) for full automation within that scope. General AI CAM (including BestAI CAM) takes a human-in-the-loop path: broader applicable geometry, but keeping human review and gatekeeping. The two are a trade-off between "narrow and fully automated" and "broad with gatekeeping," not better or worse.
When evaluating an AI programming tool, what should you most ask the vendor for?
A boundary table: the supported geometry and material forms, clamping methods, machines and controllers, size range, and "what happens outside the scope" (rejection? downgrade to assist mode?). ARUM's official video description box is a good example. Once you have the table, compare it against your own order structure, work out the automatable proportion, and then talk about value.
Get notified when new articles and video walkthroughs go live — no inbox spam, one-click unsubscribe. (newsletter in Chinese)
KNOW THE BOUNDARY
Boundaries first, demo second — validate AI's value with your own part spectrum
Take a few representative drawings (both simple and hard) and run them through the AI read-model-generate-simulate-verify flow, and see the can and can't with your own eyes.
Talk to an onboarding consultant Training courses📋 Free download: the CNC Outsourcing Checklist (printable) (中文)
06References
- ARUM OFFICIAL (YouTube). ARUMCODE1 - NCプログラミング完全自動化AIソフトウェア (published 2021-09-29). youtube.com/watch?v=0tB-es3tY4w
- Xu, X., Wang, L., & Newman, S. T. (2011). Computer-aided process planning — A critical review of recent developments and future trends. International Journal of Computer Integrated Manufacturing, 24(1), 1–31.
- Zhang, Z., Jaiswal, P., & Rai, R. (2018). FeatureNet: Machining feature recognition based on 3D Convolutional Neural Network. Computer-Aided Design, 101, 12–22.
- Chryssolouris, G. (2006). Manufacturing Systems: Theory and Practice (2nd ed.). Springer.