Updated September 2026.
Manufacturers are not training their own AI models. What is actually happening on the shop floor and in the quoting office is narrower and more concrete: AI-assisted feature recognition on a drawing, a suggested cycle time, a proposed process route, a first-pass cost estimate. Small, specific decisions that used to take an estimator or a manufacturing engineer an hour, now proposed by software in seconds.
That is useful. It is also where trust breaks if the system is treated like a normal software dependency.
The EU AI Act is part of the conversation, but it does not automatically apply the same way to every manufacturer or every use of AI. Whether a given system counts as high-risk, and what obligations follow, depends on the specific use case. What does apply everywhere, regardless of legal classification, is a simpler standard: an AI-generated manufacturing decision should stay reviewable, and a human should remain responsible for it.
Why engineering data is not ordinary input
A technical drawing, a STEP file, a customer's tolerance requirements and a factory's real cutting data are not generic text. They carry customer IP, defence- or medical-grade specifications, and years of accumulated cost knowledge. Feeding that data into an AI system, or generating manufacturing decisions from it, is not the same risk category as summarizing an email.
Any AI system that touches this data needs to be explainable, measurable and controllable before it is allowed near a quote, a routing, or a setup sheet.
Explain it
Can you show why the system proposed this cycle time, this process, or this cost, for this part revision?
Measure it
Can you tell when a suggestion's accuracy is drifting against actual production outcomes, before it reaches a customer quote?
Control it
Can an estimator or manufacturing engineer override the suggestion, with a documented reason, and have that correction feed back into the system?
Provenance is not optional
Every AI-suggested manufacturing value needs to carry where it came from: which part revision it was calculated against, which model and version produced it, which factory library or historical actuals it drew on, and how confident the system is. Without that provenance, a suggestion is just an unverifiable number that happens to look precise.
Humans stay responsible for consequential decisions
A quoted price, a chosen manufacturing process, a committed tolerance: these carry real commercial and engineering consequences. AI can propose them. A person with the relevant expertise reviews, corrects if needed, and remains accountable for what goes out the door.
Responsible deployment is architecture, not a compliance add-on
Explainability, provenance and human review are hard to retrofit after a system ships. They need to be part of how the system is built from the first version, not a policy document attached afterward. That does not mean every manufacturer needs an AI Act compliance program. It means the underlying discipline, knowing what the system decided, why, and who can override it, is worth building regardless of which regulatory bucket a given use case falls into.
Subduxion is building a world model for precision manufacturing. Our research focuses on how geometry, requirements, manufacturing processes and production outcomes can be represented and learned computationally, with provenance and human review built into the architecture from the start. Read more about how we think about this in security, or see it applied in Blake.