Generative AI is rapidly finding its place in industry, from supporting engineers to analysing measurement data. Yet the technology runs into a fundamental problem: it delivers probable outcomes, while manufacturing processes require unambiguous, hard limits.
Probability versus certainty
Models built on statistics return the most likely result. That works well for text or images, but on the factory floor an average often means nothing. A part either meets the tolerance or it does not. A temperature either stays within the safe range or it exceeds it. There is no room for "roughly correct".
It is precisely this vagueness that makes uncritical use risky. An answer that sounds convincing but falls just outside the specification can lead to scrap, rework, or in the worst case to unsafe conditions. For quality assurance and certification, a result that is "almost right" is effectively wrong.
What this means for manufacturers
Companies are wise to deploy generative AI where variation is acceptable, and to avoid it where strict limits apply. Suitable applications include:
Supporting tasks: drafting documentation, summarising reports, or generating ideas.
Preparatory work: quickly searching through data before a deterministic check delivers the final verdict.
Critical decisions: here, fixed rules, measurement systems and human oversight remain decisive.
The key is that AI outputs must always remain testable against objective criteria. By combining generative models with verifiable rules and measurable limits, manufacturers can harness the speed of the technology without giving up the certainty their operations demand.
Generative AI is therefore not a replacement for precise process control, but a tool that must be carefully embedded. The challenge for engineers is not whether to use the technology, but where and with what safeguards.