The Opportunity and Limits of AI in Product Development

What AI is good at: it generates many variants from a draft and makes comparisons possible for which there is otherwise no time in everyday project work. It reviews standards and datasheets, summarises research, delivers first drafts for the technical documentation and takes routine work off your hands. Used correctly, it saves considerable working time.

However, one should not make the mistake of leaving the content unchecked. Although the agents can take over the work of employees, they are not employees themselves. A language model all too often produces plausible-sounding text – but it may then lack truthfulness.

What AI cannot do

AI bears no responsibility: it is neither accountable for safety nor does it stand in for the person who will later work on the machine. It does not resolve conflicting objectives, because it never fully knows the real context; what it delivers is a suggestion without awareness of the consequences.

For many tasks that build on finished 3D models – renderings, colour variations and visualisations – AI is now very well suited. For drawing derivations it supports initial views, but dimensioning and tolerancing remain the designer's task. For creating functional 3D models, however, it is so far not suitable. This is partly because today's models are trained and controlled mainly with language and images, but hardly at all with parametric CAD data. Complex geometries with boundary conditions, tolerances and manufacturing requirements cannot be reliably described this way. This is where a skilled designer will "still" be needed.

Our model

We keep the roles separate: the machine delivers suggestions and drafts, the engineer is responsible for the result. Every suggestion therefore goes through a technical review before it becomes part of a product – this costs some of the time saved earlier; in our projects, a clear net saving nevertheless remains.

---

Sources

1. Xu, Jain, Kankanhalli (2024): Hallucination is Inevitable – formal proof that language models, as general problem solvers, inevitably hallucinate; the error cannot be fully eliminated.
2. Chen et al. (2025): Understanding Design Fixation in Generative AI – generative AI shows design fixation and narrows the solution space, similar to human mental blocks.
3. Vicente & Matute (2023): Humans inherit artificial intelligence biases – Scientific Reports: humans adopt the errors of an AI system and retain them even after the support is gone.