Generative AI on the Factory Floor: Where It Works and Where It Doesn't

Generative AI captured the world's imagination with chatbots and image generators. On the factory floor, the technology is finding a narrower but potentially more valuable set of applications — along with some hard limits that vendors would rather not discuss.

Where it works:

Maintenance documentation is the breakout use case. Large language models fine-tuned on equipment manuals, repair logs, and engineering specifications are becoming the first stop for maintenance technicians diagnosing unfamiliar faults. Dow Chemical reports that its internal maintenance assistant resolves 40% of technician queries without escalation to senior engineers, saving an estimated 12,000 hours annually.

Design iteration is another strong fit. Generative design tools that produce optimized component geometries for additive manufacturing are cutting prototyping cycles from weeks to days. Autodesk's generative tools, now integrated with industrial simulation software, are seeing adoption across automotive and aerospace suppliers.

Training content creation rounds out the top three. Generating standard operating procedures, safety briefings, and onboarding materials from existing documentation saves technical writing teams hundreds of hours while keeping content current as processes change.

Where it doesn't:

Real-time process control remains firmly off-limits for generative models. The stochastic nature of LLMs — the same input can produce different outputs — is fundamentally incompatible with deterministic control requirements. No responsible engineer is putting a language model in a PLC control loop.

Quality certification is another boundary. Regulatory frameworks like ISO 9001 and FDA 21 CFR Part 11 require traceable, reproducible decision-making. Generative AI's probabilistic outputs don't meet that bar, and likely won't for years.

The pragmatic approach: use generative AI for human-in-the-loop tasks where speed matters more than determinism, and keep traditional ML and rule-based systems for anything that touches product quality or safety. The boundary is clear, and the companies respecting it are the ones getting real value.