WEST Session: This presentation challenges the prevailing narrative that manufacturing must conform to AI, proposing instead that AI must evolve to meet manufacturing's unique demands. After over a decade of attempting to transplant cloud-designed AI models into manufacturing environments, it has become clear that this approach is not practical. Rather than continuing to quotes percentages of failures, we advocate for a fundamental shift in perspective. If AI's core strength lies in pattern recognition and learning, why not leverage this capability to make AI itself more adaptable to manufacturing contexts? This talk demonstrates how AI can be redesigned to thrive in manufacturing environments through concrete examples that accelerate the development of robust, continuously learning models. We examine three critical assumptions that, when reconsidered, significantly enhance AI adoption and scalability in manufacturing. First, we start with quantifying success. Time invested in understanding and quantifying the trade-offs that matter to a production line is invariably worthwhile. Consider quality control as an example: should you prioritize developing a model that catches every defect, or one that minimizes false positives by avoiding the misclassification of good products as defective? Like human decision-making, AI systems will inevitably make errors—the key is to design systems that account for and manage these errors rather than pretending they won't occur. Second, we tackle data strategy. Manufacturing data represents valuable intellectual property that demands strategic handling. Contrary to popular belief, more data doesn't always yield better results. Our experience shows that indiscriminate data usage often produces sluggish, costly models that are challenging to troubleshoot and maintain. Hence, data selection strategies play a crucial role in the long-term success of a solution. Finally, we emphasize AI's inherently non-deterministic nature. Treating AI as a deterministic tool fundamentally limits its adaptive potential. Instead of rebuilding AI systems with every product change, we propose designing solutions that inherently evolve with environmental shifts—both incremental and substantial. This approach positions AI as a dynamic partner in manufacturing, capable of continuous learning and adaptation rather than a rigid tool requiring costly reconfiguration.
