Why AI Pilots Fail in Manufacturing (and How to Fix It) with Russell Halper

AI pilots look great in a controlled environment. Then they hit the real shop floor and quietly fall apart. In this episode, Lisa Ryan talks with Russell Halper, founder and managing director of Insight Kitchen, about why so many manufacturing AI initiatives stall after deployment and what leaders can do differently before they ever write the first prompt.

Russell's path here is unusual. He started a PhD in mathematics under a founder of chaos theory, moved into operations research (one of the earliest forms of data science), worked at UPS, then went in-house as an internal consultant for a major consumer goods manufacturer. From there, he spent eight years at a boutique Palo Alto consulting firm, helped grow it from 16 to 100 people, saw it acquired by a large global consultancy, and ran its West Coast Generative AI practice. Two years ago, he left to found Insight Kitchen, a boutique data science and AI consultancy focused on operational AI in manufacturing and supply chain.

Key Takeaways for Manufacturing LeadersA pilot is not a proof point; it is a rehearsal for scale. Most organizations design pilots to prove AI works in a controlled setting. Russell argues the real work is designing the pilot to surface the change management, data, and complexity issues that will show up at full scale, before you ever roll it out plant-wide.

A 95 percent solution can be a catastrophic failure. In operational environments, the last few percentage points of accuracy often carry the most real-world risk. Before adopting an AI tool, define what "good enough" actually means for that specific decision, because the threshold changes by use case.

Watch for compounding error, not just error rate. If an AI model is 99 percent as good as your best people but making decisions much faster, small errors can compound quickly. Ask whether faster, slightly less accurate decisions are actually creating value or just creating a bigger mess sooner.

Usage is your earliest warning sign. If the people on the floor are not acting on the AI's recommendations, you already have your answer. A model nobody trusts is not a decision-making tool, no matter how sophisticated it is.

Your senior operators know things your data never captured. A 40-year machine operator who says "something's off with this machine" is picking up on signals that never made it into a sensor feed. When leadership overrides that instinct because "the dashboard is green," they lose the tribal knowledge as retiring workers w