If 90% of operational truth lives in documents that AI cannot read, how do manufacturers move from insight to action? This AI Manufacturing Day 2026 panel tackled the brownfield reality: plants never designed for AI, critical data trapped in unstructured formats, and why 60% of agentic AI projects are projected to be abandoned this year.
Hamish Mackenzie (IIoT World) moderated a discussion with Chris Huff and Anthony Vigliotti (Adlib Software) and Mathias Oppelt (Siemens). The panel dissected where AI stalls in industrial environments and what it takes to build a trust layer for agentic workflows.
Mathias Oppelt explained why general purpose AI falls short in industrial settings. Large language models trained to generate plausible text are not sufficient for industry, where reliability must be proven and a wrong answer can cause physical harm. Industrial AI must speak the language of engineers: understanding drawings, simulation models, and technical context beyond words. He shared a recent example where an 86-year-old engineer had to be hired back to resolve a line shutdown that cost $1.5 million over four weeks, because no current staff had the legacy system knowledge.
Chris Huff outlined the shift from systems of insight to systems of action. Agentic AI is real today for high-volume, lower-stakes work like procurement, contract review, and supplier onboarding. It is not yet real for high-stakes, low-volume decisions where a wrong answer means a recall or safety event. He identified three places where pilots consistently stall: ingestion of messy multi-format documents, validation and traceability back to source, and integration into PLM, ERP, and QMS systems.
Anthony Vigliotti proposed treating manufacturing documents as evidence, applying standards of relevance, reliability, authenticity, and chain of custody. He highlighted a problem many overlook: AI systems forced to clip relevant information due to context window overruns, leading to incomplete outputs that drive real operational consequences when acting on those outputs.
This session was sponsored by Adlib Software. Editorially independent.
