From Pilot to Production: Scaling Industrial AI

On October 16, 2026, IIoT World published a Q&A with Neil Smith (Segment President, CPG, at Schneider Electric) and Andrew Waycott (President and Co-founder of TwinThread). They answer six questions that emerged from the "From AI Pilots to Profit" session at the Industrial AI Summit 2026. We draw here on the first five answers: the text available to us cuts off during the fifth, on safeguards, and we have not seen the sixth, on organizational change.

The business case that persuades the CFO

According to Waycott, CFOs look first at cost avoidance: downtime, scrap, inefficiency. TwinThread typically starts by reducing process variability. Smith adds a useful distinction: cost avoidance gets a project approved, while revenue growth gets the strategy funded.

Smith cites Schneider's 2026 study on AI in the CPG (consumer packaged goods) sector:

• average ROI from industrial AI today is 24.7% in the United States and 18.7% globally;
• for the most mature deployments, ROI is expected to reach 50–74% and beyond by 2030;
• only 13% of CPG manufacturers have AI integrated end to end, so the revenue potential remains largely a projection, not a measured result.

How much data you need to get started

Waycott argues that targeted problems beat broad overhauls. On an extrusion line, he explains, 10–30 parameters are often enough to significantly reduce start-up scrap. Smith offers a counterpoint, however: 36.3% of manufacturers cite a lack of contextualized operational data as an obstacle. For many plants, the first step is making data accessible, connected and trustworthy.

Scaling, sharing data, reducing false positives

For Smith, the missing link is a connective layer on top of legacy, proprietary and siloed data. The main barrier to data sharing between OEMs and operators is trust: the party running the plant wants to retain ownership and control. His examples include AVEVA CONNECT, with governed, time-limited access for an OEM, and software-defined automation, which decouples control software from hardware.

On false positives, his thesis is that they fall as context grows: a model that also knows the product, line speed and maintenance history is better at telling a real fault from normal variation. Operator feedback should flow continuously back into the model. Smith also cites an Omdia estimate, which comes from the vendor world: modernizing closed legacy automation could save a mid-sized company around $11.28 million a year, with hidden costs averaging 7.5% of annual revenue.

Safeguards and the hardest sectors

Waycott names government and heavily regulated sectors, such as pharmaceuticals, among the most difficult. Smith calls for AI outputs to be consistent, repeatable and explainable, and for cyber risk to be addressed first. He recommends an "AI-accelerated, human-led" model and adding AI, data and models to the risk register. The original question referred to a "Hugging Face incident" involving autonomous agents in production, but the article does not explain it and we cannot verify it here.

Key takeaways

Start narrow, curate data context, share data with governance, manage AI as an asset.

For engineering and IT leaders, the practical guidance is to choose a well-bounded problem with a measurable benefit, verify that the data carries the context it needs, define access rules for vendors, and treat models and data as components subject to risk control. One caveat: the figures come from a Schneider study focused on CPG, and both interviewees are vendors. They should therefore be used as a qualitative reference, not as a benchmark for your own plant.