From Pilot to Production: Industrial AI and the Case for the CFO

On 16 October 2026, IIoT World published a Q&A in which Neil Smith (Segment President CPG at Schneider Electric) and Andrew Waycott (President and co-founder of TwinThread) answer six audience questions. The questions come from the session "From AI Pilots to Profit: Unlocking Manufacturing's Hidden Data Advantage" at the Industrial AI Summit 2026. Keep in mind that the speakers are vendor representatives, so their figures should be read with that caution. Note: this summary is based on the portion of the article available, which cuts off during the final answer, on cybersecurity safeguards; the sixth question, on organizational change, was not reviewed.

The business case: costs first, then revenue

According to Waycott, the first argument is usually cost avoidance, because CFOs already know their losses from downtime, scrap and inefficiency. TwinThread typically starts by reducing process variability. Smith adds that both levers are needed:

"Cost avoidance gets the project approved. Revenue growth gets the strategy funded."

Smith cites Schneider's 2026 study on AI in the CPG sector: average ROI from industrial AI is 24.7% in the United States and 18.7% globally, with expectations of 50–74%+ by 2030 for the most mature implementations. However, only 13% of CPG manufacturers have AI integrated end to end, so the revenue potential remains largely a projection. Moreover, the study comes from the vendor itself.

Data: less than you might think, but with context

Waycott argues that a pilot needs less data than expected: on an extrusion line, 10–30 parameters are often enough to significantly reduce start-up scrap. He prefers targeted problems to large overhauls. Smith notes that 36.3% of manufacturers cite a lack of contextualized operational data as an obstacle, and observes that "industrial AI without industrial intelligence is simply sophisticated guesswork".

Scaling: trust, shared data and false positives

For Smith, the main barrier to scaling is trust: operators want OEM expertise without losing ownership of their data. He gives the example of AVEVA CONNECT, which gives the OEM governed access to specific machine data, with the operator deciding who sees what and for how long. He also advocates open, software-defined automation that decouples control software from hardware vendors. Finally, he cites an Omdia study (a vendor-linked source, to be treated with caution) according to which modernizing closed legacy systems would save an average company $11.28 million a year.

On false positives, the answer is context: a vibration spike alone triggers an alarm, whereas a model that also knows the product, line speed and maintenance history can tell a real fault from normal variation. Continuous operator feedback turns the pilot into a reliable tool.

Difficult sectors and agents in production

Waycott names the public sector and highly regulated industries, such as pharmaceuticals, as the hardest, because of bureaucracy. On autonomous agents, Smith calls for consistent, repeatable and explainable outputs, with cyber risks addressed first. He proposes an "AI-accelerated, human-led" model and adding AI, data and models to the risk register. I could not verify what he says about IEC 62443 and the EU Cyber Resilience Act, nor the details of the so-called "Hugging Face incident" mentioned in the question.

Key takeaways

For integrators and plant engineers, the practical points are:

• choose a narrowly scoped pilot with a measurable problem such as scrap or variability;
• present the CFO with avoided cost first, then the revenue scenario, stating that it is an estimate;
• contextualize data (product, speed, maintenance) to reduce false positives;
• define governed data-access rules with OEMs;
• maintain human oversight and include AI in the risk analysis.

The topic overlaps with data readiness and the contextualization gap already covered on this site; here the focus is on the move from pilot to production and on the business case.