Industrial AI: Five Conditions to Meet Before the Payoff
Velotic CEO Brian Shepherd outlines five conditions that must be met for AI on the factory floor to deliver value. The main bottleneck remains the data layer, not the technology.
On November 17, 2026, IIoT World published an article based on a video interview with Brian Shepherd, CEO of Velotic, and on his white paper "The Journey to Value in Industrial AI". The document comes from a vendor, so its claims should be read as the vendor's point of view. It still offers a useful checklist for anyone evaluating AI projects in a plant.
Physical AI: sense, think, act
Shepherd uses the term "physical AI" for artificial intelligence that operates at plant level. It works in three stages: sensing (collecting data from sensors, machines and control systems), thinking (integrating and contextualizing that data) and acting (making deterministic changes to the process). A problem in any one stage stalls the entire implementation.
The five conditions
According to the white paper, manufacturers need to meet these requirements:
- choose the right tool for the specific operational problem;
- build a solid foundation of structured, contextualized data;
- deploy AI at the right place and the right cost, targeting bottlenecks with a clear financial return;
- build trust in from the start, so engineers can trace and verify every AI decision;
- choose between autonomous and augmented operations.
To assess maturity, Velotic proposes the "Value Cube", a three-dimensional model in which the weakest condition sets the pace for the whole organization.
Data is the bottleneck
For Shepherd, the biggest obstacle is the data foundation. Plants collect large volumes of machine data, but without context AI cannot turn it into decisions. Anyone working with SCADA, historians and the OPC layer knows this well: tags, units, machine states and links to orders and products must be modeled before any algorithm is applied.
Autonomy versus augmented operations
Fully autonomous operations remain a long-term goal. The augmented approach, with copilots, assistants and real-time recommendations that keep experienced operators in the loop, delivers value faster. According to Shepherd, in the coming years most companies will focus on putting AI tools in the hands of trusted frontline workers.
The cases cited
Church & Dwight reportedly achieved a 6% OEE improvement; Papa John's reportedly cut quality audits across 11 centers from a full week to about three clicks. Both combined contextualized machine data, worker input and change management, and both started small before scaling up. The article gives no further detail, and the OEE figure is a single number reported by the vendor, not independently verified.
"Most of the time, this is not about a technology challenge… It's about an organizational change management problem."
Background: the launch of Velotic
Velotic was launched in March 2026 and combines GE Vernova's Proficy business with PTC's ThingWorx and Kepware. Its "Industrial Intelligence Loop" pairs Kepware for data collection, Proficy for HMI/SCADA, MES and historian, and ThingWorx for applications, analytics and AI workflows. A recent ThingWorx release introduced agentic AI; similar capabilities are expected in Proficy MES and SCADA by the end of 2026.
What to do in practice
Before launching a project, it pays to assess each of the five conditions honestly and start with the weakest, almost always data contextualization. Then pick a bottleneck with a measurable return, adopt an operator-in-the-loop approach and make sure every recommendation is traceable. Even though it comes from a vendor, the underlying message holds up: success depends on the organization and the data more than on the model.