What Industrial AI Controls Today

A cement manufacturer using AI to optimize a flame that changes in fractions of seconds cannot wait for data to travel to a cloud server and back. This gap was a key topic covered at the Industrial AI Summit 2026, where David Purón of Barbara identified three factors that determine whether AI runs at the edge or in the cloud, while Jonathan Alexander of Albemarle Corporation described four levels of human-AI interaction that define how much autonomy manufacturers should grant to AI systems. At the event, Kudzai Manditereza of HiveMQ explained why agentic AI remains limited to handling operational exceptions rather than closing control loops, and Peter Sorowka of Cybus described why organizational transformation determines whether industrial AI scales beyond a single plant.

In this piece, we explore how latency, data volume and reliability decide whether industrial AI runs at the edge or in the cloud, why classic AI still outperforms agentic AI for closed-loop control, a four-level framework for human-AI interaction in manufacturing, and why scaling AI across plants is an organizational challenge more than a technical one. 

What Three Factors Determine Whether AI Runs at the Edge or in the Cloud?

Most industrial AI projects today use hybrid architectures that split processing between edge devices and cloud infrastructure, and three factors (latency, data volume, and reliability) determine how that split works in practice. A process where conditions change in fractions of seconds requires AI inference at the edge because the round trip to a cloud server introduces delays that exceed the reaction window. High-frequency sensor streams from cameras, vibration monitors, and process instruments can overwhelm network bandwidth if sent raw to the cloud, so edge processing filters and compresses before transmission. And a plant that loses internet connectivity cannot afford to lose AI capability at the same moment, so critical decision-making must function independently of external networks.

A cement manufacturer automating flame control operates at the extreme edge of this spectrum, where the AI model runs directly on the production line because the process physics do not tolerate network latency. An energy optimization model analyzing weekly consumption patterns sits at the opposite end, where cloud processing is both practical and cost-effective. Most manufacturing AI deployments fall somewhere between these extremes, with edge handling time-critical inference and cloud handling model training, long-horizon analytics, and cross-facility comparisons.

When Does Classic AI Outperform Agentic AI?

Classic AI models, including regression, classification, and traditional machine learning, deliver the majority of measurable production value in manufacturing today. A water treatment company called Acciona uses classic AI to predict chemical dosing requirements and generates more than $250,000 in savings per plant. These models work because the problems they solve are well-defined, the input data is structured, and the physics of the process are understood well enough to train reliable predictions.

Agentic AI occupies a different layer of the stack, one suited for handling operational exceptions rather than closing control loops. When an unexpected deviation occurs that falls outside the classic model’s training data, an agentic system can interpret the situation, consult documentation or historical records, and recommend an escalation path. Placing agentic AI inside the actual control loop on the shop floor introduces reliability risks that manufacturing environments are not prepared to accept. Large language models will not be embedded in control systems in the near term because production equipment needs to run continuously, and LLM inference introduces variability that process control cannot tolerate.

Manufacturers evaluating AI vendors should ask whether the solution uses classic models for the core prediction and optimization tasks, where accuracy is provable and deterministic checks are possible, and reserves agentic or LLM-based components for advisory functions, exception handling, and decision support where a wrong answer does not stop a production line.

What Are the Four Levels of Human-AI Interaction in Manufacturing?

Peter Sorowka of Cybus compared AI autonomy in manufacturing to autonomous driving, where agreed-upon levels increase gradually as organizations build experience and trust. Jonathan Alexander of Albemarle Corporation talked about four manufacturing-specific levels, where risk tolerance and regulatory context determine which level fits each operation.

Most manufacturing operations today sit at Level 1 or Level 2, where AI provides information or recommendations but a human makes the final call. Moving to Level 3 requires the kind of calibrated confidence and auditability that current industrial AI platforms are still building toward. Level 4 remains largely theoretical for manufacturing, where the consequences of autonomous AI errors carry material, safety, and regulatory costs that current systems cannot fully mitigate.

Manufacturers can use this framework to specify what they want from an AI system rather than accepting a vendor’s default level of autonomy. An organization that needs Level 2 but receives a Level 3 system has an accountability gap. An organization paying for Level 3 but getting Level 1 has a functionality gap.

What Measurable Outcomes Can Manufacturers Expect from Industrial AI?

Companies already running industrial AI in production report five to six percent improvements across different KPIs. A Deloitte report cited during the panel identified a broader opportunity of twenty percent improvement on existing KPIs when AI is applied systematically across operations rather than in isolated pilots.

Scaling from one successful plant to multiple facilities is where the outcome gap widens. A food and beverage manufacturer called GB Foods completed the engineering work in its first plant very quickly and encountered different issues at each subsequent site, including variations in equipment configuration, data availability, and local operating practices. The pattern is common: first-plant success creates expectations that the deployment will simply replicate, and each facility introduces site-specific variables that require additional engineering and adaptation.

German manufacturing COOs are planning for a workforce scenario where they need the same production output with half the number of people by 2030, driven by demographic shifts and labor market tightening rather than cost reduction targets. The technology, surprisingly, is not the hard part.  “It’s changing the organization to be digital and AI native, and this is the hard part,” said Peter Sorowka, CEO of Cybus. Organizations that treat AI as a technology purchase rather than an organizational transformation underestimate the change management, skills development, and cultural adaptation that separate a working pilot from a scaled production system.