Manufacturers are rapidly turning to edge AI to unify data, accelerate decision-making, and eliminate silos between IT and OT. As small language models and physical AI become viable at the process level, the rules of operational technology are being rewritten.
For decades, IT and OT have spoken different languages. One prioritises compute, the other control. But as AI moves closer to the production floor, convergence is no longer aspirational; it is operationally essential.
Paul Bloudoff, senior director at NTT DATA, sees a clear inflexion point. “We now have a fundamental need to act on both OT and IT data in tandem,” he explains. “Physical AI models do not just want this data; they demand it. And that is finally forcing the long-delayed integration of systems, standards and teams.”
Where earlier attempts to bridge the IT/OT divide were hampered by siloed data and organisational inertia, AI introduces a forcing function. Without unified access to sensor and enterprise data, models cannot infer, detect, or optimise in real time. For that reason, edge AI is pushing both infrastructure and culture to evolve.
This pressure is particularly acute in high-stakes environments where latency, cost and uptime are tightly interlinked. Whether in discrete manufacturing, chemical processing, automotive assembly or logistics, edge inferencing is becoming the only viable route to timely, contextual decision-making. Waiting for the cloud to respond is no longer a risk worth taking.
Small models for big decisions
The assumption that AI must be cloud-based and computationally expensive is rapidly falling apart. Large language models are not built for the factory floor. They are power-hungry, latency-prone, and often irrelevant in environments defined by physical parameters rather than natural language prompts.
Small language models designed for specific industrial tasks are proving far more effective. “When you are working with physical systems, you do not need billions of parameters,” Bloudoff explains. “Acceleration, vibration, movement patterns, these are mathematically simple phenomena. Small models can interpret them on a chip, in real-time, without relying on a data centre.”
The strength of these models lies in their minimalism. Unlike cloud-native LLMs designed for generality, small models are grounded in the physics of real-world motion. They use deterministic algorithms where appropriate, paired with minimal training data to interpret subtle changes in speed, direction, pressure or resistance. This makes them ideal for tasks such as predictive maintenance, process vision, or real-time quality inspection.
Hardware evolution also plays a pivotal role. Purpose-built neural processing units, ruggedised GPUs and low-power NPUs are now industrial-grade, compact and efficient enough to run continuous inferencing workloads on the plant floor. From machine vision for step validation to vibration analysis for gearbox diagnostics, these platforms bring AI to where the action is without compromising on speed, reliability or resilience.
This is not theoretical; these systems are already running in live production environments. They are spotting faulty components, alerting line operators to inconsistent torque levels and predicting when a conveyor belt will require recalibration. They are not replacing people. They are making their jobs more accurate, timely, and data-informed.
Infrastructure must meet intelligence
The shift toward edge AI introduces a cascade of decisions that many organisations are still underprepared to make. For inferencing to work at the process level, the underlying architecture must be redesigned to support low-latency data acquisition, contextualisation and routing.
That all starts with the data. “If sensor data is fragmented or poorly labelled, the model cannot act on it,” Bloudoff notes. “The first step in any edge AI deployment is mapping the data pipeline. Where are your sensors? How is the data transmitted? Can it be normalised and stored locally?”
However, it must be realised that the interoperability challenge is real. Legacy PLCs, proprietary SCADA systems, and vendor-specific protocols all present barriers to standardisation. But modern platforms are increasingly capable of extracting, translating and unifying this telemetry into clean datasets that can be ingested by AI models. Bloudoff cites advances in semantic tagging, digital twins and lightweight ontologies as critical enablers of this shift.
Connectivity is equally essential. Private 5G, Wi-Fi 6 and next-generation Ethernet are enabling a new level of data fluidity, bridging the gap between sensors, edge compute, local data lakes and cloud orchestration layers. Crucially, these connections must be resilient, secure and deterministic in their latency. Industrial networks cannot afford jitter, dropout or contention.
Once data is centralised, or at least indexed locally, inferencing can begin. The goal is not to replace existing systems but to augment them with faster, smarter, more adaptive decision-making capabilities. Instead of using AI as a last-mile analytics tool, it becomes a first-mile process optimiser, capturing anomalies as they emerge, not after they have disrupted throughput.
From alerts to action
Insight is only helpful if it is actionable. For frontline workers, information overload is a constant threat. Historical attempts at digital transformation have often resulted in a flood of alerts, dashboards and KPIs that confuse more than they clarify. Edge AI offers a different path. Models are now being trained not just to detect anomalies but to communicate their significance in plain language. Bloudoff points to predictive maintenance systems that generate prompts such as: “Unusual vibration detected in gearbox seven. Inspect during next scheduled downtime.” No graphs. No queries. This is just a direct, contextual recommendation.
That shift from data to dialogue is critical. AI becomes a peer to the operator, not just a passive observer. It is particularly effective in use cases like process step verification, where missed actions can result in cascading errors. A vision model can now track ten-step workflows and flag deviations in real time, giving workers the confidence that each product meets specifications without requiring post-hoc inspection.
There is also a growing appetite for conversational AI interfaces within industrial contexts. Operators can now query systems verbally or via text, asking how long a machine can run before maintenance is required or whether current throughput will meet the day’s targets. These interactions are not gimmicks. They are practical tools that improve situational awareness, reduce training overhead and capture institutional knowledge in environments where skilled labour is scarce.
Cultural shifts and trusted systems
While the technical barriers to edge AI are falling rapidly, organisational inertia remains one of the most persistent obstacles. “Many manufacturing firms still treat AI as an IT-led experiment rather than an operational imperative,” Bloudoff continues. “This disconnect manifests in project silos, lack of cross-functional governance and scepticism from shop-floor teams.
“Building trust requires more than pilot projects. Workers should not need to believe in AI. They should see it work. That means delivering systems that are explainable, predictable and, above all, reliable. If a recommendation saves time or prevents a fault, trust will follow.”
Physical AI offers a natural advantage in this regard. Unlike general-purpose LLMs that often hallucinate, these models operate in a closed loop with well-defined parameters. They are focused, narrow and deterministic. This makes them ideal for environments with stringent regulatory requirements, safety standards, or zero-tolerance for error.
Transparency is also key. Every decision made by the model, whether to alert, suppress or recommend, must be auditable. Organisations deploying decentralised inferencing must build governance frameworks that track model behaviour, data lineage and performance over time. This is not just for compliance. It is for resilience. When something goes wrong, you need to know why and how to fix it.
From POC to production at scale
Many organisations struggle to move from isolated proof-of-concepts to fully scaled AI deployments. Edge AI amplifies this challenge. Each site may have different machines, data schemas, workflows and operating constraints. Standardisation is possible, but it requires abstraction, automation and repeatability.
“Perfect standardisation is unlikely in the short term,” Bloudoff says. “What matters is building a repeatable delivery framework that allows use cases to scale without starting from scratch each time. This means decoupling data models from hardware, creating modular AI components, and investing in orchestration layers that can manage deployments across sites, regions and environments.”
Bloudoff believes that agents may play a key role in this evolution. “As multi-agent architectures mature, they will enable AI workloads to be intelligently routed between edge and cloud, based on policy, priority and availability,” he says. “These agents can act as translators, coordinators and validators, ensuring that every inference, regardless of where it happens, aligns with the broader system goals.”
Industrial AI is becoming physical
Bloudoff expects the next wave of transformation to be led not by cloud AI but by physical AI, models that understand motion, temperature, force and form. From robotics to safety compliance, these models offer a way to digitise the real world in a way that is fast, local and precise.
“We are entering a phase where AI does not just analyse data,” Bloudoff concludes. “It senses, sees and acts. And for manufacturers, that opens the door to a smarter, more responsive way of working, one that is built from the ground up, not the cloud down.”
Executives still debating AI strategy would do well to start at the edge. That is where the next industrial revolution is already being quietly, incrementally deployed, not in the headlines, but in the heartbeat of every process.
