At Industrial Wireless Forum, industry leaders explored how agentic AI running on private 5G is transforming factories and plants. The discussion highlighted not just the technology, but the architectural, connectivity, and data governance foundations that are essential to turn AI potential into real operational value on the shop floor.
In sum – what to know:
Data first, AI second – panelists emphasized that clean, high-quality data is the starting point for any agentic AI deployment. Without it, even the most advanced models fail to deliver actionable insights.
Integrated edge networks – deterministic private 5G and multi-network orchestration are critical to deliver AI decision-making close to where processes happen, reducing latency and enabling real-time action.
Trust and federation – factories must balance data centralization and distribution, ensuring privacy, sovereignty, and security while allowing AI to operate across multiple layers and sites, optimising workflow dynamically.
There was a good panel discussion at Industrial Wireless Forum a couple of weeks back about agentic AI on private 5G in Industry 4.0 – to bundle together a bunch of tech buzzwords, but also describe what is being put to work in factories and plants in the name of digital transformation. Jason Wallin, in charge of industrial networking at US machinery maker John Deere, summed it up very well, when asked about the technological pre-requisites to solve Industry 4.0 chokepoints and force-multipliers. “There’s three layers,” said Wallin. “The first is the quality of the data, right? If you start [any AI] journey with questionable data, you’re going to get questionable results, and no better.”
Really, the whole discussion could stop there; there is enough work to do in the industrial realm just to get the data right, it seems. Also on the panel, Mike Carroll, formerly with Georgia Pacific, now a research fellow at LNS Research, said: “Most industrial data is not in great shape. Predictive analytics is not reliably predictive. And when you get a prediction, the capability to do something with it is not effective… You get something that’s about a quarter effective. So how do you solve that? And it turns out agentic AI is one of the ways.” This was the setup on the panel; a response to the first question from host Rosalyn Craven, leading private networks research at STL Partners.
But we interrupt Wallin, somewhere between layer one and two; here is the rest of his explanation about how it all fits together, in service of new agentic AI applications (which are, according to Carroll, the mechanism to solve issues with data quality, as well). Wallin said: “[The point is to have] clean data and an accessible data lake to make that RAG request… The other part that can be problematic is the compute layer above, where you run your inferencing models and the LLMs… There are lots of [inferencing] cases you can use the public cloud for, but when the tolerance is tight, and the tuning is constant and changeable, then the processing needs to move very close to the edge.”
He went on: “Many use cases are compute-bound, [which throws the] process time to run through the RAG models… The other piece is the network. Deterministic cellular connectivity… is a key for those tight tolerance cases, and to be able to deliver that AI value immediately back to stakeholders on the shop floor.” But there is a bigger challenge, first, noted Wallin, who was also representing the 5G OT Alliance (“a bunch of discrete manufacturers that build things and use private networks to support their manufacturing operations”) on the panel. “Technology for the sake of technology is cool, but understanding the business value you want to extract from a process [is the key],” he said.
But for our purposes, in this article, that initial use-case focus is a given; the panel was convened to talk about the digital nuts-and-bolts of running agentic AI on private 5G in factories and plants. Carroll suggested the discipline is as much a cultural one, but also echoed Wallin. He said: “The most important thing is not the technology, because that exists. The most important thing is us – and how we apply it relative to the problem. If you take a position that reasoning at the edge is going to happen – which it will – then the question is how to make it work at the edge. [It] requires connectivity to drive decision-velocity. Latency is the biggest tax that never shows on the balance sheet.”
And he went on to explain how AI software agents, if trusted to do the work, can help enterprises smash old data silos. “Information goes up, decisions come down… The problem is the walls are transparent, but the information is opaque – because it’s trapped behind the organisational structure. To get at it, you need the architecture… to trust the agent, which is trying to shape the workflow.” As it stands, people have to approve every step, which slows decisions. Agentic AI and private 5G fix this by creating an architecture where machines can safely make routine decisions on their own, using deterministic connectivity at the edge.
Carroll said: “That’s the architecture that matters. Instead of gating every permission, because you only trust a human to make a decision… [you deploy software] that’s built to do it, with guardrails for intent. That is the problem we’re solving – where the connectivity allows the reasoning at the edge, based on an architecture of trust to manage how the reasoning actually functions.” The network lets the data flow, the compute draws the insights, and the AI makes the decisions. US private 5G vendor Celona, also on the panel, flipped the discussion the other way, momentarily, to put focus on AI for networking, rather than just networking for AI.
Puneet Shetty, head of product and field engineering at the firm, said: “[Our] Celona Orion [product] is an agentic AI platform… to make private 5G operations as autonomous as possible. The goal is to continuously monitor how devices are doing, what the experience is, and how the devices are interacting with applications – and then to look at all of these KPIs, learn from the patterns, and act on the data. So if a new device comes on the network and the workload shifts, Orion can allocate spectrum, adjust the quality-of-service, and isolate issues that could impact performance.” But Celona knows very well about the challenges on the floor side, too, which 5G can help with.
Shetty said: “[The network is often] an afterthought. We have had customers assume that the right connectivity is already there… And then when they actually start to deploy [AI applications], they can’t take advantage of them… The only way to upload data has been to go back to the office where there is connectivity. Where they needed it, they did not have it. [The same where] a manufacturing customer had these AI applications that required a certain level of [performance] that their current networks could not deliver – because they were oversubscribed or just not designed for it. So connectivity is often an afterthought. For AI to work, deterministic connectivity is non-negotiable.”
He added: “Agentic systems need these real-time observe/act/learn loops [and] private 5G networks are going to be foundational to this. The other critical part, often an afterthought, is how to get data from these legacy systems on which they are deploying AI applications – so data integration with OT systems and IoT sensors, and digital twins. Because situational awareness will be very important for AI. So those are two key things that come to my mind. Connectivity and the ability for data integration with these legacy systems.”
Also on the panel, Italian firm Adeptic Reply, part of Turin-headquartered consulting and integration firm Reply, proposed that agentic AI in Industry 4.0 is moving toward a self-optimising architecture, where intelligence is dynamically positioned across the network. Raffaele Cestari, AI specialist at the company, said: “AI can be used in a recursive way… [to] design algorithms that can help to understand the optimal place to deploy [AI] algorithms.” In other words, AI doesn’t just execute tasks, but helps to determine where other AI agents should reside and run – based on latency, data availability, compute resources, and task criticality.
This creates a layered system where “each layer is optimized… and the output of a layer is itself an optimal solution” – producing an architecture that continuously reinforces its own performance. Adeptic Reply is engaged in a couple of linked EU projects IPCEI-CIS and 8RA, to orchestrate complex multi-technology networks for the most demanding industrial environments (more here). This idea of recursive intelligence only works if the underlying connectivity is equally flexible and deterministic. Multi-network orchestration – spanning private 5G, Wi-Fi, wired infrastructure, plus edge compute infrastructure – becomes the fabric that agentic systems rely on.
Agents need to access data across machines and locations without hitting silos or bottlenecks. Decisions must be taken closer to the action, with minimal delay – as per the comments from Wallin, Carroll, and Shetty on the panel. European telcos – with embedded urban infrastructure, following European rules about data sovereignty, privacy, security – should be tapped for distributed edge data centres, orchestrated with campus edge sites, as part of a federated approach where data remains protected, but models and insights can still be shared. The tension is between centralisation and distribution of the storage and compute functions.
Cestari said: “If data is managed in a centralized manner, we have to protect just one place… but if we have issues with that place, then we might lose our data.” Federation mitigates this risk, but introduces challenges with the data transfer, where encryption and protection are non-negotiable. “Without these standards, we cannot have any solution that is based on AI… everything else becomes useless,” he said. The point is that agentic AI requires not just intelligent algorithms, but a trusted, compliant architectural foundation that lets reasoning happen at the edge while keeping data secure, sovereign, and under strict governance.
