In this episode of our podcast, David sits down (again 😀) with Aron Semle, CTO at HighByte, for a follow-up conversation roughly a year after our first podcast together. A lot has changed in a year. When we last spoke, Industrial DataOps was still a concept many people needed convincing about. AI was already making waves, but the link between data foundations and AI readiness wasn’t as obvious to most manufacturing organisations as it is today. So where are we now? And where is the hype outpacing reality? You can listen to the episode, or get the main ideas in this article! Let’s get into it. We Don’t Sell the Problem Anymore The market conversation has shifted in the last year. “We don’t sell the problem anymore,” Aron says. “People understand the problem. We have customers and prospects coming to us saying: we know we need to clean up this data to get AI-ready.” That’s a big shift. And it mirrors what we see at the IT/OT Insider as well. During the peak of the LLM hype, everyone wanted AI but nobody wanted to talk about foundations. Now? People come to us and say: we want AI, and therefore we need to build a foundation. That loop is finally closing. And AI is the catalyst. Sebastian from Frost & Sullivan, who presented at HighByte’s DataOps Days late last year, made the argument that AI and DataOps are coupled — hand in hand — and showed the market growing accordingly. He’s right. Thanks for reading The IT/OT Insider! Subscribe for free to receive our posts and support our work. Starting from Where You Are So what does that foundation look like in practice? Aron breaks it down using a digital maturity lens: If you’ve got equipment with nothing connected and everything is manual, you’re not ready for AI on the factory floor. You need connectivity first — HMIs, SCADA, historians. If you’re at that level, you’re probably ready for the next step: Industrial DataOps. The way HighByte positions it is straightforward. Industrial DataOps is not going to overlap with your existing systems that run the plant. It’s a layer you put on top — one that connects to all those systems, contextualises the data, and provides access to it from other consumers. AI being one of them. The real work, though, is in harmonisation. “How do you consistently represent equipment that could be the same or completely different across facilities?” Aron asks. That’s the scalability piece. At one factory, the production order comes from an ERP or MES. At another, it’s a CSV file on someone’s desktop. If you’re consuming that data from AI — or from anything else — you shouldn’t need to know the difference. UNS: Valuable, But It Was Always Just the Start We couldn’t have this conversation without touching on Unified Namespace. And Aron’s perspective has matured in the same way the market has. Five years ago, UNS was the start of DataOps at the edge: connect to everything, contextualise it, publish it to an MQTT broker. And for someone coming from nothing — like that bakery with independent machines and local HMIs everywhere — being able to subscribe to a broker and see the real-time state of the factory was genuinely transformative. But now? Aron is candid: “Publishing to MQTT is just the transport of real-time data out. That strategy sticks around.” With AI entering the picture, you also need historical data access, governance across multiple sites and contextual dimensions. “It’s not that anything we did is nulled,” Aron clarifies. “It was the first step of the journey towards something bigger.” LLMs and Agents: Let’s Demystify This David pushes for definitions, and Aron delivers them with refreshing clarity. * LLMs are the large neural networks most of us experience through chatbots. You put in a request, a response comes out. The context of the chat goes in each time (or a compressed version), and a new response is generated. * Tool calling is when the LLM detects it can’t answer on its own and reaches out to external tools — this is where MCP comes into play. * Agents are really just the code around the LLM. An agent is an LLM in a loop: instead of a human driving the conversation, the agent code handles tool calls, feeds results back into the context window, and decides when the task is done. How autonomous it is depends entirely on the code you write around that loop. “But make no mistake about it,” Aron says. “We’re just building text documents and feeding them into a large language model. And then stuff comes out that’s sometimes in JSON format that we can code against. It’s not beyond that.” LLMs are probabilistic. They’re hallucinating all the time and they just happen to be right a percentage of the time. In manufacturing, where determinism is fundamental, that’s a real constraint. Now imagine that kind of confident hallucination in a pharmaceutical clean room. Or on a production line with multi-million-euro equipment. “If we’re talking about fully autonomous agents that are going to get control of our machines — very bad idea,” Aron says flatly. MCP: Getting Beat Up, But Here to Stay MCP (Model Context Protocol) has had a steep rise — and Aron acknowledges it’s getting some backlash. That’s normal for anything with a sharp hype curve. His take: “MCP is remote procedure calling for LLMs.“ It provides discovery, remote calling, and a bit of instruction in between. Where it gets criticised is by technologists who look at it and say: this should just be a REST API. The funny thing? “If you go look at the latest version of MCP, it essentially is a REST API,” Aron says. “They got rid of server-sent events and it’s essentially an API now.” The only scenario where MCP loses relevance, in Aron’s view, is if LLMs get good enough to call conventional APIs directly — and those APIs evolve to be more interaction-oriented rather than CRUD-style developer tools. But for HighByte, it doesn’t matter much either way: you’re still building the Industrial DataOps layer, the contextualisation, the custom pipelines. Whether the AI calls that via MCP or REST is an implementation detail. What matters is that the tools are deterministic. “You define the MCP tool, its inputs, its outputs. If you get a call that’s trying to do SQL injection, you detect that and stop it with deterministic logic,” Aron explains. That’s why HighByte’s implementation was never a static, one-size-fits-all tool set. It’s designed so you build your own tools — like APIs — and control exactly what they do. A Quick Teaser: i3X We briefly touched on i3X, a new standard that Aron is heavily involved in alongside Jonathan Weiss and Matthew Paris. We won’t go into detail here — we’ll be recording a dedicated podcast with John Dyck and Jonathan Wise in the coming weeks — but the short version is this: Aron sees i3X as a standardised API layer for the factory. Every vendor already has APIs, but everyone does it their own way. If the industry can standardise how contextual data is accessed across vendors, it removes a massive amount of inefficiency. “If cloud vendors step up and build clients for i3X, that is going to be the ingestion highway in and out of the factory,” Aron says. Version one of the spec is expected before the end of Q2 2026. AI for DataOps: The Slider from Manual to Autonomous Beyond using DataOps to feed AI (which is the foundation story), there’s the flip side: using AI to do DataOps faster. HighByte’s latest release includes what they call a Pipeline AI Agent. Pipelines are their ETL tool for moving and transforming industrial data. The agent lets you prompt an LLM to analyse and edit pipeline configurations. What’s clever is the middle ground it occupies. Aron describes it as an AI slider: On one end, fully manual — you know what you’re doing and you want full control. On the other end, full autonomy — let the AI handle it. The Pipeline AI Agent sits in between: human in the loop. You prompt it, the UI shows you inline what it edited, you review, accept or reject, and iterate. This is where AI genuinely accelerates experts. Not by replacing them, but by handling the tedious parts so they can focus on what requires judgement. Give it a P&ID diagram, an Excel sheet, and access to an OPC server, and let it help you get started with contextualisation. Human in the loop to create deterministic output — absolutely needed — but accelerated by AI. Wrapping Up A year ago, we talked about Industrial DataOps as a discipline. Now it’s a recognised market segment. The conversation has shifted from “why do I need a data foundation?” to “how do I build one so AI can actually deliver?” The takeaway from this conversation? The foundation hasn’t changed. The urgency has. And the hype-to-reality gap on agents and autonomous AI is still very wide. The vendors and practitioners who will win are the ones who are honest about where that gap lies — and focus on the use cases that deliver real value today. HighByte will be at Hannover Messe 2026 (April 20-24) — look for Aron and his colleagues at the AWS, Siemens, and Microsoft booths. Thanks for listening — and thanks to HighByte for sponsoring this one! … and if you haven’t listened to the previous conversation we had with Aron, why not do it now? About HighByte HighByte is an industrial software company founded in 2018 in Portland, Maine USA. The company builds solutions that address the data architecture and integration challenges faced by manufacturers and industrial companies as they digitally transform. HighByte Intelligence Hub, the company’s proven Industrial DataOps software, provides modeled, ready-to-use data to the Cloud using a codeless interface to speed integration time and accelerate analytics. The Intelligence Hub has been deployed in more than a dozen countries by the world’s most innovative companies spannin
