Five technologies manufacturers should understand before scaling AI
A factory can have plenty of AI ambition and still be technically unable to scale it. A model that works against a controlled dataset in one plant may struggle when it meets different naming conventions, older equipment, disconnected applications, unreliable networks and information that was never designed to be read by an algorithm.
That is why the infrastructure around industrial AI increasingly matters as much as the model itself. Manufacturers do not need to master every emerging technology before they begin, but they do need to understand the architectural choices that determine whether a successful use case can move beyond one machine, line or site. Five technologies are becoming particularly important.
1. Industrial data contextualization
Industrial systems produce huge volumes of information, but a raw tag tells an AI system remarkably little. A temperature value becomes useful only when it can be related to the asset that generated it, the product being made, the process state, the unit of measurement and the event that followed.
This is the role of industrial data contextualization. Rather than replacing PLCs, historians, MES, SCADA or ERP systems, a contextualization layer sits across them and creates a consistent model that applications can query without needing to understand every underlying system.
Aron Semle, CTO of HighByte, argues that this missing layer is becoming more obvious as manufacturers try to apply AI across fragmented plants. “You need a contextualization layer. Underneath it is all the existing stuff you have today, because you cannot come into a plant and start ripping out PLC connectivity, HMI or SCADA. You need to hook into what is there, give it a common taxonomy, contextualize it, govern it and then provide access to it. The consumer should not have to understand all the complexity you have built over the last 10 or 15 years.”
For manufacturers, the practical lesson is to stop treating data preparation as a one-off AI project. The same contextual model should be reusable by maintenance, quality, production, engineering and future AI applications. That is what allows a use case proven on one asset to recognize equivalent conditions elsewhere.
2. Event-driven data architecture and the unified namespace
Once data has meaning, it still must move. Traditional point-to-point integration creates a direct connection between one producer and one consumer. It works until dozens of machines, applications and sites need to exchange information, at which point every new integration adds another dependency.
Event-driven architectures take a different approach. Technologies such as MQTT allow systems to publish changes as events while other systems subscribe to the information they need. A unified namespace adds a common structure so those events are presented consistently across the enterprise.
Magnus McCune, CTO of HiveMQ, sees the unified namespace less as a visualization tool than a governance mechanism. “The unified namespace is a design pattern that helps create a single source of truth for what the data is. Every site, every system and every AI agent cannot keep naming and integrating data in its own structure. It is about governing that consistency and creating a shared language, so when I refer to the temperature on a motor at site one, anyone has a quick understanding of how to find that value and where it should be.”
This matters to AI because models and agents cannot scale economically if every deployment requires a new set of bespoke integrations. Event-driven architecture separates the systems producing data from those consuming it, while a common namespace reduces the amount of translation required as applications multiply.
3. Hybrid edge AI
Sending every factory signal to the cloud before deciding what to do with it is neither necessary nor always practical. Some AI workloads can tolerate delay. Others influence quality, safety or production continuity and need to operate close to the process.
Hybrid AI architecture divides the work accordingly. Central infrastructure can support model training, cross-site learning and governance, while edge systems perform inference locally and continue operating when connectivity is constrained.
Saleh Al-Nemer, Regional Enterprise Architect and Distinguished Technologist at HPE, believes the old edge-versus-cloud argument has largely been replaced by deciding how the two should work together. “Historically, the discussion was edge versus cloud, but now it is about how the edge and cloud work together. The edge is important where the operation is time-sensitive, where you need local autonomy and where data locality matters. The cloud can provide the central controls and model training without affecting factory operations. It also allows manufacturers to bring information together across sites so engineering and product teams can use it over a much longer horizon.”
The placement decision should follow operational consequence. Al-Nemer points to an AI inspection application involving green anodes emerging from a furnace. Identifying an unsuitable anode before it continues can prevent it passing through another 25 production stages, avoiding unnecessary energy use and processing cost.
Manufacturers therefore need to think about compute, network resilience, security, power and cooling as part of the AI use case, not infrastructure details to be solved after the model has been selected.
4. Digital twins and simulation
Industrial AI is different from enterprise AI because its recommendations can change physical systems. A poor answer may waste material, damage equipment or disrupt production. Manufacturers therefore need a way to explore decisions before allowing them to alter the factory.
Digital twins and simulation provide that environment. Their value is moving beyond visual representations toward synchronized operational models that combine engineering information with live data and allow interventions to be tested before they reach the physical asset.
Dave Philp, Chief Value Officer at Bentley Systems, says this marks a shift away from what he describes as “digital theatre.” “A few years ago, you had amazing Hollywood-esque representations that were great in the boardroom but disconnected from day-to-day operational decision-making. We are now seeing digital twins much more connected to live operational data and enterprise systems, with AI increasingly coming in as decision support. The twin must keep synchronizing with the asset so it remains the twin every day.”
For AI, that creates a controlled environment in which a recommendation can be evaluated against engineering and operational constraints. The objective is not to build the most sophisticated model possible. It is to provide enough fidelity to test whether a proposed change is safe, useful and likely to improve the outcome before production pays the price for being wrong.
5. Information governance and controlled AI access
The final technology is less visible but increasingly important. AI systems need access to maintenance instructions, engineering records, quality documents, specifications and other information that may already be subject to different permissions and retention rules.
Simply connecting a copilot or agent to document repositories can expose the wrong material, produce answers from outdated files or make it difficult for employees to understand why the system reached a particular conclusion. Governance therefore has to become part of the AI architecture.
Tiina Hietajarvi, Principal Manufacturing Industry Adviser at M-Files, argues that manufacturers need more precise control than either opening everything up or locking it away. “If manufacturers want AI benefits across departments, they need to think about whether information can be shared in a controlled way rather than simply sharing a whole folder with everything in it. It needs cross-departmental thinking, defined use cases and a system that can provide the governance level required to control access and create trust.”
That includes metadata, permissions, document status, lineage and evidence of where an answer came from. For an engineer deciding how to maintain an asset, knowing that an AI response was drawn from the approved instruction for that specific product and customer is materially different from receiving an unattributed summary generated from a folder of PDFs.
Scaling requires architecture not another pilot
None of these technologies guarantees that an AI project will deliver value. Nor does a manufacturer need to implement all five as a giant transformation program before doing anything useful.
Their importance lies in what happens after a successful pilot. Contextualization makes operational data intelligible. Event-driven architecture allows it to move without multiplying integrations. Edge computing puts intelligence where timing and resilience demand it. Digital twins provide somewhere to test decisions before they become physical. Governance controls which information AI can use and helps people understand why its answers deserve trust.
A model may demonstrate the possibility. These technologies determine whether the capability can survive the factory around it and be repeated at industrial scale.
