A morning in Bilbao, discussing the future of industry

Why the next step in smart manufacturing is not just better AI models, but connecting intelligence to industrial knowledge.

Recently, I had the opportunity to participate in a roundtable in Bilbao with industry leaders from northern Spain, discussing one of the most important transformations facing industrial companies today: the adoption of Artificial Intelligence.

The event took place in the context of AMAIA, Inetum’s AI Experience Center, an initiative designed to bring AI experimentation, collaboration and practical business applications closer to organizations.

Discover the AMAIA AI Experience Center

We spent the morning exchanging ideas about the challenges facing industrial companies, from machine failure prediction and Overall Equipment Effectiveness (OEE) to talent shortages, process automation and the growing role of Large Language Models (LLMs).

What struck me most was that, despite the different perspectives around the table, we kept coming back to the same fundamental question:

How can we turn technological innovation into something that genuinely improves the way industrial companies operate?

The answer is not simply to introduce AI into existing processes. It is to rethink how people, machines, data and software work together.

Industry has been using AI longer than many people realize

When people talk about AI today, the conversation often starts with ChatGPT, copilots and generative AI. But industrial companies have been applying forms of artificial intelligence and advanced analytics for years.

Predictive maintenance is a good example.

Industrial equipment generates enormous amounts of data through sensors, controllers and monitoring systems. Machine learning models can analyze this information to identify patterns associated with equipment degradation and potential failures.

Instead of waiting for a machine to break down, maintenance teams can use these predictions to plan interventions, reduce unplanned downtime and make better use of maintenance resources.

The business impact can be significant. Every unexpected stoppage can affect production schedules, delivery commitments, labor utilization and profitability.

The same logic applies to Overall Equipment Effectiveness, or OEE. By combining availability, performance and quality measurements, companies can identify where production capacity is being lost and investigate the causes. A predictive model might indicate that a machine is likely to fail within the next 24 hours. That is valuable information, but it is only the beginning of the operational decision.

Which component is at risk? Is there a maintenance order already open? Do we have the required spare parts? Is the machine part of a critical production line? Can maintenance be scheduled without disrupting an urgent customer order?

Answering these questions requires more than a prediction. It requires connecting that prediction with the operational context of the factory. And this is where the next generation of AI applications becomes particularly interesting.

From predicting problems to helping solve them

Machine learning and generative AI solve different problems, and their combination creates new possibilities.

Machine learning models can detect patterns, estimate probabilities and forecast potential failures. LLMs can help people interpret information, interact with documentation and formulate solutions. Intelligent agents can go a step further by coordinating tools, retrieving relevant data and executing defined workflows under appropriate controls.

Imagine an engineer asking a system:

“Why has OEE on production line three decreased this week, and what should we investigate first?”

A useful industrial AI application would need to do much more than generate a plausible answer. It would have to retrieve production figures, compare them with historical performance, identify the main sources of downtime, examine maintenance records and potentially correlate the findings with quality incidents or machine telemetry.

It could then explain the most likely causes, provide evidence for its conclusions and suggest the next diagnostic steps. With the right integrations and permissions, an agent could also prepare a maintenance request or initiate an approved workflow. Actions affecting production would still require the appropriate authorization and human oversight. This is a meaningful evolution: from displaying information on a dashboard to helping people interpret it and act on it.

The objective is not to replace industrial expertise. It is to make that expertise more accessible and effective.

The talent challenge: AI must amplify engineers, not just automate tasks

One of the recurring topics in our discussion in Bilbao was the difficulty of finding qualified engineers.

Industrial companies need people who understand complex machinery, production processes, automation systems and the relationships between them. That expertise takes years to develop.

At the same time, experienced professionals are often overwhelmed by fragmented information, manual administrative tasks and the effort required to investigate operational problems.

There is also a less visible risk: critical knowledge frequently resides in the heads of a relatively small number of experienced employees.

When those people retire, change roles or leave the company, part of the organization’s practical knowledge can disappear with them.

This is where AI agents could make a real difference.

An agent connected to technical documentation, maintenance history, engineering specifications and operational systems could help an engineer find the relevant information much faster. It could guide a less experienced technician through a diagnostic procedure, summarize similar incidents or identify the documentation associated with a particular machine component.

However, there is an important condition: the system must understand the company’s actual environment.

A general-purpose LLM does not automatically know how a particular factory is configured, what an internal equipment code means or which maintenance procedure applies to a specific machine.

Connecting AI to enterprise information is necessary. Making that information understandable and meaningful is equally important.

The missing piece: industrial knowledge and context

Industrial data rarely lives in one place.

Production systems, enterprise resource planning platforms, manufacturing execution systems, maintenance applications, engineering repositories and historians all hold different pieces of the operational picture.

Even when these systems are integrated, their data may use different naming conventions, identifiers and representations of the same real-world objects.

One system might refer to a machine as M-102, another as Machine_102, while an engineering document uses a plant-specific designation. A component may be known by its internal part number, its supplier reference or its functional name.

People who know the factory can often navigate these differences because they understand the context. An AI agent cannot be expected to infer every relationship reliably.

This is why I believe industrial AI needs a semantic foundation: a shared understanding of the entities, relationships, terminology and rules that describe the business.

Ontologies and knowledge graphs can help provide that foundation.

An ontology defines the concepts in a domain and how they relate to each other. A knowledge graph represents entities and their relationships, connecting information that would otherwise remain scattered across separate systems.

Consider a simple industrial example:

A production line contains several machines. Each machine has components. Components have maintenance histories, technical specifications and associated spare parts. Machines contribute to production orders, and their performance affects the overall OEE of the line.

When this information is connected, an AI agent can reason over a richer picture of the operation.

Instead of searching for isolated records, it can follow meaningful relationships between a machine, its components, its maintenance history and its production impact.

This does not eliminate the need for accurate data, good integrations or sound reasoning. Nor does a knowledge graph automatically guarantee a correct answer. But it gives the agent a much stronger starting point.

Naming conventions are more important than they might seem

A related lesson is the importance of consistent naming conventions.

Naming standards can sound like an unglamorous data-management exercise. Yet inconsistent terminology creates real friction for AI systems, especially when agents must retrieve and combine information from multiple sources.

If equipment identifiers, business terms and relationships are defined consistently, an agent has less ambiguity to resolve. If they are not, it may retrieve the wrong records, overlook relevant information or draw conclusions from incomplete context.

The solution is not necessarily to rename every legacy system or undertake a massive data transformation before starting an AI project.

It is to establish a common semantic layer that maps existing terminology to shared concepts, preserves relationships and makes the meaning of enterprise data explicit.

That layer can evolve incrementally, beginning with a high-value use case and expanding as new needs emerge.

Bringing this approach into enterprise software

These ideas are closely related to the work we are doing at Siemens, both in our internal adoption of AI and in the evolution of our software products.

Internally, we are seeing how important it is to help people incorporate AI into their daily work, rather than treating it as a separate technology initiative. Adoption involves more than providing access to an LLM. People need to understand where it can help, how to use it effectively and how to work within the appropriate security and governance boundaries.

At the product level, we are also working to give engineers better ways to use intelligent capabilities in their applications and workflows.

Through Mendix, for example, developers can build applications that combine business processes, enterprise data and AI capabilities. This opens up opportunities to create applications in which intelligent agents support operational users instead of leaving them to navigate multiple systems manually.

One relevant area is the Mendix Enterprise Knowledge Graph.

The idea is to connect information across enterprise systems through a semantic representation of the business, making relationships and context available to downstream applications and AI agents. Mendix Graph Studio supports the creation of enterprise knowledge graphs and the mapping of relationships across data domains.

For industrial organizations, this approach can help connect information that matters to a particular operational decision, without assuming that every existing system must first be replaced. The broader Siemens context is also worth exploring through Intelligence Center X, which brings together industrial data and intelligence to support more informed action.

The opportunity is to combine industrial domain knowledge, engineering software, enterprise applications and AI in a way that makes operational expertise easier to access and use. The value is not in adding a chatbot to an existing application simply because the technology is available. It is in designing the application around a real problem and giving the AI the tools, context and permissions needed to help solve it.

From isolated pilots to industrial-scale adoption

Of course, moving from an interesting demonstration to a reliable production solution is not straightforward. Industrial environments introduce constraints that cannot be ignored: safety, cybersecurity, traceability, system availability, data quality and the consequences of an incorrect decision. A convincing AI demonstration is not necessarily a production-ready industrial solution.

Organizations need to define what an agent is allowed to access, which actions it can perform, when a human must approve a decision and how the system’s behavior can be monitored and audited. They also need to establish clear measures of success.

For a predictive maintenance initiative, this might include unplanned downtime, maintenance costs and the accuracy and usefulness of failure predictions. For an AI-assisted engineering workflow, it might include investigation time, resolution time, the quality of recommendations and the number of manual steps eliminated.

The key is to begin with a specific business problem, establish a measurable baseline and validate the results before expanding the solution. There is no need to automate an entire factory in one step. In many cases, connecting two or three relevant data domains and enabling one well-defined workflow can be more valuable than launching a broad initiative without a clear operational objective.

The next industrial revolution will be about collaboration

Looking back at our conversations in Bilbao, I think one of the most important conclusions is that the industrial AI journey is not primarily a technology problem.

The technology is advancing rapidly. Machine learning models are becoming more capable, LLMs are changing how people interact with software, and agentic systems are opening new ways to coordinate work across applications.

But successful adoption depends on how well organizations connect these capabilities with their processes, data, people and accumulated expertise.

The companies that make progress will not necessarily be those that deploy the largest number of AI agents. They will be the ones that identify where intelligence can improve a decision or a process, provide the necessary context, establish appropriate controls and measure the results.

For industrial engineers, this could mean spending less time searching for information and more time solving problems. For maintenance teams, it could mean moving from reactive intervention towards better-informed decisions. For operations managers, it could mean understanding production losses faster and coordinating corrective actions more effectively.

And for industrial companies facing a shortage of experienced professionals, it could mean making valuable knowledge available to more people, without pretending that expertise can be fully automated.

I left Bilbao convinced that the opportunity is significant, but also that we need to be pragmatic about how we pursue it.

The future of industrial AI is not just about machines that can predict what might happen. It is about people and intelligent systems working together, with the right knowledge and context, to decide what to do next.

That is where I believe the real transformation begins.