How much intelligence does an installation need?

An HVAC system automatically reduces energy consumption because the weather forecast changes. A production line schedules maintenance before a bearing fails. And an operator has a conversation with an AI assistant to investigate a malfunction.

Much of this technology already exists. The discussion is therefore increasingly less about what is technically possible. The more interesting question is:

How much intelligence does a installation actually need?

An installation does not start with AI

Anyone looking at industrial automation might easily think that everything revolves around AI these days. In reality, virtually every installation still starts in the same place: with a measurement.

Temperature, pressure, flow, level, and energy consumption are measured, transmitted, and processed. Only then do visualization, analysis, optimization, and potentially artificial intelligence come into play. After all, without reliable measurement values, there is nothing to optimize.

JUMO focuses precisely on this layer of the automation chain: collecting and unlocking reliable data via sensors, IO-Link, and Single Pair Ethernet. [1] Datacation is located at the other end of the same chain. There, the focus shifts from measuring to predicting and deciding. AI agents manage buildings based on occupancy, energy prices, and weather forecasts, or predict maintenance before systems fail. [2]

Together, both perspectives show that automation does not consist of two worlds, but of one continuous process: measuring, understanding, and deciding.

More data is not automatically more knowledge

In many technical discussions, the same assumption creeps in unnoticed:

More data → more intelligence → better decisions.

That sounds logical.

Yet, sooner or later, every engineer asks the same question: what does the next data point actually add?

An extra temperature sensor at a critical location can be of great value. Sensor number fifty might not. The same applies to AI. An additional model only makes sense if it enables a better decision.

In practice, engineering constantly revolves around trade-offs between:

• accuracy;
• complexity;
• maintainability;
• investment costs;
• operational costs.

An installation that makes autonomous decisions must not only be smart. It must also remain explainable when something goes wrong.

The forgotten layer beneath AI

This brings us to an observation that often remains underexposed.

An AI agent sees no building.

An AI agent sees no machine.

An AI agent sees no process.

An AI agent sees exclusively data.

Which data is available is determined by technical choices:

• which variables are measured;
• where sensors are placed;
• how often data is collected;
• which measurement uncertainty is acceptable;
• which parameters are considered important.

As a result, every AI application relies on a measurement architecture that is usually much older than the model itself. Without reliable sensors, consistent communication, and good data quality, even the smartest AI turns into a producer of false certainty.

When does a system justify AI?

That leads to a more interesting question than:

Can we apply AI?

The real question is:

Under what circumstances does AI yield better decisions than traditional automation?

For relatively stable processes, fixed control strategies can function perfectly well. A combination of sensors, clear limit values, and proven control technology often results in a predictable and robust system.

It is a different matter when:

• many variables influence each other;
• circumstances constantly change;
• energy prices fluctuate;
• user behavior becomes unpredictable;
• Multiple objectives must be weighed simultaneously.

This creates a playing field where AI can actually deliver added value, because classic rules fall short.

Technological proportionality

Ultimately, the discussion about AI in industry is not about being for or against AI.

It concerns technological proportionality.

Not every installation needs an AI agent.

Not every rule-based question requires machine learning.

And not every dataset justifies a cloud architecture.

The challenge therefore shifts from:

What technology is available?

Unpleasant:

Which technology actually improves the quality of the decision?

Sometimes that leads to artificial intelligence.

Sometimes to an extra sensor.

Sometimes towards a better regulatory strategy.

And sometimes the existing solutions turn out to be surprisingly effective.

Ultimately, that is not an AI question.

That is an engineering question.

An AI agent doesn't see a machine. It sees data. Curious where that data comes from? Check out the live Ethernet APL setup at WoTS 2026 and speak to the specialists who apply this technology in practice.

Register

[1] FHI, federation of technology industries

[2] The self-learning building; agentic AI in building management – FHI, Federation of Technology Industries