AI in the Water Sector: Why Value Starts with the Data Foundation

Few technology topics are being discussed as widely right now as artificial intelligence, and the water sector is no exception. At industry conferences, in association papers and in many product brochures, AI has become part of the future agenda. The promises range from automated leak detection and predictive maintenance to the intelligent control of entire network sections.

Still, many utilities remain quietly sceptical. Are the polished demos backed by robust solutions — or are they mainly a new argument for justifying digitalisation budgets? It is a fair question. AI can already create real value in the water sector today, but primarily where the foundations are in place. And those foundations are exactly what is often missing from the public debate.

Where AI Is Already Working Today

One of the most obvious and well-documented fields of application is leak detection in drinking water networks. The problem is well known across the sector: depending on the condition, age and structure of the network, a share of treated drinking water is lost before it reaches customers.

Traditional methods such as minimum night flow analysis provide initial indications at district metered area level, but they remain relatively coarse. Newer approaches combine high-frequency pressure and flow data with machine learning, allowing anomalies to be located much more precisely. Some methods do not just look at individual measuring points, but model the network as a connected system, making it possible to detect even small deviations before visible damage occurs. Several German research projects involving municipal utilities and universities are currently working to move these approaches from the lab into regular operations.

There are also robust fields of application beyond leak detection:

• Predictive maintenance for pumps, valves and system components, based on actual sensor data rather than fixed maintenance intervals
• Automated sewer inspection, where image analysis can classify damage in wastewater pipes faster and more consistently than purely manual visual inspection
• Demand forecasting, combining seasonal patterns, weather data and demographic development to anticipate peak demand earlier
• Drought and extreme weather management by linking soil moisture sensors with satellite data
• Real-time control in wastewater treatment plants, where models detect process deviations and derive control recommendations

As different as these cases are, they have one thing in common: they do not replace the experience of employees. They add a depth of analysis that simply cannot be achieved manually.

Why AI Often Does Not Fail Because of the Model

In practice, the real bottleneck is rarely the algorithm itself. More often, it is the data foundation underneath. An anomaly detection model is only as good as the granularity, frequency and reliability of the data it is fed with. Anyone who wants to locate leaks precisely first needs reliable automated remote reading of measurement data, consistent geospatial data from the GIS and sensors that actually provide the required measurement frequency.

Without that foundation, even the most sophisticated model will not help much. In the worst case, it produces false alarms. And few things damage trust in a new technology as quickly as an alarm that turns out to be wrong three times in a row. The same pattern can be seen in the wave currently occupying so many organisations: language models and generative AI. An assistant designed to handle enquiries in customer service or public administration can sound very convincing in a demo. But without embedded domain knowledge, clear rules and defined boundaries, it can also respond confidently when the factual basis is missing. The answer may sound plausible, even though it is not technically or operationally reliable.

Whether it is anomaly detection in the network or a text assistant in an administrative office, the technology itself is rarely the whole problem. What matters is the context you give it. This brings a topic into focus that is often discussed separately from AI in the industry, although the two are inseparable: an organisation’s data culture.

An AI initiative that meets a fragmented data landscape, isolated departmental silos and inconsistent standards will struggle to deliver on expectations, even with a large budget. Conversely, every investment in clean master data management, consistent interfaces and a shared data foundation across departmental boundaries does more than improve the conditions for AI. It improves day-to-day operations and increases the chances that AI projects will later become genuinely usable.

Regulation: Caution, Yes — Standstill, No

With the European AI Act, the question of how AI applications in critical or regulated areas should be classified is moving further into focus. Industry associations have pointed out that blanket high-risk classifications could place a heavy burden on small and medium-sized utilities in particular. At the same time, it is clear that applications affecting security of supply, sensitive data or automated decisions require careful assessment.

In practice, it helps to document early which data is being used, how results are checked and where human oversight remains in place. This reduces regulatory risks and builds trust internally.

What This Means in Practice

For utilities, municipalities and associations now asking themselves how to approach AI, a fairly clear order emerges. The starting point is not tool selection. It is an honest assessment of the organisation’s own data: How granular is it? How up to date? How consistent?

Only then does it make sense to define a clearly scoped, measurable pilot project — for example, leak detection in a defined network section — instead of launching a company-wide AI initiative without a clear measure of success. And finally, perhaps the most important insight: AI competence is not purely an IT task. It requires people who can assess both the technical potential and the operational limits of the models.

AI will not transform the water sector overnight. But it will gain importance over the coming years wherever utilities consistently expand their data foundation. The advantage will not come from introducing just any AI tool as early as possible. It will come from having the patience to first build the structures on which more reliable AI applications can actually run.