From chatbots to agents: AI’s next wave in water

A couple of years ago, when I first joined KWR as a young researcher, language models and generative artificial intelligence (AI) were still largely unnoticed by the wider public. They were a topic of research inside niche circles. Today, they are rapidly transforming the way we work. The next evolution of generative AI is agentic AI (or AI agents): AI systems that are (semi-)autonomous and can plan, execute tasks, and generate outputs by combining models, data, and external tools. These systems can integrate with other software and monitoring systems in the water sector, enabling the automation of workflows with limited human supervision.

This was the topic of the 2^nd Waterwise Hydroinformatics Knowledge Exchange meeting of 2026, part of a series of quarterly meetings, where participants from water utilities, research institutions and the industry discuss the use of novel technologies in practice. The presented applications and projects were met with enthusiasm, highlighting the power of this technology as well as the need for clear human oversight, continuing strong human domain expertise and ethical guardrails.

Enabling conversational hydraulic modelling

The session opened with a presentation by Christos Michalopoulos (KWR) and Dennis Zanutto (KWR) on the introduction of an Open-source Model Context Protocol (MCP) Server for EPANET, enabling conversational hydraulic modelling. MCP is an open standard and open-source framework (introduced by Anthropic in late 2024) that standardizes the way AI systems interact with external software systems and data sources. By developing an MCP server for the industry-standard Water Distribution System modeling software EPANET, natural language-driven simulations as well as AI agent system integration are enabled. In simple words, practitioners (or AI agents) can run complete hydraulic simulations and retrieve data through a simple conversational interface. The use of this technology aims to reduce the time needed to develop hydraulic simulation models, as well as aid in the development of automated decision-support workflows. Further, it allows users to express intent, with the LLM converting the user’s request into a series of formatted calls to tools, which would otherwise be hidden behind a UI. This can help operators access additional resources contextually in their workflow, for example, retrieving info about a street from a street view service when investigating a possible burst event. However, key challenges such as context bloating (e.g. hallucinations, limited context), security vulnerabilities (e.g. prompt injection, data governance challenges, premature authorization workflows) or human oversight (e.g. reliability of conversational input translation according to each user, transparency and control of actions by the agent), remain.

Image 1. Differences between API and MCP

A Hydroinformatics agentic AI web application in practice

In the second presentation of the session, Branislav (Bane) Ristić (Amazon Web Services – AWS) walked us through the foundations of agentic AI and a proof-of-concept agentic AI web application that monitors, simulates, analyzes and verifies its own recommendations across real water distribution network models. The application orchestrated 13 open-source tools from three organizations (KWR, Vitens and USEPA, including the KWR developed MCP EPANET server discussed previously) as well as a single-agent and a multi-agent model. The use of the application was demonstrated in the context of monitoring contamination events in water distribution networks and the proposal of contamination isolation actions by the AI agents, while other use cases were also proposed. Accordingly, it showcased the real value of utilizing LLM’s and agentic AI in fast prototyping, as well as the potential to aid decision-making in real operational settings. Challenges including data governance structures, risks regarding the usage of such tools in critical infrastructure contexts as well as the clear need for transparent and thorough human oversight, were discussed. Finally, what stood out was not only the technical integration, but also the speed of development: a functional prototype was created in just a few afternoons by someone without prior domain expertise in the water sector.

Image 2. Demo agentic AI web application for water distribution systems

The Role of Generative AI for Water Utilities

In the last presentation of the session, Gigi Karmous-Edwards (Karmous Edwards Consulting & NC State University) presented an overview and the findings of the WRF 5321 Research project (The Role of GenAI for the Global Water Sector). The project aimed to understand the role that GenAI powered by large language models (LLMs) could play for municipal water and wastewater utilities. The project highlighted that many use cases can already be implemented with relatively low cost and effort, providing value and enabling water utilities to work more efficiently with diverse data sources. The use cases covered a wide range, from knowledge-base chatbots to automated data input applications. At the same time, the study highlighted the importance of robust data governance structures, as well as addressing concerns around security, compliance, and workforce adaptation.

Image 3. Project summary of the WRF 5321 Research project (The Role of GenAI for the Global Water Sector)

Final thoughts

Looking ahead, agentic AI is set to become an increasingly important part of the water sector, presenting clear opportunities in automating workflows, enhancing decision-making and improving operational efficiency across utility processes. However, its adoption must be deliberate and responsible. Accordingly, robust data governance, domain expertise, clear human oversight and safeguards are necessary for successful outcomes, though precisely how to implement these remains an open research question. Ultimately, its greatest value should lie in strengthening, not replacing, expertise, while clear human oversight and control is necessary to ensure trust, resilience, fairness and sustainability.