Agentic AI In The Water Sector: From Chatbots To Digital Operators

By Ashwin Dhanasekar and Sandy Moskovitz

How utilities can move from AI-generated insight to supervised, real-world action

If you have attended a water conference in the last two years, you have heard the term "Generative AI" more times than you can count. You’ve likely seen demos of ChatGPT writing a passable cover letter or a compliance summary. And if you are like most engineers and utility managers we speak to, your reaction has been a mix of mild impressiveness and practical skepticism.

“Sure, it can write a poem about wastewater treatment, but can it optimize my aeration blowers at 3:00 AM?”

Until recently, the answer was a hard no. Now the conversation is shifting. We are moving from Generative AI, which creates content, to Agentic AI, which performs actions. In our previous article, we explored how utilities can adopt AI responsibly, with governance, transparency, and human oversight at the core. The next question utilities are now asking is: If AI can inform decisions, can it also help execute them? For the water sector, this distinction is not just semantic; it is the difference between a digital assistant that takes notes and a digital operator that turns valves.

The Shift: Thinkers vs. Doers

To understand Agentic AI, we must look past the hype surrounding Large Language Models (LLMs) and focus on how AI systems are beginning to move from generating insights to carrying out work. Generative AI is reactive. You give it a prompt, and it gives you text or an image. It is a "thinker." In the water sector, it is already proving useful for tasks such as drafting NPDES reports or summarizing lengthy EPA guidelines. However, it lives inside a chat box, and it doesn't touch anything.

Agentic AI is proactive. It is goal-oriented. You don't give it a prompt; you define an objective. An agent has the ability to perceive its environment (via SCADA data, sensors, or market prices), reason through a plan, and then execute that plan by interacting with other software or physical systems. Unlike traditional automation or advanced process control systems, Agentic AI can reason across domains such as operations, maintenance, energy markets, and administrative workflows, rather than optimizing a single process in isolation.

Think of Generative AI as a junior consultant who writes a great memo about a problem. Think of Agentic AI as a seasoned field operator who notices a pump vibrating strangely, checks the maintenance log, orders a replacement part, and adjusts the flow setpoint to compensate, all before you’ve finished your morning coffee.

What This Looks Like In The Field

The water sector is arguably one of the best testing grounds for Agentic AI. Water systems are complex, data-rich, and operate on physics-based logic. This makes them well suited for AI systems that must reason, plan, and act reliably. Here is how Agentic AI moves beyond "chat" to deliver real operational value.

1. The Autonomous Treatment Plant

In a traditional setup, a SCADA system alarms when dissolved oxygen (DO) levels drop. An operator acknowledges the alarm and manually adjusts the blower speed.

An AI agent doesn't wait for the alarm; it constantly monitors influent flow, ammonia levels, and energy prices. It can predict a spike in biological load is coming in two hours, and decide to pre-aerate slightly now, while energy rates are lower in order to buffer the system. It executes the setpoint change directly to the PLC and then logs the action and the rationale ("Pre-emptive DO increase for anticipated load spike") in the operations log for the human manager to review later.

This is not science fiction. Utilities like those in Cuxhaven, Germany, have already used early forms of this logic to reduce aeration energy by 30% using digital twins. Agentic AI builds on these foundations by tightening the decision loop and, where appropriate, increasing autonomy under defined constraints (Xylem, Cuxhaven Wastewater Treatment Plant AI Optimization case study). This autonomy operates within predefined safety envelopes and physics-based constraints, not free-form decision-making.

2. Leak Detection

If treatment plants are the heartbeat of utility operations, distribution networks are the nervous system. Current acoustic leak detection puts dots on a map. A human analyst has to look at the dots, verify them, and issue a work order.

An Agentic workflow changes the pace. The "Leak Agent" detects an anomaly in the nighttime flow data of a specific DMA (district metered area) and cross-references this signal with acoustic sensor data. When the correlation exceeds a defined confidence threshold, the agent can automatically:

• Query the GIS to identify the pipe material, diameter, and age.
• Check crew availability and skills in the work management system.
• Draft a provisional work order with the likely leak location and estimated repair parts needed.
• Send a "Review & Approve" notification to the distribution manager.

The human stops being the data collator and becomes the decision validator.

A real-world precursor to this model can be seen at VA SYD, the regional water utility serving southern Sweden. VA SYD has deployed AI-based leak detection that analyzes flow data to identify and prioritize hidden leaks across its distribution network. Rather than relying on manual review of raw sensor outputs, the system ranks suspected leaks by probability and impact, enabling faster intervention and documented reductions in non-revenue water (Siemens, VA SYD SIWA Leak Finder case study).

Agentic AI builds directly on this foundation, connecting detection, asset context, and response into a single, supervised workflow that accelerates action without removing human control. The agent accelerates response but does not replace engineering judgment or field verification.

3. Procurement and Compliance Agents

Operational gains alone are not enough. Much of a utility’s inefficiency hides in administrative workflows. Agentic AI is particularly well suited to handling multi-step workflows that consume staff time but add little strategic value. Imagine an agent tasked with "get quotes for 500 feet of ductile iron pipe." The agent can independently email approved vendors, parse incoming PDF quotes, compare the pricing against the historical database, check lead times, and present a comparison table and a draft purchase order for the best option.

Early versions of this approach are already appearing in utilities that use AI-assisted procurement and contract management tools to streamline sourcing and compliance reviews. While most deployments today stop short of autonomous execution, they demonstrate how agentic workflows can significantly reduce cycle times and free staff to focus on higher-value tasks.DC Water, for example, has begun deploying Generative AI tools internally to support adminis

trative and procurement-adjacent workflows, including summarizing vendor responses, extracting key terms from procurement documents, and accelerating internal reviews. While these systems do not execute purchases autonomously, they significantly reduce manual document handling and decision latency, freeing staff to focus on negotiation, risk management, and compliance judgment (Bluefield Research & Xylem, Reimagining Water Management report).

Agentic AI builds directly on this foundation. By connecting document analysis, benchmarking, workflow coordination, and approval routing into a single goal-driven process, procurement agents compress cycle times without bypassing governance. Humans remain accountable for final decisions but are no longer trapped in the mechanics of paperwork.

The Cautionary Tale: Why We Need "Guardrails"

If the idea of an AI turning valves or emailing vendors makes you nervous, good. It should. The water sector is critical infrastructure; we cannot afford "hallucinations" in our control logic.

Remember, Agentic AI does not replace responsible AI principles. It stress-tests them. When systems move from generating insights to executing actions, governance, explainability, and human oversight become even more critical.

There are three major risks we must manage:

1. The Infinite Loop: If an agent is trying to optimize for pressure and another agent is optimizing for energy, they could get into a "war," constantly overriding each other and causing system instability.
2. Hallucinated Actions: Generative AI might make up a fact. Agentic AI might act on a bad fact. If an agent misinterprets a sensor reading as a fire flow demand and opens a valve that drains a tank, the consequences are physical and dangerous.
3. The "Black Box" Problem: If a tank overflows and the log says "AI Agent 007 commanded valve open," but doesn't explain why, we have lost control of our system.

The Responsible Path Forward: Human-in-the-Loop

We don't need to choose between manual drudgery and terrifying automation. The most effective way to deploy Agentic AI in water is through a Human-in-the-Loop (HITL) architecture. An HITL model operationalizes responsible AI in practice by embedding governance through approval thresholds, transparency through logged rationale, and robustness through bounded autonomy. In this model, the agent does all the legwork: the monitoring, the planning, and the drafting of the action. However, it doesn't execute the final "commit" without a human sign-off, at least initially. HITL architectures can also serve as a training ground, allowing less-experienced staff to learn from AI-recommended actions while preserving institutional knowledge.

For example, the agent might say: "I have detected a drop in pump efficiency at Station 4. I recommend switching to Pump 2 and scheduling maintenance for Pump 1.

Do you approve? [Yes/No]"

As trust builds, you can move to Bounded Autonomy. You give the agent permission to act on its own, but only within strict "sandboxes."

• “You may adjust chemical dosing by +/- 10% without asking me. Anything more requires approval.”
• “You may draft emails to vendors, but you cannot hit send.”

The Path Forward

Moving toward Agentic AI is not a solo journey. It requires utilities, technology providers, and industry organizations to collaborate, test assumptions, and share lessons learned.

To support this, SWAN Americas Alliance has launched a collaborative research project focused on the role of agentic AI in the water sector. The project brings together utilities, solution providers, and practitioners to explore practical use cases, governance models, implementation pathways, and workforce implications.

As we have previously explored, responsible AI adoption is no longer optional for the water sector. At the same time, the water sector is facing a "silver tsunami" of retiring workforce and an actual tsunami of climate challenges. We are being asked to do more with fewer people. We cannot solve this problem simply by working harder or hiring more young professionals who don't exist.

We need force multipliers. Generative AI was an important first step. It showed us that computers could understand language and summarize complexity. Agentic AI is the tool that allows computers to understand work.

By carefully deploying agents, starting with administrative tasks and low-risk operational advisories, we can free our human engineers to do what they do best: solve complex, novel problems that no algorithm can touch. The future of water isn't about replacing operators; it's about giving every operator a team of digital experts working 24/7 in the background. The question for utilities is no longer “Should we explore AI?” but “Which decisions are we ready to let machines prepare, and which actions are we ready to let them take, under our supervision?”

Ashwin Dhanasekar is Operations Leader for Digital Solutions at Brown & Caldwell.

Sandy Moskovitz is Programme Director for the Smart Water Networks Forum (SWAN).

SWAN, the Smart Water Networks Forum (SWAN), is the leading global hub for the smart water sector. A UK-based non-profit, SWAN brings together leading international water utilities, solution providers, academics, investors, regulators, and other industry experts to accelerate the awareness and adoption of “smart,” data-driven solutions in water and wastewater networks worldwide. Learn more at www.swan-forum.com.