What to Actually Buy When Every Water Vendor Claims AI

At a water conference a few years ago, I watched a vendor pitch their AI platform to a room of utility managers. The slides were impressive: neural networks, real-time anomaly detection, predictive maintenance, a dashboard that looked like mission control. The Q&A was where things got interesting. An operations manager from a mid-sized utility raised her hand and asked: “What data do you need to run this?”

The vendor’s answer took about two minutes. It required SCADA integration, two years of historical sensor data in a specific format, asset records digitized to a particular standard, and a dedicated IT resource to manage the API. The operations manager nodded slowly and said: “We have some of that.”

Some of that. That phrase is the story of AI in water right now.

The technology is real. The potential is real. And the gap between where most utilities actually are and where they need to be to use it is also very real, and rarely discussed honestly by the people selling these platforms.

The problem with “AI-powered”

Every water technology vendor is AI-powered now. Leak detection platforms. Chemical dosing systems. Asset management tools. Customer analytics. SCADA dashboards. The label has become meaningless, which makes it harder, not easier, to evaluate what you’re actually buying.

When a vendor says “AI-powered,” they might mean any of the following: a machine learning model trained on sensor data to flag anomalies; a statistical algorithm dressed up in modern language; a rules-based system with an if-then logic tree; a large language model used to generate reports; or genuinely sophisticated predictive modeling that took years to build and validate.

These are not the same thing. Some of them are valuable. Some of them are expensive ways to do things you could do with a spreadsheet. The label tells you nothing.

What you need to ask is: what is the model actually predicting, how was it trained, on whose data, and what evidence exists that it works in conditions similar to yours?

Where AI is genuinely useful in water right now

There are a handful of applications where the technology has moved past proof-of-concept and into something utilities are getting real value from.

Pipe failure prediction. Machine learning models trained on break history, pipe age, material, soil conditions, and installation records can meaningfully improve capital planning. The data requirements are substantial, but many utilities have more of this than they realize; it just isn’t in one place. The models that work best in this space are built on utility-specific data, not generic industry benchmarks.

Chemical dosing optimization. Real-time adjustment of coagulant and disinfection doses based on source water quality is a mature application with documented results. It requires good sensor coverage and reliable SCADA integration, but the ROI is measurable in chemical costs and compliance risk.

Energy optimization. Pump scheduling and energy cost management based on demand forecasting and time-of-use pricing is well-established. Some utilities are achieving 10–15% reductions in energy costs with relatively low implementation complexity.

Anomaly detection in distribution systems. Pressure and flow monitoring to flag potential leaks or contamination events before they become crises. The technology works; the challenge is alert fatigue. Systems that cry wolf train operators to ignore them.

Customer analytics. Consumption pattern analysis for demand forecasting, conservation program targeting, and affordability intervention is increasingly accessible even for smaller utilities, particularly where AMI infrastructure exists.

Where it doesn’t work yet, and what vendors won’t tell you

Predictive maintenance at scale. The pitch sounds perfect: AI monitors your assets and tells you what’s about to fail before it does. The reality is that most utilities do not have the sensor density, maintenance records, or failure history to train a useful model. Without good data, you get a system that either misses failures or generates so many false positives that operators stop trusting it. The vendors selling this know this. They will tell you the model improves over time as it learns your system, which is true, but that learning period can be years.

Automated treatment decisions. AI can inform and recommend. It should not be making autonomous treatment decisions on systems that serve public drinking water. The regulatory framework does not support it and the consequences of errors are too severe. Any vendor suggesting otherwise should make you cautious.

One-size-fits-all models. A model trained on data from large urban systems may perform poorly on a small rural system with different source water, aging infrastructure, and highly variable demand. Ask vendors specifically where their model was trained and what the operational profile of those utilities looks like. If they can’t answer clearly, the model may not be as applicable as the pitch suggests.

The data readiness question no one wants to answer first

Before you evaluate any AI platform, you need an honest assessment of your own data infrastructure. Not aspirationally — actually.

Do you have reliable, continuous sensor data? How complete is the historical record? Is your asset data digitized and accurate? Do you have IT resources to manage integrations and maintain systems? Is your SCADA system modern enough to connect to third-party platforms?

If the honest answer to most of those questions is “partially” or “no,” that does not mean AI has nothing to offer you. It means the first investment is probably not an AI platform; it is the data infrastructure that would make an AI platform useful. That is a less exciting pitch, but it is the right sequence.

Vendors will tell you they can work with imperfect data. Sometimes that’s true. More often, imperfect data produces imperfect models, and you don’t discover that until you’ve spent the budget.

How to evaluate vendors in this space

A few questions that will tell you more than any demo:

Who are your current utility customers at a similar scale and operational profile to ours? Get references and actually call them. Ask specifically what the implementation was like, what data was required, and what results they’ve seen after twelve months of use.

What does implementation actually require? Not the optimistic version. The realistic version. How many hours of internal staff time? What integrations are needed? What is the typical timeline from contract to operational?

What happens when the model is wrong? How does the system handle false positives? What is the escalation process? How is the model updated when conditions change?

How is the model trained and maintained? Is it a shared model across all customers, or is it trained on your specific data? Who owns the model and the data? What happens to your data if you end the contract?

What does the pricing model look like at year three? AI platforms often have low entry costs and rising subscription fees as usage scales. Understand the total cost of ownership over a realistic engagement period.

The right frame for decisions

AI is not a category of technology you adopt or don’t adopt. It is a set of tools that solve specific problems, and like any tool, the question is whether the problem it solves is the right problem for your utility to be focused on right now.

The utilities getting the most out of AI investments tend to have a few things in common: they identified a specific, measurable problem first; they invested in data quality before they invested in analytics; they had internal champions who understood enough to evaluate what vendors were telling them; and they were willing to start small, validate results, and scale only what worked.

That is not a particularly exciting framework. But it is the one that produces outcomes rather than expensive dashboards.

---

Adam Tank is the founder of HydroKnowledge and co-founder of Transcend, a water technology company that built AI-enabled engineering tools now used by utilities serving hundreds of millions of people. HydroKnowledge helps utilities and technology companies navigate AI adoption in the water sector. Get in touch if you’re working through these decisions.