Where AI actually earns its place in water and wastewater today, and where the autonomous-plant story is still marketing.
The useful, deployed layer in water and wastewater isn't autonomous plants. It's telemetry discipline plus narrow analytics tied to crews, maintenance, energy, and compliance: leak and anomaly detection, condition monitoring on critical rotating assets, and reverse-osmosis energy optimization. For a multi-plant operator, the moat is a repeatable telemetry-to-action operating system that makes every site measurable and comparable, not a robot that runs a plant.
Real deployments exist in trunk-main monitoring, district-metered-area analytics, acoustic sensing, pressure-transient detection, and meter-data analytics. The value is measurable in non-revenue water, but only when the analytics sit on top of an honest water balance.
The "AI" label matters less than historian quality and whether an alert actually closes a work order. Instrumented active-control cases (sensors plus controllable valves that shift flow) are the clearest public proof that this works in live systems.
Pressure-exchange energy recovery is deployed at scale in seawater RO and is a genuine production-grade lever. AI process optimization sits on top of that engineering; it doesn't replace it.
Pumps, blowers, motors, membranes, and critical electrical gear (where failure has a clear cost) are the credible targets. Condition-based maintenance is a focused lever, not a generic plant-wide "AI program."
Narrow twins tied to telemetry, and human-supervised remote operation, are credible. Broad "autonomous plant" claims across regulated, multi-site portfolios remain mostly pilot or marketing.
Standardize historians, alarms, and work-order closure across sites before buying "AI." That comparable data estate is the asset that compounds, and it makes every later lever cheap to deploy.
It's the most production-ready use case, and it pays in non-revenue water you can actually measure.
Name the model boundary, the operating decision it drives, the fail-safe, and the measured baseline, or don't call it a twin.
Vendor case studies prove deployment, not board-ready returns. Use them for diligence targets, not as financial evidence.
The winners here won't be the operators with the most impressive single-plant demo. They'll be the ones who turn a fleet of plants into one comparable, governable data estate, because that's what makes every future AI lever cheap to deploy and easy to trust. The moat is the operating system, not the algorithm.
Each figure in this brief was verified against primary or source-of-record pages before publication. Confidence reflects evidence quality, not author confidence.
Credible, verified sources include Xylem Vue / GoAigua deployment reporting (via Idrica), Energy Recovery pressure-exchange data, and peer-reviewed pipe-failure and pump-optimization studies. Named active-control and monitoring cases are treated as diligence targets, not audited ROI.
A few widely-shared numbers run ahead of the evidence. Headline single-site savings figures are often secondary-source reporting, not audited ROI. Many AI-optimization studies are simulations, not live operator results. And "digital twin" is frequently a marketing label on a model with no named boundary. Treat all three as diligence targets, not proof.
✓ Storm Research v2 · 14 citation clusters verified against primary sources, July 8 2026 · reliability = evidence quality, not author confidence