← Insights
Briefing · Water & Utilities
In Water Operations, the AI That Works Isn't Autonomy
Where AI actually earns its place in water and wastewater today, and where the autonomous-plant story is still marketing.
The bottom line
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.
What's actually deployed
1
Leak and anomaly detection is the most production-ready use case, tied to water-balance discipline.
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.
2
SCADA and telemetry analytics are the real foundation.
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.
3
Reverse-osmosis energy value is real, but proven engineering comes before AI.
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.
4
Predictive maintenance works best on critical rotating assets.
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."
5
Digital twins and autonomous operation are real only when scoped.
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.
What it means for your operation
✓
Build the telemetry-to-action layer first.
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.
✓
Start with leak and anomaly detection tied to a water balance.
It's the most production-ready use case, and it pays in non-revenue water you can actually measure.
✓
Treat a digital twin as a scoped tool, not a value claim.
Name the model boundary, the operating decision it drives, the fail-safe, and the measured baseline, or don't call it a twin.
✓
Demand audited, operator-controlled ROI.
Vendor case studies prove deployment, not board-ready returns. Use them for diligence targets, not as financial evidence.
The wild card to watch
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.
Claim ledger
14/14
Checked
citation clusters traced to primary or source-of-record pages
0
Fabricated
no invented figures found
2
Corrected
qualified after source review
5
Demoted
useful signals kept out of the headline
ConfirmedXylem Vue / GoAigua (via Idrica): vendor-agnostic smart-water platform, 400+ water systems globally, with drinking-water, wastewater, and resource modules. Vendor-stated, not independent audit.idrica.com
ConfirmedAir Selangor monitoring: a state-wide program covering 8.4M residents and 6,200 km of trunk mains with leak-detection and meter-analytics modules. Benefits listed, not quantified as audited ROI.idrica.com
CorrectedSouth Bend smart sewer: real sensor-and-valve active control; the roughly $100M avoided-pipe-construction figure is strong secondary reporting, primary city and EPA sources not reachable this pass.wikipedia
ConfirmedEnergy Recovery PX: pressure-exchange energy recovery deployed at scale in seawater RO, with vendor claims of 35,000 devices, 100 countries, 99% peak efficiency, and up to 60% SWRO energy-reduction potential.energyrecovery.com
CorrectedVeolia Hassyan RO: the 818,000 m3/day capacity and 2.9 kWh/m3 claims are for a plant scheduled to begin operating in 2026 and reach full capacity in 2027, so not a completed ROI case.wsj.com
ConfirmedFilipe et al. (2019), wastewater pump optimization: a 90-day real case-study comparison reports 16.7% lower electrical energy and a 97% alarm reduction in the predictive-control scenario.arxiv.org
ConfirmedRahbaralam et al. (2020): Barcelona water-distribution pipe-failure data used for machine learning and survival analysis to inform maintenance and renovation decisions.arxiv.org
ConfirmedColwell & Abolghasemi (2024): a wastewater digital twin developed in a utility context; an extended abstract, not a portfolio ROI ledger.arxiv.org
ConfirmedSharma et al. (2020), digital-twin state of the art: identifies the lack of a universal reference framework, weak quantitative metrics, and data-security concerns as adoption barriers.arxiv.org
ConfirmedAmankwaa et al. (2021), utility digitalisation study: found limited impact so far, with value accruing mainly at the operational rather than strategic level. A useful transferability caution.arxiv.org
DemotedNon-revenue-water economics: World-Bank-attributed figures near $14B/year global cost and $2.9B in potential cash gains were reachable only through a secondary source. Directional, not audited.wikipedia
DemotedSouth Bend cost and overflow history: secondary reporting cites a roughly $150M Phase 1 and a 75% combined-sewer-overflow reduction by 2019. Context, not primary audited ROI.wikipedia
DemotedWater-sector cyber figures: the EPA Office of Inspector General's November 2024 report (25-N-0004) records a passive assessment of 1,062 drinking-water systems with 97 rated critical or high risk. Re-verified against the EPA source July 17, 2026 (the original pass could not reach it); kept as cyber-risk context only, never an ROI claim.epa.gov
DemotedRemote-access PLC exposure: public alerts on internet-exposed programmable controllers support the risk, but this brief uses it only to qualify remote-operations governance, never as an ROI claim.tomshardware
Demoted"Autonomous plant operations" as a near-term category (confidence 4/10): the evidence supports remote monitoring and bounded control loops, not autonomous multi-site plant operation across regulated regimes. Monitored, not asserted.contested
First moves before buying broad AI
01
Build the portfolio telemetry spine first.
Standardize asset IDs, SCADA tags, alarm classes, historian retention, CMMS links, lab and compliance fields, and site criticality across plants. Without this, every AI pilot becomes a bespoke data-cleaning project, and nothing compounds across sites.
02
Start leak and anomaly detection where the loss is already material.
Use district-metered areas, pressure zones, acoustic and pressure sensors, meter analytics, and repair-ticket closure where non-revenue water, bulk-water cost, scarcity, or customer penalties already hurt. Treat the software as one component of a water-loss operating program, not the program itself.
03
Instrument the top 10-20% of failure-cost assets for condition-based maintenance.
Prioritize lift-station and high-service pumps, blowers, RO high-pressure pumps, critical valves, generators, and electrical gear where downtime cost is unambiguous. Require prediction-to-work-order closure and avoided-failure accounting, not a generic plant-wide "AI program."
04
For RO desalination, separate proven energy equipment from AI claims.
Confirm the energy-recovery devices, pump curves, VFD controls, membrane cleaning regime, recovery ratio, and tariff schedule first. Then test analytics for fouling prediction, cleaning timing, setpoint recommendations, and energy-per-cubic-meter reduction on top of that engineering.
05
Frame remote operations as human-supervised portfolio command, not autonomy.
A central operating center can triage alarms, benchmark sites, guide crews, and standardize incident response. Keep local accountability and a defined fail-safe. Do not allow autonomous control without a named regulator and insurer posture behind it.
Owner, briefing, proof
Owner
A named operating owner for each deployed play (leak program, condition monitoring, RO energy), accountable for the telemetry-to-work-order loop, the vendor's scope, and the retirement decision.
Briefing
A production / pilot / marketing decision brief per vendor claim, so you buy deployed capability and hold pilots and platform language to a boundary before scaling across sites.
Proof
The chain of custody from asset tag to anomaly to decision to authorized work to verified outcome, plus audited, operator-controlled ROI rather than vendor case studies.
Where to start
Start with owner, briefing, and proof for one site and one play. If the gap is material, widen to a readiness look at a portfolio telemetry-to-action operating system, and build the operating machinery only when the operator wants it run.
Where the evidence stops
Three widely-shared claims run ahead of the evidence. Headline single-site savings figures are often secondary-source reporting rather than audited ROI. Many AI-optimization studies are simulations or short trials, not live operator results at portfolio scale. And "digital twin" is frequently a marketing label on a model with no named boundary. Treat all three as diligence targets, not proof.
What would change this conclusion
- A multi-plant operator publishes audited, operator-controlled ROI for leak detection, remote operations, predictive maintenance, or RO optimization across at least 25 sites.
- A regulator, insurer, or consent-decree authority accepts a bounded autonomous-control model for water or wastewater plant operation.
- A major vendor publishes independently audited ROI rather than case-study benefits.
- A large RO facility publishes post-commissioning energy-per-cubic-meter and lifecycle-cost data that confirms or contradicts current pre-operation claims.
- A peer-reviewed or utility-led field trial shows AI or digital twins outperform conventional advanced process control on energy, compliance, or maintenance at production scale.