Industries / Energy, water and waste / Water and wastewater
Running a plant
The daily work of keeping treatment and pumping going: testing the water, adjusting the process, keeping records, and protecting the controls.
6
Milestones achieved, from 97 CE to 2026.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
97 CE
Rome's water commissioner audits
Frontinus, appointed Rome's water commissioner in 97 CE, checked the fixed-size nozzles that metered water to users against the official standards and found that the water-men who ran the system had altered some of them and drew water off through secret branch pipes. He wrote his account, he said, to learn the job and to guide whoever came after him.
LessonMeasuring what a system actually delivers is how losses hidden by the records come to light.
Credit
Frontinus
1988
A one-day bacteria test
Coliforms are gut bacteria used as a sign that sewage may have reached water. A Yale team published a test in which a water sample, added to a powder and kept warm, turns yellow within 24 hours if coliforms are present and glows under ultraviolet light if E. coli is, with no follow-up tests; in a trial with five utilities it did not differ statistically from the standard tube method. Sold as Colilert, it is on EPA's list of approved methods for compliance testing as Standard Methods 9223 B.
Credit
S. C. Edberg, M. J. Allen and D. B. Smith, Yale University School of Medicine
2021
A hack that was not
In February 2021 the Pinellas County sheriff called a jump in the lye (sodium hydroxide) setting at the Oldsmar, Florida, water plant, to about 100 times normal, a cyberattack, and CISA, the US cybersecurity agency, issued an advisory; plant staff saw the change and reversed it. In 2023 the city manager at the time called it a "nonevent" that may have been an employee's mistake, and the FBI said it could not confirm a targeted intrusion.
LessonCISA's advice held either way: put physical limits outside the software, such as chemical pumps sized so they cannot overdose the water.
Credit
City of Oldsmar; CISA; FBI
2023
Controllers hacked through default passwords
From November 2023, attackers that US agencies tie to Iran's Islamic Revolutionary Guard Corps, calling themselves CyberAv3ngers, took over Unitronics programmable logic controllers, small computers that run pumps and valves, that were reachable from the internet with default or no passwords. A joint advisory counts at least 75 devices compromised, at least 34 of them in US water and wastewater systems.
LessonThe fix was basic: take controllers off the open internet and change factory passwords, and the agencies asked makers to stop shipping default passwords.
Credit
reported by CISA, the FBI, NSA, EPA and other agencies (advisory AA23-335A)
2024
EPA finds basic gaps
In May 2024 EPA said that over 70% of the water systems it had inspected since September 2023 violated basic requirements of Section 1433 of the Safe Drinking Water Act, which makes systems serving more than 3,300 people assess their risks and keep emergency plans. Problems included unchanged default passwords, one login shared by all staff, and former employees who still had access.
Credit
US EPA
2026
AI aeration measured at scale
Aeration, blowing air into tanks so microbes can digest sewage, uses more than half of a typical plant's electricity. Economists studied an AI system from the French firm Purecontrol that Veolia switched on at French sewage plants from October 2022, and found that across 164 plants, from January 2022 to July 2025, full-time AI control cut electricity use by 5.4% and energy costs by 8.2% while the cleaned water's quality improved (NBER working paper, July 2026).
LessonThe gains were near zero at the plants that were already the most efficient.
Credit
Purecontrol and Veolia (system); Mathias Abitbol, Philippe Aghion, Céline Antonin, Lint Barrage and Benjamin Lengereau of the Collège de France, ETH Zurich and other institutions (study)
Watch for
AI pilots at Hampton Roads
Hampton Roads Sanitation District in Virginia, which runs 14 wastewater plants, has piloted generative AI to cut power use and chemical dosing, and has worked with other utilities on a shared agentic AI system, software that carries out multi-step tasks with tools; the post gives no results.
Achieved whenthe district or an outside party publishes measured power or chemical savings and how they were measured.
Claimed by
Xylem and Bluefield Research, in a post naming the district, July 2026
Open
01
Agent runs a simulated plant
An AI agent that operates a standard computer model of a whole wastewater plant through a simulated year, setting aeration and other controls and logging each decision with its reason.
The briefCredit · unclaimed02
Independent test of dosing AI
An outside before-and-after study of AI dosing at a drinking water plant: how much coagulant it saves and whether the treated water stays as clear, measured by people with no tie to the vendor.
The briefCredit · unclaimed
Open 01 · the brief
Agent runs a simulated plant
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An AI agent that operates a standard computer model of a whole wastewater plant through a simulated year, setting aeration and other controls and logging each decision with its reason.
Why it is still open
A 2026 review in Water Research says plants still rely heavily on human-supervised control and that AI agents need domain-specific benchmarking before they are deployed. A search found neural-network and reinforcement-learning controllers tested on the International Water Association's benchmark plant models, BSM1 and BSM2, but no published test of a language-model agent operating them.
Built with what exists
bsm2-python, an open-source Python version of BSM1 and BSM2 under a BSD license that reports effluent quality and operating cost indexes; an AI model that can call tools; the package's default controllers as the baseline.
First step
Install bsm2-python, run its closed-loop example, and expose the aeration setpoints as tools an agent can call.
Done looks like
Twenty independent runs of the same simulated year, each scored on the benchmark's effluent quality and operating cost indexes against the default controllers, with the number of runs in which the agent did as well as the baseline on both (see the Note "All twenty times").
Open 02 · the brief
Independent test of dosing AI
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An outside before-and-after study of AI dosing at a drinking water plant: how much coagulant it saves and whether the treated water stays as clear, measured by people with no tie to the vendor.
Why it is still open
Vendors report savings, such as the 23.3% cut in alum use that Envirosuite's own case study gives for Hong Kong's Tai Po plant, where the software advised operators and no outside reviewer is named. The economists who measured AI aeration in France call theirs the first academic study of an at-scale industrial AI deployment, and it covers sewage plants only.
Built with what exists
Hourly plant records shared by a utility (dose, flow, raw and treated turbidity); periods with the AI on and off; the statistical methods in the NBER paper, which compare periods and plants with and without the AI; Python.
First step
Ask utilities named in vendor case studies for anonymized hourly data and permission to publish the results.
Done looks like
A public report on at least one plant and twelve months of data, giving coagulant dose per cubic meter and treated-water turbidity with and without the AI, written by authors with no tie to the vendor.