{"resourceId":"veolia-ai-aeration-control-causal-energy-savings-2026","versions":[{"version":"external-32167e5f1154b74aef923952ff0fd42e5702c45593e5fdf89b1bd373341a65c0","resource":{"id":"veolia-ai-aeration-control-causal-energy-savings-2026","title":"Green AI in Industry: quasi-experimental evidence that AI aeration control cut wastewater plant electricity use 5.4%","organization":"Collège de France, ENS-PSL, Sciences Po (OFCE), ETH Zurich, INSEAD; data from Veolia Water France and Purecontrol","sector":"Municipal wastewater treatment (activated sludge plants under delegated operation)","geography":"France","publishedAt":"NBER conference paper, dated 17 Jul 2026","publicationDate":"2026-07-17","eventDate":null,"sourceName":"National Bureau of Economic Research (conference paper)","sourceLabel":"Academic working paper","sourceUrl":"https://conference.nber.org/conf_papers/f240807.pdf","evidenceClass":"academic-research","outcomeClass":"effective","topics":["infrastructure","operating-model"],"finding":"Using staggered roll-out across more than 200 plants and unplanned interruptions of AI control, the authors estimate that full-time AI aeration control reduced electricity use 5.4%, carbon emissions 6% and energy spending 8.2%, while effluent quality improved. Electricity used by the AI computing represents less than 1% of the savings. AI-equipped plants were more resilient under extreme weather and pollutant peaks and shifted load from peak to off-peak hours. The analytical sample is 164 plants (139 AI-managed by the end) observed January 2022 to July 2025. Eligible plants were 10,000 to 50,000 population equivalent using activated sludge; plants volunteered and early ones needed no IT changes.","sledRelevance":"Why now: K-Water's autonomous treatment results were reported on 31 Oct (10% power cut at one pilot, self-reported). This paper is the strongest available independent, causal estimate of the same class of AI control, and it gives US wastewater utilities a more conservative planning figure. It is not in the archive.","evidence":"Quasi-experimental panel study (staggered adoption and within-plant interruptions) with operator monitoring data; conference paper, not yet peer reviewed. Data and support came from Veolia and the technology provider. The authors note that external validity beyond these sectors and electricity systems still needs testing; broader DICE welfare figures are extrapolations.","architectureImplications":"Plant-level digital twins drive aeration setpoints through existing automation interfaces; first-wave plants needed no IT changes, which suggests the control layer can sit on existing SCADA where interfaces exist.","governanceImplications":"Volunteer plants and technical-readiness selection mean results apply to well-instrumented plants; utilities should require baselines and interruption logs to verify savings.","securityPrivacyImplications":"Remote AI control of aeration adds a write path into plant OT; the paper does not assess cybersecurity.","caveats":"Published July 2026, so older than the window; included only as the independent counterpoint to K-Water's 31 Oct claims (see sledRelevance). French plants of a specific size and process, run by a private operator under delegated contracts, so transfer to US municipal plants is not guaranteed. Not peer reviewed.","streamIds":["utility-management"],"roles":{"sales":"Sales framing: this gives a credible, conservative number. A causal estimate of 5.4% lower electricity use from AI aeration control, with effluent quality held, is a planning figure that wastewater directors and finance staff can test. Ask about aeration's share of plant energy and the plant's instrumentation. A bounded engagement is a measured pilot at one or two activated-sludge plants with an agreed baseline. Do not claim the same result for other processes, plant sizes or US conditions without local measurement.","engineering":"Engineering framing: the method is the lesson. The authors used staggered adoption and unplanned AI interruptions to separate AI effect from weather and load. For a US pilot, confirm activated-sludge process, plant size, dissolved-oxygen and ammonia sensing and an automation interface for setpoints. Build in planned off periods or a matched control plant so savings can be measured rather than asserted, and track the AI's own computing energy.","delivery":"Delivery framing: the owner is the plant superintendent, with process engineering and OT security. Work is to baseline energy per unit treated, install the control layer, log every AI on and off period and report monthly against permit limits. Proposed acceptance criteria are measured energy reduction with no permit exceedances. Risks include sensor drift degrading control, choosing only easy plants and overstating results, and a new OT write path without security review."},"roleDetails":{"sales":{"customerProblem":"Aeration is a major wastewater energy cost and utilities lack independent evidence on AI control savings.","stakeholders":["Wastewater utility director","Plant superintendent","Finance / energy manager","OT security lead"],"discoveryQuestions":["What share of plant electricity goes to aeration?","Which plants have DO and ammonia sensing and setpoint interfaces?","Would you accept planned AI off periods to verify savings?"],"valueHypothesis":"AI aeration control may reduce plant electricity use by a few percent while holding effluent quality, as measured in French plants.","potentialEngagement":"A one- or two-plant measured pilot with baseline, interruption logs and monthly reporting.","unsupportedClaims":["5.4% savings at any given US plant","Savings for processes other than aeration","Results for plants outside the 10,000-50,000 PE range"]},"engineering":{"fit":"Fits activated-sludge plants with reliable sensing and automation interfaces, similar in size to the studied plants.","architectureIntegration":"Digital-twin control layer writing aeration setpoints through existing automation; monitoring data feeds the model.","prerequisites":["DO and ammonia sensors","Setpoint interface to blowers","Baseline energy and flow data"],"constraints":["Different US permits and electricity prices","Plant size and process limits","Study not peer reviewed"],"security":"Treat the control layer as an OT write path: segment it, restrict remote access and log all setpoint changes.","proofOfValue":["Planned AI on/off periods or matched control plant","Energy per unit treated against baseline","Effluent compliance tracked throughout"]},"delivery":{"work":"Baseline, install control layer, log AI interruptions, report energy and effluent monthly for at least a year.","dependencies":["Instrumentation health","OT security review","Permit compliance monitoring"],"ownership":"Wastewater plant superintendent, with process engineering.","skillsAdoption":"Operators trained to supervise and override AI setpoints and to maintain sensors.","governanceCheckpoints":["OT security approval before go-live","Monthly savings and compliance review","Annual decision on roll-out to other plants"],"acceptanceCriteria":["Proposed: measured aeration energy reduction against baseline over 12 months","Proposed: zero permit exceedances attributable to AI control","Proposed: all AI on/off periods logged"],"risks":["Sensor drift","Selection bias toward easy plants","New OT attack surface"]}},"retrievedAt":"2026-11-02T08:35:54-06:00","enrichedAt":"2026-11-02T08:41:03-06:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Not assessed; implies operators supervise AI setpoints rather than adjust aeration manually.","procurementImplications":"Supports outcome-based contracting tied to measured energy per unit treated, with interruption logs as evidence.","operatingModelImplications":"AI control was rolled out in waves by readiness; a phased, measured roll-out is a transferable operating approach.","peerReview":"preprint","sourceVerification":{"openedUrl":"https://conference.nber.org/conf_papers/f240807.pdf","referenceExcerpt":"The additional electricity demand generated by AI models represents less than 1% of these savings.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}