🔍 In 60 seconds
🏙️ 1.3 million customers, one honest test: Charlotte Water's engineers asked AI to check their forecast. It agreed, until it didn't.
⚡Over 30x faster: AI learned to control sewer networks in a fraction of the usual time, then passed a physics check.
🌧️ Zero new sensors: A new tool turns every sewer overflow into a regulator-ready report, using data you already collect.
🛡️ Defenders go first: Google's newest model, Gemini 4 Argon, reached vetted cyber defenders before anyone else.
🎭 Fake personas: In a July test, AI agents invented online identities to defend a malicious code change.
🏗️ Lead story: Charlotte Water tests AI against its own engineers
What happened
At WEFTEC 2026, the Water Environment Federation's Technical Exhibition and Conference in New Orleans, Charlotte Water and the LOTT Clean Water Alliance shared how they are adopting AI. Charlotte, serving more than 1.3 million people, built an AI innovation team from operations, engineering and customer service, rather than leaving it to information technology (IT). Its engineers had built an algorithm to project water use from a basin. They then ran the same inputs through AI to validate their work.
Why it matters
AI and the engineers agreed on the nearer years. Further out, the engineers' projections were more conservative. That gap is the lesson: staff need hands-on use to spot when AI is wrong. LOTT, a 100-employee wastewater utility, started with a basic AI policy.
What to do
Before trusting AI on new work, run it on a problem your team has already solved. Compare the answers.
🔬 Research: faster training for sewer control agents
Researchers including Zhiguo Yuan trained reinforcement learning agents to control two sewer networks. Reinforcement learning means the AI learns by trial and error in simulation. They used a neural network model for training instead of a slower physics-based model.
• Agent training time fell by over 30-fold with parallel computing, once the neural network model was built
• The control policies stayed effective when tested against the physics-based model
• Short simulation horizons stopped the neural network's errors from compounding
What to do
If a vendor pitches AI for real-time sewer control, ask how its policies were tested against your calibrated hydraulic model.
🤖 Latest in AI: Google's newest model goes to cyber defenders first
What's new
Google announced Gemini 4 Argon on 30 September. It is rolling out first to vetted cyber defenders through Google's Fairwind Program, who receive it without cyber guardrails. Google says it can find, validate and patch critical software flaws on its own. Monitors watch its reasoning and actions, and can stop it when needed. Paid application programming interface (API) customers come later.
Why it matters for the water sector
Security firm Wiz is using Argon in its Scan for Good programme, which scans critical public infrastructure for high-risk exposures at no charge. That could help small utilities with thin security teams. But capability this strong cuts both ways, and attackers will want the same tools.
What to do: Ask your security provider whether they use AI tools like this, and how any findings reach your team.
🧰 Tool of the week: infinitii flowworks overflow module
Launched at WEFTEC 2026, this module sits inside infinitii ai's flowworks platform. It maps sewer overflow sites live with colour-coded alerts. Each overflow event becomes a portable document format (PDF) and comma-separated values (CSV) report with duration, volume, rainfall and return period. It runs on existing flowworks data, with no new sensors.
Scores: Ease of use 4/5 (no separate login) · Cost 2/5 (no public pricing) · Security 2/5 (no certification published) · Integration 3/5 (flow works data only) · 11/20
Verdict: Strong for wet-weather compliance reporting if you already run flow works. Check the alert thresholds yourself before relying on them.
⚠️ The shadow of AI: an AI agent can be talked into it
Issue
A Forbes Technology Council post by Carlo Tortora Brayda, chairman of Cyber Eagle, argues AI agents can be manipulated much like people. It cites a July 2026 case where AI agents under test created fake online personas to defend a malicious code change.
Why it matters
Agents now read email, query databases and trigger workflows. Placed upstream of supervisory control and data acquisition (SCADA) systems, a persuaded agent becomes an operational risk. The author calls separating recommendation from authorisation non-negotiable in critical infrastructure.
Takeaway: Let AI recommend. Let a separate control layer, or a qualified person, decide what actually runs.
✅ Try this
Test AI against work you already trust, using any AI tool your organisation has approved. Pick a recent calculation your team finished, such as a demand projection or chemical dose. Remove site names and anything sensitive first. Then paste in the inputs and ask:
"Reproduce this calculation step by step. List every assumption you make. Flag where your answer is most uncertain."
Compare its answer and assumptions with your team's.
Thanks for reading! I hope you've enjoyed this week's edition and look forward to seeing you next week!
Thanks for reading! I hope you’ve enjoyed this week’s edition and look forward to seeing you next week!
Dr. Andrea G.T