Closing the Loop: Practical AI for Water Asset Management

Imagine the future to act now

I was nominated to contribute to the State of AI by RISE newsletter, where each issue highlights a researcher’s Futureframing—a short reflection meant to spark new perspectives on AI. My theme is asset management in water and wastewater sector, a sector often hidden underground yet vital for our daily lives. In this reflection I look ahead to how AI can help close the loop: turning data into decisions, and decisions into learning. Below, I expand on this vision in more detail, showing how agentic AI, human governance, and continuous feedback could reshape the future of our utilities.

The challenge we must solve

Sweden’s water and wastewater utilities face a familiar squeeze: aging networks, climate-driven extremes, and rising expectations—without a matching rise in budget or specialist capacity for example Sweden’s water and wastewater infrastructure has an annual under investment of 10 billion SEK with a growing investment debt according to Svenskt Vattens investment report 2023. Beneath that headline are a handful of stubborn blockers to AI-enabled asset management to support municipalities in facing these challenges:

• Data silos, data quality, data readiness, trust, restrictive data security classifications and cybersecurity concerns .
• Pilot-itis and ROI uncertainty. One-off experiments rarely cross the chasm into daily routines; utilities struggle to show measurable impact that convinces operations, finance and boards.
• Trust, safety and compliance. The EU AI Act, NIS2 and CER demand governance, documentation and resilience that many organisations are only starting to operationalise.
• Competence constraints. Small and medium municipalities in particular need practical, “learn-by-doing” pathways—not just tools.
• Risk blindness. Renewal plans still lean heavily on age and anecdote; uncertainty is rarely quantified, making prioritisation debatable and slow.

If we want AI to meaningfully reduce bursts, localise leaks faster, and stretch renewal budgets, we need to turn data into decisions—and decisions into documented learning.

What we’re doing now at RISE

Across our projects we are building the capabilities—technical and organisational—that close this loop.

1. Readiness → Adoption → Impact (the RAI Loop).
   We start by framing AI as a management discipline, not a model hunt. The RAI Loop aligns a real problem and outcome metric (e.g., “time to localise a leak”), embeds decision support in existing tools (SCADA/GIS/CMMS), and demands proof of impact against a baseline. Lessons learned feed back into strategy, so the next project starts stronger, not from zero.
2. Gamified capacity-building.
   Through our Water Infrastructure Management Game and workshop series under development, we hope to bring and sandbox where cross-functional teams rehearse real trade-offs—inspection vs. renewal, mapped vs. unmapped zones, budget shocks, cluster bursts—inside a safe sandbox. Participants see how data quality affects model outputs, and how different models (statistical, ML, hydraulic) can disagree. The goal is confidence with human-in-the-loop decision-making, not blind automation.
3. Governance and assurance by design.
   We’re packaging practical templates—model cards, data lineage, access control, bias/drift monitoring, fail-safe modes—so utilities can meet AI Act expectations alongside NIS2/CER resilience. If a sensor dies or a model drift, the playbook still works.
4. From pilots to a Learning Network.
   Working with municipalities and partners by creating communities of practice, where we’re shaping a shared Water Learning Network: common artefacts, repeatable pipelines, and case evidence that accelerates uptake across utilities with different contexts.

The next frontier: agentic, human-governed AI (2035 outlook)

Looking ahead to 2035, we see a system-of-systems where agentic AI supports people across perception, reasoning and governance:

• Perception at the edge. Swarms of ground, aerial and aquatic robots (UGV/UAV/ROV) run calibrated vision models to detect cracks, infiltration, corrosion or biofouling; fixed acoustics and pressure arrays add context.
• Reasoning with LLM agents. A planning layer—tool-using LLM agents with access to SCADA, GIS and CMMS—triages risk, proposes actions (inspect/patch/renew), and explains why in plain language. It cites the data used and shows expected impact ranges, not just a single score.
• Humans govern and teach. Operators approve actions, set policies, and correct mistakes; that feedback becomes labelled data. Each intervention updates a digital twin and retrains models on a cadence utilities control.
• Continuous learning in operations. The loop tightens: tomorrow’s patrol route is smarter because today’s work order was approved (or overridden) and the outcome was captured. Over time the organisation’s time-to-insight shrinks and confidence grows.

Taking the next step together

For utilities that want to move forward, the best place to start is small but meaningful: choose one use case that matters, and build readiness around it. At RISE, we work alongside utilities on this journey—co-designing pilots, exploring governance templates aligned with the AI Act, validating models, visualising uncertainty, and even using serious games to strengthen decision-making across teams.

The aim is simple: to turn data into decisions, and decisions into shared learning. By connecting local results into a broader Learning Network, each step strengthens not just one utility, but the sector as a whole.

Explore more of my work at RISE:

• AIX – AI Predictive Assessment & Sewer Index: building data-driven models to map risk across sewer networks.
• Data-Driven Asset Management through Gamification: using a serious game to help utilities internalise and practise data-driven decision making.