Fifty Steps Back, Five Hundred Steps Forward

Put the AI down for a second. I build it for a living, for water utilities specifically, so this is either the worst sales pitch you have heard all year or the most honest one.

I have spent the last few weeks reading. Excerpts from the proceedings of the IWA Leading Edge conference in Houston. The recent work coming out of SWAN. Page after page from people who have spent whole careers keeping water safe and clean on a fraction of the money, the staffing, and the public attention the job deserves. I came away holding two feelings at once. One was respect, the kind that is hard to put into words. The other was a worry I cannot shake.

The worry is this. Every room in our field now arrives at the same comfortable sentence within ten minutes: AI should augment the people who run our utilities, not replace them. True, and by now obvious. Everyone nods, including me. Then the conversation tends to stop at the exact moment it should start, because the real question is not whether AI should help. It is whether we are ready to be helped.

We are playing catch-up. Catch-up from consent decrees we are still working off. Catch-up from a funding gap that is not a rounding error: the AWWA's 2026 assessment puts U.S. drinking water needs above two trillion dollars over the next 25 years, while current capital spending runs at barely a third of what is required each year. Catch-up from sea level rise, which does not consult our capital schedule. And while we run to catch up, the workforce that knows how the system actually behaves is walking out the door. Roughly a third of our water and wastewater operators are eligible to retire within the decade, and the knowledge that should outlive them sits scattered across systems that were never built to talk to each other.

Investment, operating obligations, and the loss of institutional knowledge compound one another.

Figure 01

Read the pressures together

So before we reach for the newest thing, I want us to be honest about the ground we are standing on.

Here is what almost no one says out loud. Utilities have never been given the attention they deserve. I say that as an engineer and I say it as a citizen. I turn on the tap and forget about it like everyone else. It is human to take a system for granted exactly when it is working well. The water shows up, so we stop thinking about the people and the plant that make it show up. That detachment is not a character flaw. It is the strange reward reliable systems earn: our inattention.

Figure 02

Make quiet expertise durable

The people who run water have been quietly innovating for years, without nearly enough credit for it.

And inside that inattention, the people who run water have been quietly innovating for years. I have read their work closely. The men and women publishing through IWA, SWAN, the Water Research Foundation, WEF, and AWWA. Operators, engineers, modelers, and researchers who have moved this field forward on a fraction of the budget, the staffing, and the credit they were owed, and who have never once done it at the expense of the people they serve. The opposite, in fact. Every careful page they wrote made someone's water a little safer. I am indebted to that body of work. I have spent my own years trying to add to it. So nothing here is a critique of them. It is a plea on their behalf.

They did all of it while meeting a list of obligations that only ever grows, with IT and OT systems that were never designed to cooperate. Not pressures. Requirements. The board and the council. The citizens. The regulators. Operations. Capital planning. The bond rating. Risk and resilience. Each one legitimate, each one non-negotiable, and the list never closes.

I want to name that experience precisely, because the name changes what we do about it. The resilience field gave us a useful pair of words: shocks and stresses. We are good at the shocks. A shock is a hurricane, a main break, a contamination event, and we drill and budget for those. The stresses are what quietly win. The constant, open-ended adaptation to an unknown and expanding set of requirements is not a shock. It is a stress, the chronic kind that wears a system down between the headlines. When the EPA finalized national PFAS limits for drinking water in 2024, that is what it felt like from inside a utility: a fully formed national obligation landing on top of every plate that was already full. Nobody got to finish the old list first.

So my message to the sector is not "move faster." It is the opposite, and it only sounds backward until you sit with it.

Take fifty steps back so you can take five hundred steps forward.

In plain terms, that is what a resilient system actually is. Not the one that reacts fastest to the next shiny thing. The one that pauses, finds its footing, and then moves with a quantitative method, a feedback loop, and a habit of continuous improvement underneath it. A system that steers itself, instead of being yanked around by whichever requirement arrived this quarter.

Which is why I will say something that may sound strange from a person who builds AI for water. Be careful what you put on autopilot. Pilot, experiment, learn, by all means. A pilot is often how you find out where your data is broken in the first place. But do not hand an autonomous agent the controls of a live treatment process while your understanding of that process lives in one retiring operator's head. The line is not "no AI." The line is no autonomy on ground you have not checked.

Start with education. Not a vendor demo. A real understanding of what AI is, what "agentic" actually means, and how these things work under the hood. We are more than capable of this. Look at who works in a utility: chemical engineers, electrical engineers, civil engineers, modelers, planners, operators who can hear a failing pump before a sensor catches it. We already fold half a dozen engineering disciplines into one working system every single day. A field that does that has no business telling itself it cannot understand a model. We can. We simply have to learn it before we buy it.

Education before automation, every single time, because understanding the tool is what keeps you in control of it.

Figure 03

Understand before handing over control

Then build self-reliance. A workforce that understands its own tools does not get captured by them. The same holds for any business. Self-reliance is the whole difference between using a technology and being used by one.

Let me make "fifty steps back" concrete, because otherwise it is just a nice phrase. Picture two treatment plants in the same utility. The same model of pump runs in both. In one plant the maintenance system calls it "Pump 3A." In the other it is logged under three different names, depending on who entered the work order that day. Now you want an agent to predict failures across your fleet. It cannot. Not because the AI is weak, but because it cannot reliably distinguish each physical pump or connect the correct records to it. The step back is boring, and it is everything: agree on what to call the pump. Standardize the tags. Clean the asset register. Make the record match reality. Do that across your plants and you have not slowed your AI down. You have built the only ground it can stand on. And notice what it costs. Agreeing on what to call a pump is a meeting, not a capital program. This is the rare fix the smallest utility in the country can begin tomorrow, with the people it already has, while the biggest one still has not.

Two plants, same utility, same AI tools available, and the quality of the foundation changes what those tools can reliably do.

Figure 04

Identity comes before prediction

The hype has one thing backward. It treats AI as the starting line. AI is not the starting line, it is the payoff. You would not hand a new analyst the keys to the trading account on day one, even if you let them watch the screens and learn while you still hold the keys. We should not hand an autonomous agent our treatment process before our data can be trusted to describe it.

Notice that none of the steps I just described require my company, or any company. Agree on your naming. Standardize your tags. Clean your asset register. Write down what the retiring operator knows before the retirement party. Teach your people what these systems actually do under the hood. You can begin all of it on Monday, with the staff you already have. I will still tell you where I have put my own hands, because hiding it would be dishonest: I founded APAS to do exactly this readiness work, and the name was deliberate, since in Sanskrit apas means water. But here is the test of whether this is advice or a sales pitch. If you do the list above and never call me, the advice still worked. That is the only kind of advice worth writing down.

None of this happens utility by utility, alone. The reason we stay in catch-up is not only money and not only retirements. It is that we keep having the same conversation in a hundred separate rooms. The Water Research Foundation, WEF, AWWA, IWA, the operators, the boards: all of us are circling the same problem and solving it in pieces. If we want the five hundred steps forward, we take them together or we do not take them at all.

That points straight at the next thing I want to write, so I will leave it as a question and pick it up in the next piece. We keep saying our data is fragmented. I think that is the symptom, not the disease. Data is not fragmented. People are fragmented. More on that soon.

So here is the question I would put ahead of the one everyone is asking. We keep asking what AI can do for our utilities. The question that comes first is why: why are we asking a machine to understand our utility before we have understood it ourselves? Your data is your utility. The condition of your pipes, the behavior of your plant, the memory of every operator who ever solved a problem at 2am. That is what the data is. If we, the humans in the loop, do not understand our own data, we are handing the interpretation to a system that does not yet know what it is looking at.

It is a little like your own health. The best doctor in the world can only work with the history you are able to give. Walk in knowing nothing about your own body, with no sense of what is normal for you and no record of what changed, and even a brilliant fifteen-minute visit is half blind. Walk in as a partner who knows your own baseline, and that same doctor can do extraordinary things. AI is that doctor for your utility. It is not there to understand your system in your place. It is there to do far more with what you already understand, and it can only reach as far as your own grasp of your own data will let it.

So take the fifty steps back, and start by understanding your own data, because your data is your utility. The five hundred steps forward, AI included, are waiting on the other side. They are only worth taking once you know where you are standing.

The arc from fifty steps back to five hundred steps forward is not a straight line, but it is absolutely a navigable one.

Figure 05

The step back makes progress possible

---

Sources: AWWA, "Beyond the Replacement Era," 2026 (U.S. drinking water needs of roughly $2.1 to $2.4 trillion over 25 years, against current spending near a third of the annual need). U.S. EPA, "America's Water Sector Workforce Initiative" (about one-third of water and wastewater operators eligible to retire within ten years). U.S. EPA PFAS National Primary Drinking Water Regulation, finalized April 2024.

What to take into your next meeting

Practical steps you can begin with the team you already have.

Email takeawaysShare on LinkedIn

1. Choose one useful step back. Start with a concrete problem such as inconsistent asset names instead of an organization-wide transformation.
2. Keep each asset identifiable. Give each physical asset a unique identity and connect its local names to that record.
3. Capture the reasoning people carry. Ask experienced operators to document what changed, why they acted, and what happened next.
4. Learn through a supervised pilot. Let staff compare AI outputs with operating reality while retaining control of consequential decisions.
5. Measure before expanding. Agree on a baseline and a useful result, then use the evidence to decide what to do next.

One question to ask: Which decision could we improve by understanding our existing records better?

Fifty Steps Back, Five Hundred Steps Forward
Hardeep Anand · The Systems Lens
https://hardeepanand.com/writing/fifty-steps-back/