For a while now I’ve been looking for an AI course.
Not another Udemy playlist on how to prompt ChatGPT. Not a YouTube “10 AI tricks that will change your life.” I have nothing against those, but that is not what I need.
What I wanted was simpler to describe and much harder to find.
How does AI actually transform water on a real plant, with real membranes, real pumps, real chemistry, real operators?
I looked. I didn’t find it.
So I decided we would build it.
We’re releasing the specific AI course, led by Sandro Hansen. He brings the real, hands-on expertise; I help bring our community and audience together; and we’re looking for a few members to join a small first cohort.
Your queries and specific questions are welcome — in fact, we’d prefer it, because we want to focus on tackling what you actually need. And I’ll be in there too, as another member just learning.
⚠️Register of interest form at the end of this publication.
When we agreed to do this, I asked Sandro for two promises:
Keep it human, not technical. I don’t want to spend the course learning Python — my brain is not wired to digest that, and I doubt I’m the only one.
Make it real. Specific cases, clear situations, what to actually do — in language any of us can understand well enough to take action.
Quick note: this is the start of a new format for us. I don’t want to be the bottleneck for those of you who need specific topics drilled into far deeper than I can take you.
This conversation is just the beginning of exploring this topic and sharing why it matters to us. I'd genuinely love to hear your thoughts, experiences, and feedback.
Starting from the plant floor
The honest place to start is operations, because that is where water either works or it doesn’t.
And here I have to be equally honest about myself. I am a civil engineer. My world is Business Development and EPC delivery,
AI in the control room is not my domain of expertise, and I am not going to pretend otherwise.
So I did what I always try to do, I went to someone who actually knows.
I picked Sandro Hansen. I’ve listened to Sandro at events and congresses for a while now, and he has quietly become one of my favourites: someone with genuine, hands-on experience applying AI in water operations, not just talking about it in the abstract.
Just as important for us, Sandro shares the value this whole community is built on — knowledge sharing. Why we both believe this is necessary for so many of us, right now.
“Start now, start small, start with a real problem”
The first thing I asked Sandro was the question I imagine a lot of you are quietly asking yourselves: am I too late? Or is what we see today going to look nothing like what we’ll see in two years?
His answer was that now is a good moment to start.
The technology has been around for a long time. What changed is that computational capacity is finally accessible, and large language models arrived.
He put it in terms I understood: before, to use AI you needed to program, to do data science the hard way — like the old DOS days, where you typed every command into a black screen.
Then Windows appeared, the mouse appeared, and suddenly everyone could enter the computer. LLMs are that moment for AI. You no longer need to be a programmer to interact with the models and get real things done.
But the more important part of his advice was about how to start.
Start now, start small, and — this is the one to underline — start with a real problem.
According to Sandro, roughly 90% of AI failures in our industry come from a single mistake: people using the technology for the technology.
They want to say they’re “doing AI,” so they launch a project without any clarity about what they’re actually trying to solve.
A few months later they announce that AI gave them nothing. But the technology was never the problem.
The question is whether you had a real problem to begin with — and whether you were clear about it.
A while back I ran a poll here asking what an AI course should focus on. One of the options simply carried the word problem, against others like optimization.
Almost everyone clicked problem. We all know we have them. The work is turning that instinct into a well-defined starting point.
AI is not magic, do your engineering homework first
If there is one message from Sandro I would tape above every control room, it’s this:
AI won’t fix a design problem.
His example was beautifully concrete. A filter calculated to pass 300 m³/h will not suddenly pass 4,500 m³/h because you pointed an algorithm at it.
If your plant has a physical limitation, no amount of software solves it. You have to do your homework first.
He’d just visited a plant that did exactly that. They reworked the pretreatment, then moved from 400 m² to 440 m² membrane elements — higher production, lower energy.
Their consumption fell from 3.3 kWh/m³ to 2.8 kWh/m³, and here is the punchline: that was pure engineering, no AI at all.
Then AI comes in — trimming from 2.8 to maybe 2.6, or shifting production to the hours with more solar generation, so the same water costs less.
AI is the last percent of optimization on top of good engineering, not a substitute for it.
It’s (almost) all about data
To optimise any process, you need to see what goes in, what comes out, and how it performs.
For an ultrafiltration system, the outlet turbidity alone tells you almost nothing — you need the inlet turbidity too, because that’s what lets the system reason about whether to do a backwash, a chemical-enhanced backwash, or a full CIP, and which of those is the smartest economic decision at that moment.
And then he made a point that will sting a little for those of us trained to save every euro in a bid: don’t skimp on sensoring.
A pressure switch tells you when you’ve hit a limit. A pressure sensor tells you, in real time, whether fouling is developing faster than planned — and the price difference between the two is not large.
Same story with a level switch versus a level sensor: the switch stops your pump or prevents an overflow, but without continuous level data you can’t close a water balance inside your own plant.
We optimise so hard on capex that we quietly starve our future selves of the data that makes everything else possible.
He also told a story. A plant had its SCADA data retained for only 30 days.
It started as a small site, sensors kept getting added, memory stayed the same — and one day anything older than a month was simply gone.
Before you invest a single euro in “AI,” Sandro’s advice is to make sure the knowledge your plant already generates is stored properly and accessible. That’s business safety.
His sequence, honestly:
Invest in planning first, put in the sensors you’re missing, then wait — collect a real season of data before you expect anything.
A UF cycle might need 45 to 90 days. A Mediterranean or North-African desal plant needs both summer and winter, because train an algorithm on August water and unleash it in January and it will fall over.
Do that groundwork, and a well-planned project can reach payback in around two years. Skip it, and of course the payback never shows up.
So what happens to the operators?
This is the question that comes up every single time, usually with some anxiety behind it, so let me say clearly where I’ve landed after this conversation.
We humans are not good at repetitive tasks. As Sandro mentioned, we get bored, we drift, we make mistakes.
But on the other side, we have something no model has: imagination. Intuition.
That “prickle” at the back of your neck that says something here is strange before any threshold is crossed. No AI is going to replace that.
So here is what we believe actually happens.
AI won’t substitute the operator — it will run in the background, analysing information faster and continuously, and hand insights back to the person doing the job.
The difference won’t be operators versus machines.
It will be operators with a co-pilot versus operators without one.
The co-piloted operator is more productive, makes fewer errors, and — this is the real prize — perceives anomalies before the alarm goes off.
The plant becomes more efficient and more effective, and the human stays firmly in charge.
Sandro sees the same shift in the tooling. We’ve moved from AI that only analyses to agents that can act — always within limits you define.
You might let an agent nudge chemical dosage within a set range if it decides that’s warranted, while you keep control of everything else.
And increasingly the interface is a conversation: an operator starts a shift and asks the system where the two most important problems are today, or “I need 10% more water tomorrow — which train, in the most economical way?”
Watch the pressure climb on an RO and ask why, and the model might answer that it’s 70% biofouling, 30% scaling — so you know where to look first.
As Sandro put it, this is not the future. It’s already happening. ⏰
This is exactly the knowledge gap
Regular readers know I keep returning to one theme: the water sector’s problem is not only an investment gap, it’s a knowledge gap.
Sandro reached the same conclusion from the operational side without me prompting him.
The reason projects so often start and finish with an outside consultant — never taking root in the plant — is that the internal team was never given enough knowledge to carry them.
He said, plainly, that you cannot find all of this in one place: you have to stitch it together from data science, from storage, from sensoring, from the algorithms, from the water itself.
That scattered, un-findable knowledge is precisely the course we’re building, a practical middle level that takes you from the designing engineer through construction, operation, and the supervisors and managers who have to plan it all.
Enough to make a good decision.
The invitation
We’re opening the first cohort now, and we’re keeping it small on purpose.
If you want in, the best thing you can do is send us your real problems and specific questions. Sandro and I genuinely want to build the sessions around what you need to solve, plant by plant, rather than around what we assume.
I’ll be in the room too, by the way. Not as a host performing expertise, but as another member, learning alongside you.
One last note about where this is going. This is the first of a new format for The Water MBA: courses led by real experts, on the specific topics you’re hungry for.
I don’t want to be the bottleneck. There are subjects many of you need drilled into far deeper than I can take you — so from here on, when the right expert and the right demand meet, we’ll bring them together.
That’s the whole point of this community. This is simply the first one.
🚨If you're interested in joining this course, please submit your expression of interest (below) along with some key information, and we'll get back to you as soon as possible. Let’s see what we get.
A few quotes
A few lines from the conversation worth keeping.
On when to start
“Now is a good moment. The technology has been around for years — what changed is that you no longer need to be a programmer to use it.”
On where to begin
“Start now, start small, and start with a real problem. The technology will work. The question is whether you have a real problem to solve — and clarity about it.”
On why projects fail
“Around 90% of failures come from launching an AI project just to say you’re doing AI, without any clarity about what you’re solving.”
On engineering first
“AI won’t make magic. A filter calculated for 300 m³/h will never pass 4,500. Do your homework before you reach for the software.”
On the real prize
“They went from 3.3 to 2.8 kWh/m³ with pure engineering, no AI. Then AI takes you from 2.8 to 2.6.”
On data
“If you don’t have data, you can’t do anything. And to optimise, you need to see what’s coming in — not just what’s going out.”
On sensors
“People try to save money on sensoring. But that’s exactly what steals the biggest benefit from AI later.”
On planning and payback
“Invest first in planning. A well-planned project pays back in about two years. When people say the payback never came — well, how did they start?”
On operators
“AI won’t substitute the operator. It runs in the background and hands the insights back. The difference isn’t operators versus machines — it’s operators with a co-pilot versus operators without one.”
On the co-pilot’s edge
“The real prize is perceiving the anomaly before the alarm goes off.”
On agents
“We’ve moved from AI that analyses to agents that act — always within the limits you define. This is not the future. It’s already happening.”
On the knowledge gap
“You can’t find all of this in one place. That’s why projects start and finish with an outside force — the internal team was never given enough to carry them.”
