Latest / The Operations Podcast with Fexingo: Process, People, and Profit in Modern Business / How a Farm Cut Water Use by 40 Percent with AI Irrigation
Transcript
- Lucas: There's a farm in California's Central Valley that grows almonds on four thousand acres. And over the past two years, they cut their water use by forty percent without reducing yield. Luna: Forty percent — that's huge. How? Drip irrigation upgrades? Switching to less thirsty crops? Lucas: Neither, actually. They already had drip lines. What they did was replace their irrigation scheduling — which was based on the farm manager's gut instinct and a few soil moisture probes — with an AI system that decides, tree by tree, when and how much to water. Luna: So it's not about the hardware, it's about the decision logic. Lucas: Exactly. The system uses three data layers. First, there's a network of wireless soil sensors at different depths — four hundred of them across the farm — measuring moisture, salinity, and temperature. Second, they feed in hourly weather forecasts from a private provider, down to the one-kilometer grid level. Third, they use satellite imagery every three days to measure the canopy temperature and the normalized difference vegetation index — that's a proxy for how stressed the trees are. Luna: And the AI takes all that and says: water this block on Tuesday for eighteen minutes, skip this one until Thursday? Lucas: More granular than that. The farm is divided into irrigation zones of roughly forty trees each. The model runs every six hours and outputs a prescription for each zone. It accounts for things like soil type variation across the field — one area might be sandy loam that drains fast, another is clay that holds moisture. The old schedule treated the whole farm as uniform. Luna: So the big win is eliminating overwatering in the clay zones while making sure the sandy zones don't get under-watered. Lucas: Right. And the data shows that before the AI, the farm was over-irrigating by about twenty-five percent on average — just watering everything the same because it was simpler. The farm manager, a guy named Dave, had been doing it that way for eighteen years. He knew the soil, but he couldn't know every tree. Luna: I imagine there was some resistance. You're telling a veteran farmer that a computer knows better. Lucas: There was. Dave told me he spent the first three months fighting the system. He'd override the recommendations about thirty percent of the time. But then they did a controlled test on one hundred acres — half run by the AI, half run by Dave's usual method. They measured water use and yield independently. The AI side used forty-two percent less water and produced the same yield. Dave's side used more water and actually had slightly lower yield because of some overwatering stress. Luna: That's the kind of evidence that changes minds. Lucas: It did. By the end of the first season, Dave was trusting the system about ninety-five percent of the time. He still overrides it during extreme heat events — when temperatures hit one hundred and five, he adds a little extra because the model sometimes underestimates canopy cooling needs. But that's the right kind of human intervention. Luna: So the model got smarter, or the human got more comfortable? Lucas: Both. The AI is a machine learning model that retrains every week on the new data — it learns from its own mistakes. For example, early on it underestimated how much water was needed after a windy day because evaporation rates spike. After a few cycles, it adjusted. Meanwhile, Dave learned to read the model's confidence scores — when the model is ninety percent sure, he goes with it; when it's only sixty percent, he double-checks. Luna: That sounds like a good human ai collaboration model, not just automation. Lucas: Exactly. And the economics are compelling. The farm spent about two million dollars on the sensor network, the software licenses, and the integration. They save roughly six hundred thousand dollars a year in water costs — that's the water itself, plus the energy to pump it. So the payback period is around three and a half years. Luna: And water prices in California are only going up. So that payback probably shortens. Lucas: The farm's water district raised rates twelve percent this year alone. The almond industry is under huge pressure to demonstrate sustainable water use, especially after the drought years. This system gives them a verifiable record — they can show regulators exactly how much water went to each tree and why. Luna: Let's zoom out. This is a farm, but the operational principle — using granular sensor data and machine learning to replace a uniform, intuition-based allocation system — that applies to a lot of industries. Lucas: Sure. A hospital allocating nurse shifts, a warehouse distributing picking tasks, a server farm routing cooling — anywhere you have a variable resource and a fixed schedule that's based on averages, there's probably a similar gain. The key is the data granularity. The farm could have put in one sensor per ten acres and gotten maybe a ten percent improvement. They went to one per ten acres and got forty percent. Luna: But there's a cost to that granularity. More sensors, more data, more complexity. How do you know when you've gone far enough? Lucas: That's the question the farm's CFO asked. They did a pilot with one sensor per acre, then doubled it to one per half acre. The marginal improvement from doubling was only about five percent — not worth the extra hardware and data processing. So they settled at the current density. That kind of marginal analysis is exactly what operations people should do before scaling any sensor system. Luna: So the lesson is: start with a small, controlled test, measure the actual improvement, and scale only the sensors that pay for themselves. Lucas: Right. The farm also learned that not all sensors are equal. The satellite imagery gave them a lot of value for the money — it's a subscription service, no hardware to maintain. The soil sensors, by contrast, require battery changes and calibration. They had about a five percent failure rate per year. That's a maintenance cost they didn't fully anticipate. Luna: Give me a sense of the cultural shift. You've got a workforce of irrigators — people who physically walk the fields and adjust valves. What happens to their jobs? Lucas: The farm didn't lay anyone off. The irrigators became field data technicians. Instead of turning valves all day, they now do sensor maintenance, check for leaks, and handle the physical repairs that the system flags. The farm manager told me it actually made their jobs more interesting — they went from manual labor to problem-solving. One irrigator, Miguel, had been there fifteen years. He's now the go-to person for interpreting the sensor data in Spanish for the rest of the crew. Luna: That's a nice outcome. So the operations redesign wasn't just about technology, it was about role redesign. Lucas: Exactly. And that's a big reason the adoption stuck. If they'd just put in the system and told the irrigators to follow orders, it might have failed. Instead, they involved Miguel and others in the pilot, asked for their feedback, and incorporated their knowledge about which parts of the field drain poorly after heavy rain. The AI model actually got better because of that local expertise. Luna: So the AI didn't replace knowledge, it amplified it. Lucas: That's a perfect way to put it. And I think that's the real operational lesson here — not the tech, but the integration of tech with human judgment. The farm spent two million dollars on hardware and software, but the biggest return came from the change in how decisions were made, not from the sensors themselves. Luna: You know, thinking about the role redesign part, it reminds me of something that might resonate with our listeners who work in operations. A lot of the value from these systems comes from freeing people up to do higher-level work. And if that has moved your thinking forward in some small way, honestly, it's worth a coffee. You can find us at buy me a coffee dot com slash fexingo. Lucas: Yeah, listeners like you are the reason we can keep doing ad-free deep dives. So thanks for that. Luna: Anyway — back to the farm. One thing I want to ask: did they see any unintended consequences from cutting water that much? Like, did the soil salinity change? Lucas: Good question. They monitored salinity closely because with less water, salts can accumulate. What they found was that because they were watering more precisely — not just less — they actually reduced salinity in the root zone. The old over-irrigation was leaching salts away, sure, but it was also waterlogging some areas. The new system maintains a more consistent moisture level that prevents salt buildup at the surface. So that was a positive side effect. Luna: That's interesting. So the system didn't just save water, it improved soil health. Lucas: And that shows up in the tree health data. The satellite imagery over two years shows a steady improvement in canopy density. The trees are less stressed, which also makes them more resistant to pests. The farm has reduced pesticide use by about fifteen percent, though that wasn't a primary goal. Luna: So the benefits compound. Less water, healthier trees, fewer chemicals. Lucas: Right. And that's the kind of system-level thinking that good operations enable. The farm is now looking at whether they can use the same sensor network to optimize fertilizer application. They think they can cut nitrogen use by maybe twenty percent, which would save money and reduce runoff into groundwater. Luna: So this is just the first phase of a broader operational transformation. Lucas: That's exactly how the farm's CEO sees it. He told me: 'We started with water because it was our biggest cost and biggest risk. Now we have a data platform. We can layer other optimizations on top.' Luna: That's a smart approach. Start with the highest-impact variable, prove the model, then expand. Lucas: And they're already talking to other growers in the region about licensing the system. If even a fraction of California's almond acreage — about one point three million acres — adopted this, the water savings would be enormous. In a state that's constantly battling drought, that could make a real difference. Luna: Alright, so to wrap up: the key takeaway for me is that operational improvement isn't just about buying fancy tech. It's about redesigning roles, testing marginal returns, and blending data with human expertise. Lucas: And starting with a problem that actually hurts — in this case, water scarcity and cost — rather than implementing tech for tech's sake. The farm didn't wake up and say 'let's do AI.' They said 'we need to use less water,' and AI turned out to be the best tool for the job. Luna: That's the right order. Thanks, Lucas. Lucas: Thanks, Luna. See you next time.