AI Agents in Golf Course Maintenance: The Sensor Comes First

Before an AI agent can watch a golf course, something on that course has to be measuring it, and at most facilities nothing is. In-ground soil moisture sensors sit at 3% of US golf facilities and have not moved in fifteen years, according to the water survey the Golf Course Superintendents Association of America runs on its own membership. The instrument comes before the agent. That ordering is the whole of the problem, and most writing on this subject gets it backwards.

The backhoe shows up on a Friday afternoon. A sprinkler head on the 7th fairway has been leaking for three weeks, and now the valve box is under six inches of water. Nobody knew until a member drove a cart through it.

The real cost is not the backhoe. It is three weeks of wasted water, the weak turf that will need reseeding, the cart damage and the member who spent Thursday morning waiting while the pro shop tried to find someone to fix it. The repair is the smallest part of the bill.

The usual next paragraph says the data was there all along and nobody looked. Almost always, it was not there.

What golf courses actually measure

The GCSAA has surveyed irrigation scheduling practice at US facilities three times, in 2005, 2013 and 2020, and published the series in its Golf Course Environmental Profile. Here is what the survey found.

That table is from the Phase III water report, and the 2020 column is the most recent complete profile of scheduling practice in the survey series.

The 2024 wave moves two of those lines and confirms the rest. Hand-held moisture sensors passed half the profession for the first time, at 55%, and scheduling from an on-site evapotranspiration station rose to 21% against 14% in 2005. That analysis was published by J. Bryan Unruh and Travis Shaddox in Golf Course Management in December 2025.

Read the two instruments that produce a continuous, machine-readable stream. On-site evapotranspiration stations: four facilities in five have none. In-ground soil moisture probes: three facilities in a hundred, a number that has not moved since the survey began. Drones: one in a hundred. Meanwhile more than half the profession still schedules irrigation partly by looking at the ground, which is a skill rather than a data source.

Two caveats. This is a survey of US facilities, and no comparable census exists for Britain, Ireland, continental Europe, Australasia or the Gulf, so a reader elsewhere should treat the shape as indicative rather than local. And it measures what superintendents report scheduling from, not what a facility owns.

The implication for anyone selling monitoring software, and for anyone buying it, is the same. The bottleneck is not attention. It is that at most clubs there is nothing generating a signal for software to read.

The instrument the profession has already adopted

There is an encouraging half to this, and it is the reason the article is worth writing at all.

Hand-held soil moisture meters went from 29% of facilities in 2013 to 39% in 2020 to 55% in 2024, on the same survey instrument. That is the fastest adoption of any scheduling method the GCSAA has tracked. And in the Phase III report, 92% of facilities using moisture sensors said the impact on their operation was positive.

Superintendents adopt instruments that earn their place. They adopted this one quickly, on their own budgets, without a software vendor involved. Anyone who tells you the profession is resistant to technology has not read the series.

The hand-held meter is also the ceiling, and that is the interesting part. A reading taken with a hand-held probe exists for about four seconds, in one person's hand, on one green, and then it exists in that person's memory or on a clipboard. It cannot be trended. It cannot be compared with last August. No software can watch it, because it was never a stream. The profession has proved it will pay for the measurement. What it has mostly not bought is the part that keeps the measurement.

That gap is where an agent enters, and it is a much narrower opening than the category's marketing suggests.

Irrigation, and the sentence the USGA has already written

The strongest claim in this whole subject was made by the governing body five years ago and in ten words. From the USGA Green Section Record of 18 December 2020: pump failure is usually avoidable, or at least predictable.

That is worth sitting with, because it concedes the entire premise of predictive maintenance from a source with nothing to sell. The USGA's route to it, though, is not the one a software company would draw. The December 2020 issues prescribe routine performance testing and maintenance, carried out on a schedule, with vibration analysis as the diagnostic. Serious problems come from neglecting irrigation and well pumps until a catastrophic failure occurs, and the guidance is that testing and maintenance reduce that risk and extend pump life.

Continuous monitoring is the same insight applied every day rather than every few years. Pressure sensors, flow meters and current transformers on a pump motor produce a trend, and a trend is what makes an anomaly visible. Rising power draw against falling hydraulic output points at mechanical drag rather than a hydraulic change, which is a coherent thing to watch.

Now the honest part. No published work benchmarks continuous pump telemetry against the USGA's testing regime on golf courses, and nobody can tell you how much earlier the sensor sees it than the annual test does. Anyone quoting you a lead time in weeks has invented it. The case for telemetry is that it is cheap relative to a pump station and that it removes the failure mode where the scheduled test is the thing that gets deferred, and that case does not need a fabricated number to stand up.

The same water report gives the size of the gap it would fill. Only 15% of US facilities have a preventive irrigation maintenance plan at all. In the Southwest region it reaches 25%, and every other region is lower. That single figure says more about the state of maintenance planning than any assertion about what most superintendents do with their day, and it is measured rather than asserted.

Signature-level diagnostics deserve more caution than they get. Pressure falling while flow holds steady is consistent with rising friction loss, and so with a partial obstruction. That is an example of the kind of correlation worth trending. It is not a published diagnostic rule for golf irrigation.

Turf health, and what a real disease model does

Start by unlearning something the source version of this article had backwards, because the physics matters for anything built on top of it.

Warmer turf means drier turf, not wetter. When a plant runs short of water, its stomata close, transpiration stops and the evaporative cooling that came with it stops too, so canopy temperature rises above that of well-watered turf nearby. Thermal imaging research on turfgrass has measured drought symptoms coinciding with a surface temperature increase of roughly 1 to 7 °C, or 2 to 13 °F, depending on solar intensity. That is the basis of the crop water stress index and of every thermal drought-detection method built on it, and it is documented in the small-unmanned-aircraft thermal imaging work of Hong and colleagues in 2019 and in the 2021 field radiometry study in Agronomy.

It also tells you when to look. A thermal signal exists because transpiration differs between stressed and unstressed turf, so the useful window is the middle of the day, when the canopy is loaded and transpiring hard. At sunrise the mechanism that produces the signal is switched off.

For disease, there is a validated model, it is public, and naming it is more useful than any amount of talk about correlating data streams. The Smith-Kerns dollar spot model, developed by Damon Smith and Jim Kerns and hosted by the University of Wisconsin-Madison Turfgrass Diagnostic Lab, runs on two inputs: a five-day moving average of daily relative humidity and daily average air temperature. Nothing else. It was validated through years of field research, primarily in Wisconsin but also in Oklahoma, Pennsylvania, Mississippi, Tennessee, Connecticut and New Jersey.

Three things about it discipline the whole conversation.

It is site-level. It runs on weather at your location and produces one number for the property. No published model resolves dollar spot risk to a range of holes, and an alert that claims to is telling you something its inputs cannot support.

It is a daily probability derived from a trailing average, not a forecast window. There is no seventy-two-hour horizon in it.

And the lab is explicit that the model does not tell users when to spray. It gives a probability. The published starting point is a 20% threshold on creeping bentgrass, which each course is expected to calibrate to its own tolerance and its own history, and the model is treated as inactive below about 10 °C or above about 35 °C.

So the agronomy is not the software's job, and the model is not the software's insight. What software is genuinely good at is dull and valuable: pulling the weather feed every day without fail, running the model against it, holding the trailing five-day window, remembering what the club decided its threshold was, and telling a named person on the morning the threshold is crossed. A superintendent will not forget to do that in June. It is late August, after eleven weeks of it, that a daily discipline quietly becomes a weekly one.

Imagery is further from deployment than the drone photography in vendor decks suggests. The most credible current work is WinterTurf, a USDA-NIFA funded project running from 2021 to 2026 out of the University of Minnesota with partners including Wisconsin, Michigan State, Iowa State, Oregon State, Rutgers and the Norwegian Institute of Bioeconomy Research. In work published on 8 May 2025 by Ce Yang, Xuechen Li, Alireza Sanaeifar, Aleksei Rozanov and Bryan Runck, multispectral drone imagery was analysed with an AI segmentation model, integrating NDVI, GNDVI and NDRE to flag variation in turf health, alongside custom sensor nodes monitoring conditions on greens.

Read the scope, because it is narrower than the enthusiasm around it. That is winter damage on greens in northern climates, not a general disease-detection capability, and with drone-based scheduling reported at 1% of facilities in the same survey series it describes something one course in a hundred is positioned to do at all. Excellent research, and not yet a product decision.

Mowing, where the machines are actually running

Robotic mowing is the one area where the equipment has genuinely arrived, and it is also where the sourcing needs the most care, because almost everything published about it originates with a manufacturer.

Golf Sustainable reported in August 2025 that Husqvarna is represented on more than 130 golf courses in Germany, and that around 30% of those courses mow their fairways using robots alone. Those figures come from Husqvarna, the manufacturer, and appear in an article about a study Husqvarna co-funded, so read them as a supplier's account of its own installed base rather than as neutral market data.

The study in question is ROBO-GOLF, run by the Norwegian Institute of Bioeconomy Research under Trygve S. Aamlid and funded by the Scandinavian Turfgrass Environmental Research Foundation together with the Husqvarna Group. Its finding on quality is the one worth having, with that funding disclosed: daily cutting produced better grass and fewer weeds on the fairways, and golfers described the surfaces as more even and dense.

That is the real argument for autonomy, and it is not the labour argument. A robot mows the same pattern at the same time every day. It does not rush to finish before an outing goes off. The consistency is the product.

On equipment classes, precision matters more than brand. Toro, another manufacturer and therefore another interested party, publishes an autonomous range that separates by job: a fully robotic line aimed at rough mowing with satellite and real-time-kinematic navigation, a distinct autonomous fairway mower and a greens mower that both retain manual operation, and a separate machine for picking the driving range. Fully robotic and operator-optional are different purchases with different labour implications, and a fairway machine is not a rough machine. An equipment manager will know this in the first minute of a demo.

The USGA has been more careful than the manufacturers, and recently. Its Green Section Record of 4 March 2022 records interest in autonomous mowing growing as the technology advances and labour pressure persists, and then asks the question that has not been answered: whether the technology will be viable at the smaller scale of a golf operation. Hold both positions at once. The machines work. Whether they work economically at eighteen holes with your labour market and your fairway acreage is a question your own numbers answer, not a question the category has settled.

The 31% water figure, and the part nobody quotes

Golf's most-cited environmental number needs its footnote attached every time.

The GCSAA announced on 30 December 2025 that US golf courses used 31% less water in 2024 than in 2005. The survey covered nearly 1,700 facilities. It was analysed by Travis Shaddox of Bluegrass Art and Science and J. Bryan Unruh of the University of Florida with the National Golf Foundation, and submitted for peer review to HortTechnology. That is a serious instrument, seriously analysed, and the figure is real.

Here is the sentence that almost never travels with it. Roughly one third of the national reduction in irrigation volume is attributable directly to facility closures. Courses that shut down stopped using water. That is arithmetic, not agronomy.

The remaining two thirds came from efficiency, and the practices credited are worth reading in order. GCSAA's own release lists wetting agents, hand watering and maintaining turf under drier conditions first, with hand-held moisture sensors, irrigation audits and on-site evapotranspiration data as contributors. Sensing helped. It was one input among several, and the leading practices were agronomic decisions made by people.

Anyone selling you monitoring software on the back of that 31% is quoting a number that is a third demolition and two thirds superintendents changing how they water. The honest version still supports the case, because a third of the efficiency gain came from decisions that better measurement makes easier. It just does not support the version where sensors did it.

What does connecting any of this to your club system actually change?

Here is where an article like this normally claims that a unified platform solves the problem, and the honest claim is much smaller than that.

Once an instrument exists on the course, the marginal cost of watching it is an integration. Not a capital project, an integration: one authentication, one data format, one thing that breaks when a vendor changes an API. The number of integrations is the number a general manager can actually forecast, and it is the number that decides whether the second sensor and the third ever get added. Clubs that run one system add streams one at a time. Clubs that run six find every addition needs its own negotiation, and mostly stop after the first.

Now the boundary, stated plainly, because this is an industry analysis and not a product page. Links Meridian holds club-side operational data: the tee sheet, point of sale, member records, billing. It does not read irrigation controllers, soil probes, weather stations or mower telemetry, and nothing in this article should be taken to say otherwise. Anyone in this category who implies that agronomic telemetry is already flowing into their club management system should be asked to name the controller makes and models, and to do it in the demo.

What is true, and what the two halves of a club have in common, is that a maintenance decision becomes an operations event within the hour. A hole comes out of play and the tee sheet has to be reshaped, the affected players have to be told, and the revenue consequence lands in somebody's month-end report. That half is club-side, it is where a management platform lives, and it is the part of this problem that is solved today. The agronomic half is a sensor and an integration that most clubs have not bought yet.

Is your club instrumented enough for any of this to be worth buying?

Five questions, answerable from where you are sitting, before anybody quotes you for anything.

1. What on your course produces a reading more than once a day without a person walking to it? If the answer is the irrigation controller and nothing else, that is your starting inventory and the honest scope of any monitoring conversation this year.

2. Does your irrigation controller expose its data to anything, or does it only expose it to itself? A controller that logs internally and shows a screen in the shed is not a data source until something can read it out.

3. Would your superintendent describe maintenance planning at your club as mostly scheduled or mostly reactive? Only 15% of US facilities report having a preventive irrigation maintenance plan, according to the GCSAA water report, so the answer many clubs give is the common one rather than a failing.

4. Pick your own number for a maintenance surprise that hurts, in your own currency. Now count how many times you were surprised by one last year. Two or more and the gap is visibility, not budget, and the cheapest fix is instrumenting one system rather than all of them.

5. If you added a sensor tomorrow, what would receive its output, and who would read the alert at half past six in the morning? An alert with no named recipient is a notification nobody acknowledges by the third week.

A club that answers question one with "nothing" has not found a software problem. It has found that the first purchase is a probe, a gateway and somewhere for the readings to land, and that is a smaller and more defensible first spend than any platform.

Where this leaves a general manager

The version of this story usually told is that courses are drowning in unwatched data. The census says otherwise, and the superintendent in the room already knows it.

The version that is true is more useful anyway. Almost nobody is instrumented. The profession has already shown it will buy an instrument that earns its place, because it did exactly that with hand-held moisture meters inside a decade. The governing body has already said pump failure is usually predictable. There is a validated disease model in the public domain that any club can run for the cost of a weather feed. What is mostly missing is the unglamorous layer that keeps a reading instead of taking one, and tells a named person when a threshold the club set for itself has been crossed.

An AI agent in this context is software that watches a data stream, detects a deviation from a baseline and escalates it to the person who can act. It should not be sold as anything more autonomous than that, and on a golf course, where the decision is agronomic and the consequences are seasonal, the escalation is the point.

So start with the measurement, not the intelligence. Instrument the pump station, because it is the most expensive machine on the property and the one the USGA has already told you is predictable. Run the dollar spot model against a site weather feed, because it costs almost nothing and it is real. Then, only then, ask what software should do with the streams you have created. Buying that in the other order is how a club ends up with a dashboard that has nothing to show.