The farm is not a demo: which agricultural technologies earn their place?

By Jordan Gold and Valery Gold

An evidence-led technology review for Legends Only Racing.

Research cutoff: November 7, 2026 (CST). Observation periods differ by study. This is a selected narrative review, not a systematic review, product ranking or complete survey of equipment available in 2026.

The most important question about a farm technology is not whether it can recognize a weed, draw a beautiful map or drive a straight line. It is whether the whole operation works better when the weather changes, the crop closes its canopy, a component breaks and the invoice arrives.

That is a higher bar than a demonstration. It is also a more hopeful one. Farmers deserve tools that respect their knowledge, protect their independence and leave land and water in better condition—not another obligation dressed up as innovation.

This review examines four connected areas: guidance and variable-rate inputs; sensing and irrigation; targeted weed control and robotics; and artificial-intelligence decision support. It draws on selected government and extension research and peer-reviewed field studies, not an exhaustive systematic review. The examples concern arable systems in the United States, Germany and Australia, including commercial-field strips and research-station plots; these are not equivalent tests of whole-farm commercial performance. They do not establish results for every climate, crop or smallholding. Sources are linked below; supporting passages are retained in our editorial verification records.

The short answer is conditional: guidance can reduce measured coverage waste, deficit irrigation can reduce applications, and targeted spraying can reduce herbicide use. None of those measurements alone proves a profitable purchase or ecological recovery. The failures are as instructive as the successes. S1 S2 S3 S4 S5 S6

The central distinction is between doing a task more precisely and knowing which task should be done. A controller can place fertilizer accurately while following a poor prescription. A sensor can measure something real without identifying a profitable intervention. The evidence repeatedly makes that distinction consequential. S1 S2 S3 S4 S5 S6

1. Guidance: start with an observable problem

Automatic guidance has an unusually tangible proposition: reduce unnecessary overlap and missed ground. Unlike a yield forecast, its immediate output can be checked against the path the machinery actually traveled.

In central Iowa, Matt Darr’s extension account describes monitoring more than 2,500 acres of field operations in planting and spring tillage during 2011 on two producers’ operations without precision-agriculture equipment. Average planting overlap was 3.3%; spring-tillage overlap was 7%. These were measurements of an existing problem, not a randomized demonstration that buying autosteering would eliminate it. The planters and tillage implements were large, and the sample was small. The account’s dollar calculations used historical input and crop-price assumptions; they are not current equipment quotations or universal payback estimates. S1

A separate peer-reviewed study by Tulsi P. Kharel and colleagues compared fertilizer application on six fields and spraying on four fields with and without tractor guidance. USDA’s technical abstract reports an average spatial-coverage efficiency improvement of 8%, with overlap reductions up to 6% of field area and gap reductions up to 16%. Those maxima are not average savings in fuel, fertilizer or labor. Nor does an 8% coverage-efficiency gain establish an 8% yield increase. Our access for this study was the USDA abstract and interpretive summary, not the publisher’s full methods; that limits how confidently we can assess trial allocation and generalizability. S2

The practical lesson is to measure the baseline before choosing the receiver. Separate pass-to-pass guidance from automatic section control, which switches parts of an implement off where treatment is already complete. Headlands, field shape, implement width and the operator’s existing accuracy matter. Darr’s account identifies different causes of planting and tillage overlap; a single assumed savings rate would hide those differences. S1

For an operator, reduced fatigue is a legitimate reason to value assistance. It should be recorded as an operator benefit, however, rather than silently converted into an unmeasured cash saving. USDA’s adoption research documents fatigue reduction as a motivation, not an experimentally measured return. S8

2. Variable-rate inputs: a map is not a prescription

Guidance asks where the machine should go. Variable-rate technology asks how much to apply at each location. The second question requires more agronomy.

USDA economist David Schimmelpfennig’s 2016 report used U.S. corn-farm survey data from 2010 and a treatment-effects model intended to address selection into adoption. Table 10 reports modest estimated percentage changes in net returns: 1.5% for guidance and 1.1% for variable-rate technology. These are not additions of 1.5 or 1.1 percentage points to a farm’s profit margin. These are historical model-based estimates from observational farm data, including overhead—not randomized causal effects, current returns or guaranteed margins. The report notes that capital requirements can offset gains. S3

The most instructive counterweight comes from Australian nitrogen research. A. F. Colaço and colleagues compared thirteen approaches, including four reference/benchmark methods and nine digital alternatives, using twenty-one commercial-field wheat and barley strip trials conducted during 2018–2021 across four states. Their participating farmers were already strong managers. These were farmer-compatible adjacent strips, not randomized, replicated rate experiments. Economically optimal rates were estimated from quadratic partial-profit curves using three nitrogen rates; they were fitted benchmarks, not an independently known agronomic truth. Most sensor-based methods did not outperform farmer management. Recommendations with substantial error could perform better when averaged across a field than when applied site-specifically. In other words, finer control can distribute a bad recommendation more elaborately. S4

That finding does not prove variable-rate fertilizer never works. It concerns particular nitrogen methods, crops, sites and participating growers. It does not test every approach to lime, phosphorus, seed or other inputs. Nor should older U.S. observational estimates and newer Australian field-response analysis be forced into a single verdict: their questions, baselines and methods differ. S3 S4

Before paying for variable-rate capability, ask whether management zones have distinct, repeatable responses to the input—not just different colors on a yield map. Ask what evidence links the proposed rate to that response. Retain an appropriate uniform-rate comparison and evaluate harvested output, input quantity and the cost of making the prescription. These are our recommendations, informed by the field study’s warnings about recommendation error and the limitations of extrapolating beyond trial strips. S4

A prescription that raises input use can still be economically rational; one that lowers it can still lose money if it sacrifices valuable production. Resource efficiency and profit must therefore be reported separately, with the environmental objective stated openly rather than inferred from the word “precision.”

3. Sensing and irrigation: useful feedback, conditional savings

A sensor becomes valuable when it changes a decision at the right time. Cotton research from Bushland, Texas, provides a useful example—and an equally useful warning about attribution.

Susan A. O’Shaughnessy, Paul D. Colaizzi and Craig W. Bednarz tested an automated scheduling system in 2021 and 2022 on a research-station cotton field. The cropped half-pivot covered 10.5 hectares. Treatments were randomized with four replications. For the hybrid scheduling method, one soil-water station supplied each irrigation level’s four plots, rather than an independent soil-sensor station in every replicate. Plant-temperature feedback and a hybrid incorporating soil-water measurements were compared with manual scheduling based on neutron-probe readings. S5

At the treatment level designed around 75% replenishment, the authors reported irrigation savings of at least 20% in 2021 and 16% in 2022 relative to fully irrigated plots, with no statistically significant seed-cotton yield penalty. Crucially, manual deficit scheduling also achieved savings: the authors report 24% in 2021 and 16% in 2022 for M75 relative to the fully irrigated M100 control. Seed-cotton yields did not differ significantly among scheduling methods at the same irrigation level; some irrigation amounts and evapotranspiration comparisons did differ. “Crop response was the same” would be too broad. It would be wrong to attribute the entire water reduction to automation, or to claim that adding soil sensors necessarily beat careful manual management. S5

“No statistically significant difference” does not prove identical yields or eliminate downside risk. The study reports variability and year-specific tests rather than a portable confidence interval for a farmer’s future savings. Rainfall timing differed substantially between seasons; a pump variable-frequency-drive failure affected operations in the second year. These are not embarrassing footnotes. They are part of what a working technology must survive. S5

The case for automation here is partly operational: potentially replacing a demanding manual measurement and calculation routine while supporting deficit irrigation. The authors describe the neutron-probe method as labor-intensive and expensive, but this review does not convert that description into measured hours saved or a commercial payback period. The system’s developers participated in the research; this is public-sector field evidence, not an independent commercial product comparison. S5

For purchasing decisions, separate sensing from actuation. Can the existing irrigation equipment respond to the recommendation? Who checks installation, failed communications and implausible readings? What happens during cloud cover or missing data? The Texas study explicitly discusses redundancy and fallback behavior; the sensible buying question is how those safeguards translate to the proposed installation. S5

Less irrigation is not automatically more water for the river

Farm-level application and basin-level conservation are different outcomes. Frank A. Ward and Manuel Pulido-Velazquez’s Upper Rio Grande basin analysis explains the accounting problem: more efficient irrigation can reduce valuable return flows and aquifer recharge, so reduced applications do not guarantee reduced depletion. Its abstract describes an integrated policy analysis, not an observed sensor-adoption experiment. We independently verified the abstract and bibliographic record, not the full paper; no numerical model prediction is used here. The mechanism warrants caution, not a claim that every irrigation upgrade worsens scarcity. S10

Our recommendation is to report both the farm’s metered withdrawal and the relevant wider water-accounting boundary. A pump-energy saving is valuable. An ecological conservation claim additionally needs attention to evapotranspiration, return flows and whether saved capacity supports expanded production. Caring for a watershed requires asking who receives—or loses—the water downstream. S10

4. Targeted weed control: the timing matters as much as the camera

Among the studies reviewed, targeted spraying offers a particularly clear example of real savings and real failure in the same experiment.

Alicia Allmendinger and colleagues conducted four maize experiments on farms in southwestern Germany in 2023. Each used a randomized complete-block design with four repetitions; individual plots were 3 by 10 meters. Drone imagery generated weed maps for subsequent spot spraying. Early post-emergence spotting saved 47% relative to the corresponding broadcast application. Savings were calculated from applied spray volume at the study’s specified mixture and rate—not from a toxicity-weighted indicator or a whole-farm annual pesticide budget. Weed-control efficacy was 86% for early spotting versus 76% for early broadcast; the difference was not statistically significant. Here efficacy meant the change in weed density from before treatment to 14 days afterward, not the percentage of weeds correctly recognized in an image. S6

Later treatment was a different story. Late spotting achieved only 20% weed-control efficacy and 12% herbicide savings. The authors identify crop leaves obscuring newly emerged weeds after rainfall as the likely explanation; they also discuss uneven dose distribution at spot edges as a possible contributor to poor late control. The attractive early result was therefore conditional on an agronomic window—not simply ownership of a high-resolution camera. Maps were prepared two to four days before application, introducing a practical interval between seeing the field and treating it. S6

Silage yield did not differ significantly among treatments, including the untreated control. The authors suggest weed competition was too limited to cause a significant yield loss in that season. That weakens any inference that equal yield proves durable protection under heavier weed pressure. The university-led study reports public funding and no declared conflicts; a commercial imaging provider generated the UAV images. Its methods also report removing outliers and pooling experiment means, and raw data are available only on request. Those features do not invalidate the result, but four small-plot experiments in one region and year are not an independent, multi-season operating record for a whole commercial fleet. S6

Reduced spray volume is also not a measured biodiversity recovery. This study measured weeds, spraying and crop outcomes—not long-term effects on beneficial organisms, herbicide resistance or whole-farm emissions. The environmental promise should motivate those measurements, not replace them. S6

Robots do not abolish capacity constraints

Roland Gerhards and colleagues evaluated seven robotic weeding systems in five randomized-block experiments with four replicates, in sugar beet and winter oilseed rape at two southwestern German locations in 2021 and 2022. Systems differed in autonomy and mechanism; not every system appeared in every experiment. Table 5 reports roughly 92–94% mean weed control for several in-row systems. Unlike the maize paper, this study calculated efficacy against weed density in untreated plots, so the percentages are not directly interchangeable. S7

Reading beyond the abstract changes the crop-safety story. The abstract’s reassuring description of mostly low crop losses omits a substantial Farmdroid loss reported in Table 5 and the results. The Farming Revolution W4 sugar-beet treatment showed 39.5% crop stand loss relative to the untreated comparison, after frost had already reduced stands across treatments. Table 5 also gives the Farmdroid hoeing-only treatment a 20.7% mean crop stand loss, with a 0.9–40.4% range. Other passages give different oilseed-rape loss figures that this review cannot reconcile without underlying data. We therefore do not offer a definitive crop-loss ranking. The supported conclusion is narrower: high weed removal did not guarantee crop safety. S7

The cost comparator also needs correction. Table 6 gives camera-guided inter-row hoeing with finger weeding a calculated €220.8 per hectare; inter-row hoeing alone is listed at €102.5 for sugar beet and €205 for oilseed rape. In-row systems cost about €555–804 per hectare, versus approximately €307–383 for broadcast herbicides. These are historical study-calculated treatment costs, not audited farm invoices, current equipment prices or purchase costs. The calculation includes herbicides, fuel, depreciation, labor, interest and maintenance, with some robot costs estimated. For autonomous operations it assumes one operator can supervise five robots, allocating only 20% of labor cost to each—not a demonstrated commercial staffing ratio. S7

This was academic field research involving a sugar-company research-farm coauthor and commercial equipment support; it should not be called wholly independent of industry. It reports useful tradeoffs in crop losses and cost, but its internal inconsistencies make table-level checking essential rather than optional. S7

An operator evaluating a robot should demand hectares completed within the actual weed-control window, including transport, setup, charging or refueling, supervision, breakdowns and rescue treatments. Those are proposed acceptance criteria, not quantities established across the machines in this review. Peak-season capacity matters more than an impressive uninterrupted demonstration.

A wider check reinforces the need for complete economics. A 2026 review by Vladyslav Pitsyk, Johanna Pfrombeck and Markus Gandorfer synthesized 27 economic evaluations of site-specific weed management. It found that technology or management costs were often omitted or understated, and that real-world multi-year evaluations remained scarce. It also identified more recent positive results alongside missing biological and longer-term costs. This is a secondary methodological synthesis, not 27 newly verified experiments or a pooled estimate of savings; we read the review, not every underlying paper. It is reason to test costs and delayed weed escapes, not to declare the technology unprofitable everywhere. S13

5. AI decision support: the best-looking result needs the strongest qualification

The Australian nitrogen study also provides this review’s clearest lesson about AI. Its best-performing machine-learning approach reduced root-mean-square recommendation error from 42 to 15 kilograms of nitrogen per hectare in a “data-abundant” scenario. That result was not evidence that a generic AI adviser had learned to run an unfamiliar farm. S4

The scenario randomly split observations for training and validation, so the training data included information from the same sites and seasons. The authors explicitly describe it as a hypothetical, well-informed upper-potential case. When validation instead relied on other sites—the “data-limited” scenario—performance was much less compelling. Many observations also came from spatially overlapping analysis windows, not independent farms. S4

The often-quotable gain of up to A$47 per hectare needs the same care. It was expected partial profitability derived from fitted field-response functions; the paper expressed the comparison using an assumed A$1,000-per-hectare normalized-profit baseline. It was not a measured, booked cash gain from deploying a commercial service across farms, and it did not establish net return after purchasing an AI system. Recommendations were evaluated within trial strips, not across entire fields. S4

This is promising research precisely because the authors explain its constraints. The useful question is whether a model has been tested on sites and seasons it has not effectively already seen, against the farmer’s existing practice, with implementation costs included. Better predictive accuracy does not by itself establish a better management decision: the same paper found that improving yield prediction through machine learning barely improved a particular nitrogen recommendation framework. S4

Our recommendation is supervised decision support with an explicit boundary of validity, preserved input data and an auditable recommendation history. Require a human decision-maker to approve consequential changes until local evidence justifies more autonomy. This review found no basis in its selected studies for replacing agronomic judgment with a general-purpose chatbot, or for promising savings from a model’s test-set score.

6. Who controls the machinery—and the information?

A farm can become more precise and less independent at the same time. That is a procurement risk, not an allegation of misconduct.

The U.S. Government Accountability Office identifies high acquisition costs, farm-data ownership concerns and missing interoperability standards as barriers to adoption. It calls for understandable, accountable data terms and considers standards and governance among possible policy responses. These are government assessment findings, not measured yield effects. S9

Before signing, request written answers: Can raw observations, prescriptions and machine logs be exported in usable formats? Who may use them for model training or share them onward? What survives cancellation? Can another adviser read the files? Can the machine perform essential tasks if the subscription or connection fails? These are our due-diligence recommendations, not an assertion that every vendor denies those rights. S9

Repairability belongs on the same checklist. The Federal Trade Commission’s 2021 Nixing the Fix report identifies restricted access to parts, manuals and diagnostic tools, and software-related constraints. It found scant evidence supporting manufacturers’ broad justifications for repair restrictions. This was a cross-industry report, not a ruling that a named agricultural supplier acted unlawfully, and not an inventory of current state repair rights. S11

Ask who can diagnose, calibrate and restore a machine during a critical operating window. Demand clarity about documentation, parts availability, software support and safe independent service. Evaluate cybersecurity protections on their merits rather than treating either “security” or “right to repair” as a reason to avoid specific technical questions.

Cybersecurity is an operating issue, not merely a privacy policy. The FBI’s 2022 notification described ransomware affecting agricultural cooperatives and warned about vulnerability around planting and harvest. That is direct evidence of supply-chain disruption, not a quantified probability that a particular smart sprayer will be hacked. Its recommended measures include offline backups, segmentation, patching, multifactor authentication and a plan for manual operations. S12

Finally, access matters. USDA’s report using 2023 farm data shows guidance use on 9% of small, 52% of midsize and 70% of large-scale crop-producing farms; variable-rate use was 5%, 32% and 45%. These are shares of farms within economic-size categories, not acreage shares; the crop-technology estimates exclude farms that harvested no cropland. The family-farm categories use gross cash farm income: under $350,000 for small, $350,000–$999,999 for midsize, and at least $1 million for large-scale farms. “Small” here concerns gross cash farm income, not a universal acreage threshold. Adoption differences do not prove small farmers are uninformed or that adoption is always desirable. S8

Shared equipment, contractor services and cooperative data arrangements deserve evaluation as alternatives to ownership—not unsupported promises of instant affordability. Public-interest research should measure whether such arrangements deliver timely service, fair terms and usable local knowledge to farms that cannot carry a large capital investment. S9

7. An operator’s investment test

Our evidence-led recommendation is to buy a measurable improvement, not a category of technology.

Define the problem. Is it overlap, mistimed irrigation, weed escapes, labor bottlenecks or uncertain nitrogen response? Establish the baseline in physical units and operating hours. Do not budget a yield gain just because a demonstration looked cleaner.

Trial the decision, not only the device. Use matched or randomized comparisons where feasible, repeat across representative conditions and record crop quality as well as yield. Keep a failure and rescue-treatment log. The German canopy failure and the Australian out-of-site validation gap show why convenient test conditions are insufficient. S4 S6

Separate capital from recurring expense. Obtain a dated, itemized quote covering hardware, integration and installation. Then identify subscriptions, positioning services where needed, connectivity, sensing consumables, calibration, training, insurance, repairs, supervision and specialist analysis. Compare ownership with a service quotation. These are budgeting categories, not asserted universal charges.

Use a partial budget and a cash-flow test. Count added revenue and avoided variable costs; subtract added operating costs and any lost production. Then compare the residual with annualized capital cost, using a stated useful life, cost of capital and residual value; for an actual investment, also examine discounted cash flows over that life, taxes where relevant and the risk of an early replacement or unusable resale market. Separately check actual debt-service timing and liquidity. Do not double-count depreciation and full loan principal as annual economic costs, or count an operator hour as both eliminated payroll and newly earned revenue without an actual operational change.

EXAMPLE—symbolic break-even template, not a forecast: let F be annualized incremental fixed cost; s the locally measured net benefit per treated hectare after recurring per-hectare costs. Break-even area is F/s only when s is positive. If additional annual subscriptions are not already included in F, add them before dividing. If the required area exceeds what the machine can treat within the critical window, nominal payback does not make the plan workable. No price, benefit or acreage is assumed here.

Protect access and people. Compare an owned machine with retrofit, custom-hire and shared-service options on the same work scope. For a contractor, agree on the agronomic deadline, queue priority, rescue responsibility, data rights and cancellation terms. Count training and skilled supervision rather than assuming workers disappear from the budget. A lower-capital arrangement is not accessible if it arrives after the crop’s useful window. These are procurement recommendations, not measured affordability outcomes.

Stress-test the downside. Recalculate with no yield improvement, lower utilization, a service outage, a missed application and the cancellation cost. Keep unmeasured benefits visible but separate. Fatigue relief may justify expenditure even where cash savings do not; say so honestly.

Set ecological and independence conditions. Track water withdrawals and relevant depletion effects, herbicide quantities and follow-up control, and the ability to retrieve data and repair equipment. Do not claim carbon, biodiversity or watershed benefits that the trial did not measure. S6 S9 S10 S11

For a farm with documented overlap, guidance is a candidate to evaluate; for one constrained by irrigation decisions, feedback may be useful. Those are starting hypotheses for a local budget, not purchasing instructions. Targeted weed control deserves local trials with timing and crop-safety safeguards. Variable-rate prescriptions and AI advice deserve stronger proof that the recommendation—not merely its delivery—is sound. That ordering is an editorial judgment from these selected studies, not a universal league table.

Technology should make a farmer more capable, not more captive. Progress means a healthier operation and a healthier place: fewer avoidable burdens on workers, neighbors, soils and water, with benefits that survive scrutiny. We welcome corrections and contrary field evidence through our existing response form. Integrity requires keeping the failures in the story—and revising the story when better evidence arrives.

References

• [S1] Matt Darr. Precision Ag Technology Savings. 2012. DOI: Not assigned (extension article). Source.
• [S2] Tulsi P. Kharel; Amanda J. Ashworth; Aaron Shew; Michael P. Popp; Phillip R. Owens. Tractor guidance improves production efficiency by reducing overlaps and gaps. 2020. DOI: 10.1002/ael2.20012. Source.
• [S3] David Schimmelpfennig. Farm Profits and Adoption of Precision Agriculture. 2016. DOI: Not assigned (USDA ERS ERR-217). Source.
• [S4] A. F. Colaço; B. M. Whelan; R. G. V. Bramley; J. Richetti; M. Fajardo; A. C. McCarthy; E. M. Perry; A. Bender; S. Leo; G. J. Fitzgerald; R. A. Lawes. Digital strategies for nitrogen management in grain production systems: lessons from multi-method assessment using on-farm experimentation. 2024. DOI: 10.1007/s11119-023-10102-z. Source.
• [S5] Susan A. O’Shaughnessy; Paul D. Colaizzi; Craig W. Bednarz. Sensor feedback system enables automated deficit irrigation scheduling for cotton. 2023. DOI: 10.3389/fpls.2023.1149424. Source.
• [S6] Alicia Allmendinger; Michael Spaeth; Marcus Saile; Gerassimos G. Peteinatos; Roland Gerhards. Agronomic and Technical Evaluation of Herbicide Spot Spraying in Maize Based on High-Resolution Aerial Weed Maps—An On-Farm Trial. 2024. DOI: 10.3390/plants13152164. Source.
• [S7] Roland Gerhards; Peter Risser; Michael Spaeth; Marcus Saile; Gerassimos Peteinatos. A comparison of seven innovative robotic weeding systems and reference herbicide strategies in sugar beet (Beta vulgaris subsp. vulgaris L.) and rapeseed (Brassica napus L.). 2024 (2023 online/copyright; 2024 volume 64:42–53). DOI: 10.1111/wre.12603. Source.
• [S8] Katherine Lim; Jonathan McFadden; Noah Miller; Katherine Lacy. America’s Farms and Ranches at a Glance: 2024 Edition. 2024. DOI: Not assigned (USDA ERS EIB-283). Source.
• [S9] U.S. Government Accountability Office. Precision Agriculture: Benefits and Challenges for Technology Adoption and Use. 2024. DOI: Not assigned (GAO-24-105962). Source.
• [S10] Frank A. Ward; Manuel Pulido-Velazquez. Water conservation in irrigation can increase water use. 2008. DOI: 10.1073/pnas.0805554105. Verified abstract.
• [S11] Federal Trade Commission. Nixing the Fix: An FTC Report to Congress on Repair Restrictions. 2021. DOI: Not assigned (government report). Source.

• [S12] Federal Bureau of Investigation. Ransomware Attacks on Agricultural Cooperatives Potentially Timed to Critical Seasons. 2022. DOI: Not assigned (PIN 20220420-001). Source.

• [S13] Vladyslav Pitsyk; Johanna Pfrombeck; Markus Gandorfer. Economics and adoption perspectives of site-specific weed management: A review. 2026. DOI: 10.1007/s11119-026-10381-2. Source.