John Deere (Deere & Company)

Mixed evidenceThreshold is the controlDecision rights moved

In-season post-emergence herbicide application (See & Spray targeted spraying)

That a sprayer cannot tell a crop plant from a weed while moving through a field, so the only reliable way to control weeds is to broadcast herbicide across the entire field regardless of where weeds actually are.

Agriculture · United States (corn, soybean and cotton fields); independent field trial in Arkansas soybeans

Collections: Human still decides · Embodied work · Negative results

Uniform herbicide broadcast across a field contrasts with camera-guided nozzles targeting individual weeds under a grower-set threshold.

Executive brief

The operating-model shift, in one view.

The transferable lesson is not the 59% saving, it is what determines whether you get it. 49 per acre at high sensitivity, and a 280% annual increase in pigweed at low sensitivity - and the researcher attributes that difference to the grower's threshold setting, not to the model misclassifying weeds. When you move a decision to a model, the human control usually collapses into a small number of configuration choices made before the work starts, and those choices carry consequences the operator will not see for a season or more. Most organisations govern the model and leave the threshold to whoever runs the machine.

AI value · Herbicide volume applied per acre versus whole-field broadcast application

Company-reported

59% average herbicide savings

Company-reported and company-calculated.

Before

Decides the post-emergence herbicide programme for the field based on expected weed pressure across the whole area. → Drives across the field applying herbicide uniformly because the sprayer cannot distinguish crops from weeds. → Receives herbicide across the treated area, including ground with no weeds. → Pays for…

After

Sets the spray threshold that determines how aggressively detected plants are treated as weeds. → Scans over 2,100 square feet per second at up to 15 mph while the sprayer moves. → Deliver a precise dose of herbicide only where a weed is recognised, leaving unweeded ground untreated. → Monitors weed escapes and…

Human boundary

Why it matters

This case is also unusually well evidenced for a commercial AI claim: a land-grant university measured it independently, published counter-evidence, and estimated that improper use could cost Arkansas producers more than the midpoint of the potential savings.

How the work changed

Before

How the work ran before the change.

Step 1 of 4
Grower / agronomist
Decides the post-emergence herbicide programme for the field based on expected weed pressure across the whole area.
ControlLabel rates, agronomic recommendation, whole-field application plan.
Step 2 of 4
Sprayer operator
Drives across the field applying herbicide uniformly because the sprayer cannot distinguish crops from weeds.
ControlBoom height, speed and nozzle calibration; no plant-level targeting is possible.
Step 3 of 4
Field
Receives herbicide across the treated area, including ground with no weeds.
ControlChemical cost and environmental load scale with acres, not with weed presence.
Step 4 of 4
Grower
Pays for herbicide across every acre sprayed and absorbs crop injury from herbicide applied to unweeded ground.
ControlInput budget.

What changed

A moving sprayer can distinguish crops from weeds plant by plant.

Decision rightThe grower sets the threshold; AI decides plant by plant

After

How the same work runs now.

Step 1 of 4
Grower / agronomist
Sets the spray threshold that determines how aggressively detected plants are treated as weeds.
Step 2 of 4
See & Spray camera and vision system
Scans over 2,100 square feet per second at up to 15 mph while the sprayer moves.
Control120 ft boom camera array; ExactApply nozzle hardware.
Step 3 of 4
ExactApply nozzles
Deliver a precise dose of herbicide only where a weed is recognised, leaving unweeded ground untreated.
ControlIndividual nozzle actuation.
Step 4 of 4
Grower / agronomist
Monitors weed escapes and seed-bank consequences across seasons and adjusts sensitivity, residual herbicide strategy.

Process model built from the published workflow evidence for John Deere (Deere & Company). Every step, actor, and control appears in full below.

Every step, actor, and control

Exception path

Agronomic rather than transactional. Independent University of Arkansas research establishes the required exception handling: residual (pre-emergence) herbicides should be broadcast rather than targeted, and only post-emergence chemistry should be targeted. On the single-tank Premium configuration this forces the grower either to apply residuals through the targeted system or to make two separate passes. Weed escapes surviving a targeted pass cannot be controlled by a follow-up application and must be managed as a seed-bank problem in subsequent seasons. Deere's commercial exception path is financial rather than operational: customers pay only for acres where the technology is used, and from 2025 an Application Savings Guarantee means customers pay for the technology only when an application saving is achieved.

Work removed

Herbicide applied to ground carrying no weeds - the majority of most treated fields
Whole-field uniform application as the default execution model
Purchase and handling of herbicide volume for untreated-equivalent acres (an estimated 8 million gallons of mix across more than 1 million acres in 2024)

Decision authority

The per-plant application decision transfers to the onboard model. The grower retains authority over the sensitivity threshold that bounds the model's behaviour, the herbicide programme, and the pass structure. This is a genuine transfer of an operational decision that was previously made once per field by a human and is now made thousands of times per second by a model, with no possibility of per-decision human review. The boundary is set once per pass rather than negotiated per case, which makes the threshold a control of unusual leverage - independent research quantifies the consequence of setting it wrongly.

Before

01
Grower / agronomist
Decides the post-emergence herbicide programme for the field based on expected weed pressure across the whole area.
Control: Label rates, agronomic recommendation, whole-field application plan.
02
Sprayer operator
Drives across the field applying herbicide uniformly because the sprayer cannot distinguish crops from weeds.
Control: Boom height, speed and nozzle calibration; no plant-level targeting is possible.
03
Field
Receives herbicide across the treated area, including ground with no weeds.
Control: Chemical cost and environmental load scale with acres, not with weed presence.
04
Grower
Pays for herbicide across every acre sprayed and absorbs crop injury from herbicide applied to unweeded ground.
Control: Input budget.

After

01
Grower / agronomist
Sets the spray threshold that determines how aggressively detected plants are treated as weeds.
Control: Sensitivity setting; independent research finds the high sensitivity setting is best in most scenarios and that a low setting increases the weed seed bank.
02
See & Spray camera and vision system
Scans over 2,100 square feet per second at up to 15 mph while the sprayer moves.
Control: 120 ft boom camera array; ExactApply nozzle hardware.
03
Onboard machine-learning processors
Determine plant by plant whether each detected plant is a crop or a weed, and issue commands to individual nozzles - the application decision moves from the operator's whole-field plan to a per-plant machine judgement made in real time.
Control: The grower's sensitivity threshold is the operator-set boundary within which the model decides; no per-plant human review is possible or attempted.
04
ExactApply nozzles
Deliver a precise dose of herbicide only where a weed is recognised, leaving unweeded ground untreated.
Control: Individual nozzle actuation.
05
Grower / agronomist
Monitors weed escapes and seed-bank consequences across seasons and adjusts sensitivity, residual herbicide strategy and pass structure accordingly - a new agronomic control loop that did not exist under broadcast application.
Control: Independent guidance to broadcast residual herbicides and target only post-emergence chemistry; multiple passes required on the Premium configuration.

Work that left the path

Herbicide applied to ground carrying no weeds - the majority of most treated fields
Whole-field uniform application as the default execution model
Purchase and handling of herbicide volume for untreated-equivalent acres (an estimated 8 million gallons of mix across more than 1 million acres in 2024)
Crop injury from herbicide applied to unweeded ground (associated with an observed yield increase of about 4 bushels of soybean per acre in the Arkansas trial)
Payment for technology on acres where it is not used, and from 2025 payment where no application saving is achieved

Human role before

The grower and operator decided herbicide application at the field level and executed it uniformly. The operator's skill lay in calibration, timing and coverage; no plant-level decision existed because none was possible.

Human role after

The human no longer decides where herbicide goes. The grower's role moves upstream to setting the machine's decision threshold and downstream to managing the multi-season agronomic consequences of that threshold - weed escapes, seed-bank accumulation and resistance risk. Independent research establishes that this configuration decision, not the vision model, determines whether the technology creates or destroys value.

AI roleReal-time computer-vision and machine-learning classifier making per-plant crop-versus-weed determinations at scale and actuating individual nozzles accordingly. It holds the application decision for each plant within the sensitivity boundary the grower sets; it does not choose the herbicide, the timing or the threshold.

Outcomes

Herbicide volume applied per acre versus whole-field broadcast application

Company-reported

Broadcast spraying - herbicide applied uniformly across the entire field, calculated by Deere at an average of 15 gallons per acre of non-residual herbicide product applied→Average herbicide savings of 59%, and an estimated 8 million gallons of herbicide mix saved

The 2024 growing season · More than 1 million acres applied across corn, soybean and cotton fields in the United States, using See & Spray Premium and Ultimate configurations

Company-reported and company-calculated. Deere's footnotes disclose the basis honestly: the 59% is 'Compared to broadcast spraying and based on use of See & Spray Premium and Ultimate', and the 8 million gallons is 'Calculated by using an average of 15 gallons per acre for nonresidual herbicide product applied' - so the gallons figure is a derivation from an assumed application rate, not a measured volume, and Deere labels it 'estimated'. The saving applies to non-residual (post-emergence) product only. Independently corroborated in direction and magnitude by a three-year University of Arkansas field trial finding a 43-59% reduction versus broadcast, though that trial covers Arkansas soybeans only and used a prototype machine.

Herbicide use reduction versus broadcast application, independently measured

Verified

Broadcast application across the entire plot area→43% to 59% reduction in herbicide use

A three-year field trial begun in 2022, reported 2025-03-10 · Arkansas soybean plots at the University of Arkansas System Division of Agriculture Northeast Research and Extension Center, Keiser; small-plot research conducted at 8 to 12 mph

This is an independent land-grant university measurement, not a restatement of Deere's figure, which is what makes it valuable - but its scope is narrower than the commercial claim. It covers soybeans in one US state, and the machine used was a Blue River Agricultural Test Machine, a scaled version of See & Spray Ultimate, operated as a single-tank system to simulate the more widely adopted Premium configuration. It is therefore corroboration of the mechanism and magnitude rather than verification of the commercial product's performance across corn, soybean and cotton.

Per-acre economic return at the high sensitivity setting

Verified

Broadcast application method, assuming a $25,000 See & Spray upgrade cost from a comparable John Deere sprayer and incorporating applicator efficiency→Saved $30.49 per acre over the entire season with no increased risk relative to the broadcast method; the sprayer pays for itself after treating 819 acres for a $25,000 upgrade or 2,460 acres for a $75,000 upgrade if used at both early and mid post-emergence in a single season

One full season within the three-year trial, reported 2025-03-10 · Arkansas soybean production economics, modelled from the field trial using American Society of Agricultural and Biological Engineers standards for machine costs and speed

An independent economic analysis by a PhD researcher, based on stated assumptions about upgrade cost and applicator efficiency, and specific to Arkansas soybeans. It is a modelled return under one configuration and one crop, not a realised financial result, and the authors were still producing additional scenarios for other crops and cost structures at time of publication. The 'no increased risk' finding applies specifically to the high sensitivity setting.

Palmer amaranth pigweed population change at the low sensitivity setting

Verified

Weed population under broadcast application→Increased by 280% each year, because only larger weeds were sprayed and the escapes produced seed, increasing the soil seed bank

Measured year over year across the three-year field trial begun in 2022, reported 2025-03-10 · Arkansas soybean plots at the Northeast Research and Extension Center, Keiser

This is counter-evidence and is recorded as an outcome deliberately, not buried. It is independently measured and material: the same technology that delivers a 43-59% herbicide reduction can multiply the weed seed bank nearly fourfold per year if the grower sets sensitivity low. Critically, the researcher attributes the mechanism to the operator's configuration choice, not to a failure of the vision model - 'When a grower lowers the sensitivity setting, only larger weeds are sprayed' - which makes this a governance and configuration lesson rather than a system defect. Arkansas soybeans only. Researchers estimate improper use could cost Arkansas producers upwards of $60 million per year through accelerated herbicide resistance or additional in-season applications.

What leaders can reuse

Anti-pattern

Optimising the machine's threshold for the immediate, visible metric. Setting See & Spray to low sensitivity produces exactly what a per-pass cost report rewards - the least herbicide used - while multiplying the Palmer amaranth seed bank by 280% each year, escapes that a follow-up application cannot control, and accelerated herbicide resistance. The researchers estimate improper use could cost Arkansas producers upwards of $60 million per year against modelled savings of $1.6 million to $48 million, so the downside can exceed the upside. The second anti-pattern is treating the vendor's headline as the whole finding: Deere's 59% figure is honestly footnoted to broadcast spraying, non-residual product and specific machine configurations, and independent research separately establishes that residual herbicides should be broadcast rather than targeted, which requires multiple passes on the widely adopted Premium configuration. Quoting 59% without the configuration and pass-structure conditions overstates what a buyer will experience.

Questions

01Where have we moved a high-frequency decision to a model, and who owns the threshold that bounds it?
02Does the person setting that threshold see the consequences of setting it wrongly, and on what time horizon?
03Which of our AI metrics would improve if someone configured the system for the short-term number at the expense of a second-order outcome?
04Do we have any independent measurement of an AI outcome we are citing, or only the vendor's figure and our own?
05What is our equivalent of the seed bank - the slow-accumulating consequence our primary metric will not show us for a year?
06Would our AI vendors price on realised outcome rather than on capability, and have we asked?

Portability conditions

The decision being automated is high-frequency, low-individual-consequence, and aggregates into a material outcome
Real-time sensing can distinguish the target from the non-target reliably enough at operating speed
A single configuration threshold can be defined that meaningfully bounds the model's behaviour
Someone owns and is accountable for that threshold, and understands its downstream consequences over a longer horizon than the immediate result
Independent measurement of the outcome is available, or can be commissioned, rather than relying solely on the vendor's figure
Second-order effects (here, seed-bank accumulation and resistance) are monitored on their own cycle, not inferred from the primary metric
The vendor is willing to price on realised outcome rather than on capability

Reputation risk

low - both sources are public and directly retrievable, the primary claim is honestly footnoted by Deere with its baseline and configuration disclosed, and the independent corroboration comes from a land-grant university that also published counter-evidence. The material obligation is to carry the 280% pigweed finding and the residual-herbicide constraint alongside the savings figure; presenting the 59% saving alone would be a selective reading that the independent research directly contradicts in part. The Arkansas trial's narrower scope - soybeans, one state, a prototype machine simulating the Premium configuration - must be stated rather than implied to cover the commercial claim.

Evidence and authority

What the public record supports.

Current · updated 2026-08-22

1 independent, 1 primary; publication outcomes are verified and reported.

Bundle 1.0.0 · reviewed 2026-08-23 · stable ID caa3124cf4457913

Sources

See & Spray Customers See 59% Average Herbicide Savings in 2024John Deere (Deere & Company) · 2024-09-18 · primary
Precision agriculture research measures effectiveness of See & Spray technologyUniversity of Arkansas System Division of Agriculture, Arkansas Agricultural Experiment Station · 2025-03-10 · independent

Read the evidence, freshness, caveat, and version policy.

In-season post-emergence herbicide application (See & Spray targeted spraying)

The operating-model shift, in one view.

How the work changed

Outcomes

Herbicide volume applied per acre versus whole-field broadcast application

Herbicide use reduction versus broadcast application, independently measured

Per-acre economic return at the high sensitivity setting

Palmer amaranth pigweed population change at the low sensitivity setting

What leaders can reuse

What the public record supports.

Related transformations

AI-counted pink-bollworm traps and threshold-based spray advisories

WPS Analytics full-population inspection

Real-time proprietary sepsis alerts

AI straight-through processing for low-value cheques

Sources