The Second Signature: Who Owns an AI Fungicide Call That Misses
Photo: OSU Special Collections & Archives : Commons, No restrictions, via Wikimedia Commons
A fungicide decision used to have a short chain of custody. A grower, a crop adviser, a label, and a sprayer. In 2026, the chain has more links: a retailer portal that scores disease risk, an independent app that blends weather and satellite imagery, an equipment dashboard that nudges a pass, and a language model that summarizes the whole thing into a confident paragraph. The recommendation may be useful. It may also be the least accountable document on the farm.
The Recommendation Chain Has More Links Than the Label
The first mistake is treating an AI recommendation as a single object. It is usually an assembly. A data vendor supplies weather. A satellite provider supplies imagery. A model developer supplies a disease-pressure score. A retailer wraps that score in a portal. A local agronomist reviews it, sometimes in bulk, sometimes after the fact. The grower receives a notification that reads like advice but behaves like a prompt.
That matters because pesticide labels are legal documents. A model output is not. When a fungicide pass misses, the label still governs the product, the rate, the interval, and the restricted-entry language. The model governs nothing unless a contract says it does. Most contracts say the opposite. Terms of service commonly disclaim agronomic advice, warranty fitness for a particular field, and liability for crop loss. The grower may have paid for a recommendation that the vendor describes, in fine print, as an informational signal.
Three Failure Modes That Matter in a Wet June
AI disease models fail in ordinary ways. The first is provenance. A model trained on trial data from the Pacific Northwest may not understand a humid Midwestern June, a Southern rust year, or a field with a history of strobilurin resistance. Training data is not a neutral fact. It has geography, weather, crop genetics, and management bias baked in.
The second is input lag. A weather station eight miles away is not your field. A satellite image from four days ago is not today's canopy. A model can be mathematically elegant and still be late. In a disease window, late is expensive.
The third is label confounding. Yield maps, spray records, and scouting notes are often used to train or validate models, but those records reflect what the grower did, not what would have happened otherwise. A field that was sprayed early and stayed clean does not prove the model was right. It may prove the grower was cautious.
The Contract Language You Will Not Hear in the Demo
Ask three questions before relying on an AI fungicide recommendation. First, what data trained the model, and how much of that data comes from your region, your crop, and your disease pressure? Second, who reviewed the recommendation before it reached you, and what is their license or certification? Third, what does the vendor's contract say about liability when the recommendation is followed and the field still breaks?
The answers are often revealing. Some portals require a certified crop adviser to sign off. Others route recommendations through a retailer's agronomy desk with variable review depth. Still others present a score with no human signature at all. A score is not a prescription. A prescription should have a name on it.
A Practical Pressure Test Before the Sprayer Moves
Growers do not need to reject AI advice. They need to pressure-test it. Start with local validation. Has the model been checked against scouting records from your county or a comparable geography? Ask for a backtest, not a demo. A backtest shows how the model performed in past seasons with data that was actually available at the time.
Next, demand provenance. Which weather source, which satellite product, which soil layer, and which trial network feed the recommendation? If the vendor cannot name the inputs, the output is a guess with a user interface. Then ask about confidence. A probability score is more useful than a binary spray-or-wait command, but only if the grower knows how the probability was calibrated.
Finally, keep an audit trail. Save the recommendation, the date, the model version if available, the human reviewer, and the field conditions at the time of application. If a dispute arises, that trail is worth more than a screenshot of a green checkmark.
The Human Sign-Off Is Not a Formality
The most practical rule for 2026 is simple: the model can triage, but a qualified human should sign. That does not mean every pass needs a fresh field visit. It means the responsibility should be assigned before the sprayer rolls, not reconstructed after a loss. A licensed agronomist or certified crop adviser who understands the field, the label, and the local disease history is still the accountable link.
AI can make agronomy faster. It cannot make accountability disappear. The second signature is the one that matters.