Build the evidence before the model.
Accuracy comes from field protocol and labelled outcomes first. AI earns more responsibility only after it performs on independent farm data.
Months 0-3 - data foundation
Define common zone, crop, device, observation, action and outcome fields. Collect baseline and require source/quality metadata.
Months 3-6 - explainable product
Use rules, thresholds, data-quality checks and human-approved workflows. These are auditable and do not need a large training dataset.
Months 6-12 - labelled pilot learning
Growers and advisers record what a signal meant after inspection. This creates the labelled examples needed to evaluate anomaly detection or image support.
Months 12-18 - controlled model evaluation
Test candidate models on held-out farms/seasons, not the same data they learned from. Report false alerts, missed events, precision and recall by crop and region.
Release only with guardrails
Retain confidence state, evidence source, human override, model version, incident review and a clear rollback path.
Important: predictive AI is a future pilot-stage capability, not a current claim.