Agriculture AI has moved from a research topic to a shipping product category over the last three years. Satellite imagery at sub-meter resolution is affordable. Vision models for crop health and yield are accurate enough for commercial use. Multilingual LLMs enable farmer-facing advisor apps in markets where earlier generations of software failed. This post is the stack we deploy for agri-tech clients, and the adoption realities behind the tech.
The four layers
Layer 1: Data — satellite, drone, IoT, weather
ESA Sentinel data is free, 10-meter resolution, 5-day revisit. Commercial high-resolution (Planet, Maxar) costs more but reaches sub-meter. Drone imagery covers the gaps at farm scale. IoT soil moisture and weather sensors are standard equipment at commercial farms. Aggregating these feeds is standard infrastructure work; most agri-tech companies have this layer solved.
Layer 2: Vision models — crop health, yield, disease
NDVI and related indices have been used for decades. Deep-learning models for disease detection, growth-stage classification, yield prediction have become production-grade. Accuracy depends heavily on crop type and region — a disease classifier trained on US corn won't work on Indian rice without substantial retraining.
Practical tip: partner with agronomic institutions for labeled data. Academic and government extension services often have decades of labeled imagery. Licensing agreements with these institutions is often faster than building your own labeling pipeline from scratch.
Layer 3: Recommendation engine
Given current crop state, weather forecast, soil conditions, and growth stage: when to irrigate, how much fertilizer, which pests to watch for. Recommendations combine vision model outputs with agronomy rules (region-specific, crop-specific). This layer is where domain expertise matters most; a well-intentioned generic recommendation engine produces agronomically wrong advice with real yield consequences.
Layer 4: Farmer-facing app
The critical layer. The best tech stack below means nothing if farmers don't use the app. Design rules that matter: vernacular languages (Hindi, Telugu, Marathi, Swahili, Spanish-by-region), voice-first interfaces (farmers are not typing), offline capability (rural connectivity is unreliable), extension-officer integration (local trusted intermediaries onboard and support farmers).
The adoption reality
Smallholder farmers — the majority of global agriculture by headcount — adopt tools that immediately reduce risk or increase yield in ways they can verify. They don't adopt tools whose value is in aggregated data for someone else, or whose benefits are in 'premium pricing via sustainability certification' that require faith in distant market dynamics.
Product patterns that work: pest alerts with high confidence and low false-positive rates; irrigation recommendations that save water costs; price-of-produce information with regional mandis; micro-insurance linked to parametric triggers (rainfall, temperature).
Product patterns that don't work (in smallholder contexts): full farm-management suites (too much to learn), subscription models (cash-flow mismatched to crop cycle), carbon-market tools (too abstract, payback too slow).
Commercial agriculture is different
Large commercial farms in North America, Brazil, Australia are a different market. They have capital, technical staff, and integration with equipment (John Deere, Trimble, Bayer). Products here are SaaS with seat-based or acre-based pricing; decision support around planting, input optimization, and harvest timing. Adoption patterns look more like enterprise B2B.
The financing angle matters
In emerging markets, the agri-tech products that scale fastest are often linked to financing. Input loans, crop insurance, and working-capital facilities provide the cash flow that makes advice actionable. A pest alert is useful; a pest alert plus a loan for the pesticide is adopted. Agri-tech companies that ignore the financing layer often plateau at pilot scale.