Agritech pitch decks promise AI that tells every farmer exactly what to plant, when to water, and how much they'll harvest. Reality is more modest, and more interesting. Working at Dimitra, an AI-powered agritech and compliance platform, I've seen where the value actually shows up.
What works in production
Satellite crop monitoring
Vegetation indices like NDVI from Sentinel-2 show crop health over time across whole regions. Uses:
- Spotting stressed fields early
- Monitoring crop progress across a season
- Checking land use for compliance
It works because the data is free, global, and regular. See Remote Sensing with Machine Learning.
Plot boundary detection
Segmentation models can detect field boundaries from imagery, speeding up farm mapping. Smallholder plots with irregular boundaries are harder, so pair this with human correction in a field app.
Compliance and deforestation checks
Overlaying plots with forest data for regulations like EUDR. This is AI-assisted rather than fully automated. See Detecting Deforestation.
Document and data extraction
LLMs extracting structured data from certificates, invoices, and farm records. Unglamorous, but it saves a lot of manual work.
Promising, still maturing
- Yield estimation: reasonable at regional scale. Field-level estimates for smallholders lack ground truth.
- Pest and disease detection from phone photos: works in demos, but struggles with real-world lighting, varieties, and look-alike symptoms.
- Advisory chatbots: promising for reach, risky if advice is wrong or not localised.
The real constraint: data
| Problem | Why it matters |
|---|---|
| Ground truth | Models need verified labels. Field visits are expensive |
| Fragmented records | Cooperatives, exporters, and governments keep separate data |
| Connectivity | Field apps must work offline (offline-first guide) |
| Trust | Farmers need to know how their data is used and what they get back |
| Local context | Models trained in one region often fail in another |
Advice for engineers entering agritech
- Start with the workflow, not the model. Who will act on the output?
- Design for low connectivity and cheap phones.
- Keep humans in the loop for decisions that affect income.
- Measure accuracy by region, not just overall.
- Give farmers value back, not just data extraction.