Chapter 04 · Section I · 17 min read
Agriculture: yield, pest, and price
Two thirds of working Nepalis still farm. The AI applications that could meaningfully help them are not the ones with the loudest pitch decks. A practical tour of what works on a smallholder plot and what does not.
Roughly two-thirds of working Nepalis still farm — most of them on small, scattered plots of less than a hectare, often planting two or three crops on the same land across the year. This is the largest single economic activity in the country, and arguably the one where AI has the most distance left to cover. It is also the area where pitch decks most often outrun reality.
Three problems that fit AI well
Within the long list of agricultural problems an AI system could address in Nepal, three stand out as good fits — the data exists, the user need is real, and the technology has matured enough.
Pest and disease identification from a phone photo. A farmer notices yellow spots on a maize leaf. Today, the best they can do is ask the neighbour, or wait for the next krishi sewa kendra (agricultural extension) visit. A model trained on labelled images of maize, paddy, potato, tomato, and a few other staples can identify common pests and diseases at accuracy that, for the top dozen problems on each crop, exceeds the average extension worker. PlantVillage and similar open datasets give a starting point; the gap is curating Nepali-context images, especially for the diseases that look different on local cultivars.
Yield estimation from satellite imagery. Sentinel-2 satellite imagery, free and updated every five days, lets a model estimate paddy or wheat yield at the ward level with respectable accuracy. ICIMOD and a few Nepali researchers have published on this; the gap is operationalising it into a tool that ward agriculture officers actually use. Once that gap is closed, a province can know — in October — roughly what its winter wheat harvest will look like in March.
Price forecasting at the farmgate. Tomato prices in Kalimati doubled and halved twice over the past quarter. A farmer planting in Dhading has no clear way to know whether to plant for an October harvest or a January one. A model that combines historical price series from the Kalimati market with weather forecasts and acreage estimates can give probabilistic forecasts that, even if rough, are better than the guess most farmers are working from. This is straightforward time-series ML; the rate-limiter is reliable price data feed.
What does not work as well as the pitch suggests
Equally important: three pitches you will hear that, on inspection, do not yet hold up.
“AI-driven precision irrigation” sounds excellent and is at the heart of many proposals. The Israeli-style precision agriculture stack — soil sensors, drip timers, computer-controlled valves — requires a per-hectare capital investment that almost no Nepali smallholder can recover. The technology works, but the economics don’t. Where this stack makes sense is on the few hundred large commercial farms in the country, not the millions of smallholders.
“AI-powered farm management apps” are easy to ship and hard to keep used. The history of agritech apps in Nepal is full of products that achieved 50,000 downloads, 5,000 active users in month one, and 200 by month three. The product almost always loses to “ask the neighbour” because the neighbour is faster, free, and embedded in a trust network the app cannot easily replace.
“AI-trained drone surveys” are mostly marketing. Drones are useful for very specific tasks — large-area mapping, disaster assessment, plantation surveys on commercial estates — but the assumption that a fleet of drones can cost-effectively monitor smallholder plots is wishful. The unit economics of a drone passes do not work below a certain plot size.
What a good agricultural AI product looks like
Three traits, drawn from the small set of products that have actually held users past the first season:
-
Works on a low-end Android, offline, with intermittent connectivity. The product runs on whatever device the user already owns and downloads model weights once. It does not require a data connection at the moment of use.
-
Solves one specific problem decisively. A pest-ID app that identifies twenty pests with high accuracy is better than a “farming assistant” app that tries to do everything and does each thing badly.
-
Integrates with a trusted human in the loop. The most successful Nepali agricultural apps are those used by extension workers or village-level animators, who in turn talk to farmers. The app extends the expert; it does not replace them. Farmers trust people they have met more than apps from companies they have not.
Where the data should come from
Nepal does not need to build agricultural AI datasets from scratch. ICIMOD, DOA (Department of Agriculture), NARC (Nepal Agricultural Research Council), and several donor-funded studies have collected useful data over the past decade. What is missing is the work of cleaning, schematising, and publishing these into a single national platform that a model team can actually use.
A practical agenda: a small NARC–university partnership that maintains a public Nepal Agri Data Hub, with labelled image sets for the top dozen crops, time-series of price and acreage by district, satellite-derived vegetation indices, and weather. None of this is technically hard. It needs a leader, a budget, and a few years of focused work.
Check your understanding
Quick check
—Which of the following is the best fit for AI in Nepali smallholder agriculture, given current technology and farmer economics?
Quick check
—Why does the pattern of “AI plus a trusted human” — extension worker, village animator, or cooperative officer — tend to succeed in Nepal where pure direct-to-farmer AI apps often do not?
What comes next
Agriculture is one face of the land. Disaster is the other — floods after the monsoon, landslides on the mountain roads, the next earthquake whose timing we cannot predict but whose response we can prepare. The next section is about AI for early warning, risk mapping, and post-disaster response.