When a woman farmer near Bannerghatta calls for help with her tomato plants, she does not say "My tomato has a problem."
A male farmer might though. There are enough datasets to show how men are terse while engaging with the AI advisory systems. She says, “My tomato has black spots, I have been spraying, I have been tilling, I have been removing the weeds, I have been doing everything I know to do, so why is this still happening?” She has already run her own field experiments before the advisor walks up. She is asking for the next intervention, not the first.
When women in Indian smallholder collectives engage with agricultural advisory systems, they arrive as field researchers with diagnostic intuition built from years of soil contact and a working notebook of attempted interventions.
Every agricultural AI product currently being marketed at them is built for a user who describes a symptom and waits for a prescription. That user does not exist on these farms.
In vast majority of Indian farms, the woman runs the field while the man runs the purchase order. What happens when the entire agricultural AI sector has built its product line for the buyer and not the executer?
This and several more questions surfaced during my second edition of “AI in Ag Podcast”.
This behavioural insight was surfaced in beautiful detail by Preethy Iyer who joined us to share the AI engine she has been building for women farmer collectives at Kai Thota. She was joined by Rhishi Pethe, a dear friend and senior advisor at the Gates Foundation, and Sheriff Babu, who runs an agent-swarm system on WhatsApp for Indian farmers.
India's agricultural AI sector is now large enough to attract serious capital and serious policy attention. Sovereign dialect models, smallholder advisory tools, autonomous implements, voice agents, WhatsApp bots, are all being built simultaneously, often with foundation or government money or private venture capital, depending on its public or private orientation.
The metric that gates whether these investments continue is the wicked question of impact. If the metric is broken, the entire allocation is broken.
The dominant story in agricultural AI today is that the limit is technology. Better models, better data, faster inference, cheaper compute, more dialects, and the smallholder advisory problem is solved.
That story falls apart at the first field visit.
Preethy described arriving in Kai Thota with a textbook ag-tech stack: sensors, soil-data collection, a recommendation engine for what to grow. The women collectives she works with looked at it and said “we don’t need it. We touch the soil, we know.”
They had years of context on which the textbook stack had nothing to add. What they did not know was how the city worked, what the urban subscriber wanted, what the market would absorb next month.
The bottleneck was knowledge of the city, not the land.
This inversion shows up across the entire conversation. Sheriff, building from the other end, found that grape farmers reporting yellow patches were universally being told by general models that they had a disease. None of the models identified the much commoner cause, zinc deficiency. Rhishi described a “barbell distribution” of adopters in the sector: a cluster of early enthusiasts and a much larger cohort still asking where to begin, what the ROI is, whether this is just the last decade of agritech disappointment in a new wrapper.
We are in 2026 and the median Indian agritech buyer has not budged.
Now consider how impact gets measured.
A funder wants to know whether its capital is producing better outcomes for smallholders. The cleanest signal is product purchase data. Did the farmer buy the recommended input? Did the seller move volume? Did the advisory tool drive a transaction?
Then ask who, inside an Indian smallholder household, actually executes a purchase.
Preethy described this pattern with absolute specificity. Men make the visible strategic decisions: what to plant, what to spray, what to buy. Men also run the EMI payments on the phone and the bike. Women weed, sow, harvest, run the home, take backpack sprayers into the field, and crucially, take the loans that fund the men’s EMIs.
Men buy whatever the sales pitch recommends, frequently while drunk, frequently because a subsidy exists, frequently with no view on whether the product helps.
This means the purchase-data layer is measuring decisions made by people who are not in the field. The yield-data layer, which arrives two or three years later, is too lagging to course-correct any specific advisory product. The intermediate layer that would actually measure whether the right person in the household received useful advice does not exist in most reporting frameworks.
When the metric counts the husband, the industry builds for the husband.
A Coimbatore startup recently launched a battery-powered weeder for Indian smallholder farms where the operator works it with a joystick, from a distance, never bending down
Watch a woman weed her plot and you can see whom the product was actually drawn up for. She squats, pulls the deep-rooted grass by hand, and composts it, because the grass becomes the nutrient that feeds the next crop. She has already optimized the system. The weeder is solving a problem she does not have.
A different startup brought a lightweight battery-powered weeder to the same collective and asked the women to evaluate it. They tested it, found that walking with it shifted the battery weight in a way that made the weeds fly rather than separate, and handed it back. Their feedback was technically precise and product-killing. They were never the customer the company had imagined.
The misallocation does not stop at the product layer. It runs all the way down to the language layer that every Indian agritech company is currently rebuilding privately.
India has, through Bhashini and several other initiatives at IIT Madras, built some of the most ambitious linguistic public infrastructure of any government in the world. Thousands of hours of emotionally tagged training data across Indian languages, transcribed and verified. Meanwhile, every agricultural AI company is rebuilding its own private dialect corpus and calling it a moat.
The vocabulary of Indian farming changes every 150 kilometres.
A wild berry called sundakkai in a Tamil village becomes chikka badne kai in a Kannada village two hours away. A berry has been reclassified as a brinjal. No general-purpose LLM resolves this, because the data is local, oral, and not on the internet.
The country has the public infrastructure to solve this for everyone as a commons. The industry is solving it a thousand times in parallel.
Preethy wants a thousand Kai Thotas, each focused on its local context, all running on shared public infrastructure. Sheriff estimates the one-time public investment to build a serviceable Indian agricultural foundation model at ₹150 to ₹200 crore.
There is one technical question whose answer reshapes the entire investment case, and the panel disagreed on it openly.
Will general-purpose frontier models eventually absorb enough agricultural context that smallholder advisory becomes a feature inside a horizontal product like Gemini or Claude?
Rhishi assigns this some non-trivial probability over the next five years. His position is that even granting the possibility, the right move is to build for today’s farmer with today’s tools, because waiting is a moral failure when a million advisory conversations can be improved right now.
Sheriff assigns the probability close to zero. He has attempted distilling the agricultural portion of a frontier model and found nothing to distill. His argument is architectural. The transformer’s pretraining diet does not contain the kind of contextual diagnostic reasoning that distinguishes a zinc-deficient grape leaf from a fungal one, and the architecture cannot manufacture that reasoning out of general text on the internet.
If Rhishi is right, every rupee spent on sovereign Indian agricultural models is a depreciating asset waiting to be obsoleted by the next frontier release. If Sheriff is right, every rupee spent waiting for ChatGPT to figure out Indian smallholders is wasted runway, and Indian capital should be deploying into Indian models today.
We do not yet have enough evidence either way. The entire capital stack of Indian agricultural AI for the next decade rides on which of the two views turns out to be correct.
What follows from all of this is a short list of moves that should be obvious by now.
Foundations and Impact funders should retire product-purchase data as a primary impact metric for smallholder agricultural AI. The metric is reading the wrong household member. Replace it with measured reach to women specifically, measured trust, feasibility of acting on the advice given, and observed capability to act on it. These metrics are harder to collect, produce smaller dashboards, and correspond to the actual farm.
Treat the Indian agricultural dialect corpus as Digital Public Infrastructure.
Fund the Bhashini-equivalent for agriculture as a one-time public good. End the parallel private rebuild and free a hundred companies to compete on what they should be competing on, which is the quality of advice and the trust of the woman in the field.
Stop designing implements for the operator who is not present. The next weeder, sprayer, or harvester that enters an Indian smallholder farm should be co-designed with women operators in the room, walking with the prototype, holding the load, and rejecting it freely when the battery placement is wrong.
What three very different practitioners converged on was the same finding, rendered three ways: the models are not the bottleneck; the way we count impact, the household member we count it through, and the feedback we let into the next iteration are.
If you don’t want CC hassles, you are most welcome to use paypal or UPI (venkat.raman.kr@icici) and pay the annual subscription (8500 INR/95 USD) with your email in the comment. I will enable access immediately.
P.S. Supporting this work doesn’t have to come out of your pocket. If you read this as part of your professional development, you can use this email template to request reimbursement for your subscription.
So, what do you think?
How happy are you with today’s edition? I would love to get your candid feedback. Your feedback will be anonymous. Two questions. 1 Minute. Thanks.🙏
💗 If you like “Krishi.System”, please click on Like at the bottom and share it with your friend
