Before the compute arrives, the extension agent

The most consequential classroom for AI in Latin American agriculture this year is not on a university campus. It is the passenger seat of an extension agent's pickup on a dirt road in the interior of São Paulo state, or the terrace of a packing shed near Chillán where an agronomist opens a laptop next to crates of pre-export fruit.

Start with the paper. On 20 August 2026, Embrapa's Semear Digital centre published a plain-language readout of a peer-reviewed methodology it developed with researchers at the University of Campinas to predict whether a specific piece of digital agricultural technology will actually be adopted in a specific farm's context. The paper, "Diagnostic framework for assessing Smart Agricultural Technology adoption", sits in Smart Agricultural Technology volume 14. It combined a bibliometric review of 814 publications with a shortlist of ten existing indices, including the World Bank's Agriculture Digitalization Index and Embrapa's own Ambitec-Agro. Twenty-two agronomists, engineers, statisticians and computer scientists tuned the weights. Two contrasting technologies were put through it as a proof of concept: an Embrapa aquaculture management app called Aquicultura Certa and a small hardware device that clips to a harvest bag and counts fruit as it lands inside.

Now the finding you should sit with. When Aquicultura Certa was scored by its developers, it came out at 3 on a 1-to-5 scale of adoption feasibility. When it was scored by farmers who were actually running the pilot, the same technology came out at 4. Thais Dibbern, the Unicamp researcher who led the study, points to a simpler explanation than a measurement error. The developers imagine every possible farm, including the ones with unreliable power and thin connectivity. The farmers doing the trial live in places where those barriers have already been overcome. The developer sees the ceiling of the technology's reach. The adopter sees only the room they are standing in. On the harvest-bag counter, both scored 4, because the device runs on Bluetooth and asks nothing of the farm it lands on. The methodology accidentally names the biggest single lesson of five years of digital agriculture in Brazil: what a farmer will pay attention to has almost nothing to do with what a demo shows.

The classroom you cannot see from a keynote

Semear Digital itself is the more instructive object. Launched in April 2023 by Embrapa Digital Agriculture and the São Paulo Research Foundation (Fapesp), it runs ten Agricultural Technology Districts across the five Brazilian regions. Each DAT is a small piece of infrastructure grafted onto a municipality: an antenna to widen the connectivity footprint, a set of validated tools chosen for the local crop mix, a training programme aimed at rural producers, technicians, extension workers, and consultants. Reading the DAT reports is a corrective to almost every corporate AI announcement I have seen this year. The classroom is not virtual. The teacher is often the extension worker. The syllabus is the crop calendar.

SENAR, the older public rural-training body that pre-dates the digital push, now reaches roughly 100,000 producers through its Conecta Produtor app and runs more than 120 free distance courses alongside a much larger in-person programme. Its handbook covers more than 300 rural skills, and its instructors have trained more than a hundred million producers and workers cumulatively since 1991. If you want to know where an AI curriculum for Brazilian agriculture actually lands, it is inside SENAR's course catalogue, next to hand milking, canned in fifty-minute modules, delivered by a person the farmer already trusts.

Chillán's blueberry report

South of the border, INIA Chile presented in late May 2026 a tool called Reporte Digital, developed by the Data Science Unit at INIA Quilamapu in Chillán. It fuses IoT orchard sensors, physiological models of fruit ripening, and a data-science backend into a per-lot maturity forecast for blueberry exporters. The point of the tool is not the algorithm. The point is that a Chilean exporter can walk into a Rotterdam or Philadelphia buyer meeting with a forecast of ripeness uniformity per lot, and that this forecast reduces the buyer's discount for uncertainty. That is the shape of the value.

Two weeks earlier, INIA's digital agriculture specialist Stanley Best flew to a Taiwan workshop on AI applied to agriculture and came back with an observation the marketing decks tend to skip. Taiwan's agri-AI teams, he said, are running most of their farmer alerts and recommendations over WhatsApp, because that is the surface the farmer is already looking at. The technology sits behind the message. The message arrives on the app the farmer opens anyway. This is a lesson so old it feels embarrassing to repeat, and it is the lesson the region's more expensive rollouts keep missing.

A note from a coffee co-op

I visited a small cooperative in the Sul de Minas coffee belt a few years ago on a Real AI trip. The co-op had bought a soil-moisture and micro-climate sensor package for a subset of member farms, on a subsidy from a public-private programme. The dashboards were beautiful. The alerts came through fine. Six months in, the farmers who had the boxes were checking the numbers roughly twice a week. Twelve months in, the numbers were being read once a month by the co-op's own agronomist and no longer by the growers themselves. The single reason, given plainly by two of the growers over coffee, was that the person who came out to install the boxes had never come back to explain what a soil-moisture number should change about the day's work. There was no second visit budgeted in the pilot. The technology worked. The training did not compound.

That is what hides inside Dibbern's polite methodology. What determines adoption is the network of humans around the tool. The Brazilian and Chilean examples above work because Embrapa and INIA both maintain that network as a public good and both have decades of institutional memory to draw on. The private-sector rollouts I have seen fail have failed in almost the same way every time. Someone pays for the sensor and the demo. No one pays for the third visit.

What the money says

Brazil's compute-side ambitions are large enough that the extension gap should be visible from the finance ministry. BNDES is preparing to launch an AI and data-centre investment fund in early 2026, with an initial contribution the bank's own planning director places between R$500 million and R$1 billion. By early August 2026, BNDES had committed roughly half of Brazil's US$2.1 billion in cumulative AI-related public funding, most of it aimed at capacity, chips, and data-centre buildout. The line item for agricultural extension in that same envelope, so far as I can find in the public materials, is not a line item at all. Embrapa's own operating budget grew this year and its Research and Agricultural Innovation Program received a serious top-up, which is welcome. It does not close the gap between compute financed and adoption financed.

The optimistic reading, and there is a real one, is that private capital will do the extension work. The CGIAR Initiative on Digital Innovation, in a widely-cited December 2024 note by IRRI's Shalini Gakhar and Niyati Singaraju, argued that public-private partnerships and human-centred design would carry AI extension across the smallholder line at scale. I do not think that analysis is wrong. I think it undercounts the drag. Their own workshop conclusions named the digital divide as the largest single obstacle, and named capacity-building programmes as the largest single solution. Both of those items are unglamorous, off-invoice, hard to measure quarterly, and, in Brazil at least, still under-funded relative to the compute side of the same bet.

What the paper asks of you

Read Dibbern and her co-authors' five-dimension diagnostic against any Latin American agri-AI pitch you are handed this year. The dimensions are: the technology's own requirements; the property and its geography; the readiness of digital infrastructure; the farmer's own profile and capability; and economic viability. There is no clever weighting that hides how much the third and fourth items dominate the outcome. If the infrastructure is thin and the training is thin, no amount of model performance moves the field. If both are strong, even a mediocre model can produce a visible year-on-year improvement in a blueberry pack-out, a coffee-cherry ripeness call, or a fish-tank oxygen adjustment. This is what a public agricultural research service knows and a data-centre financier is still learning.

What sits with me

Sit with the timing for a moment. Embrapa's Semear Digital methodology paper landed in the same fortnight that a BNDES fund is being drawn up to underwrite the compute layer of Brazil's next agricultural decade. If both efforts hold, the country will end this decade with more sovereign compute for agricultural analytics than it can plausibly use, and, unless the extension budget follows, the same uneven adoption curve every previous wave of agricultural technology has produced, from soil-testing kits to GPS receivers. The pattern is not new. Neither is the corrective. Somewhere between the paper's polite Score of 3 and its Score of 4 sits every farm in Latin America that will or will not use the tools this decade is about to build.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan, an Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.