Main body
To find out whether an AI agronomy tool knows your field, ask it a question where local conditions must change the number — a nitrogen rate, a planting date, a spray threshold — and ask it twice, for two different fields. Then check four things: whether it used local weather and soil, whether it named a source you can verify, whether it admits what it does not know, and whether it holds up in the language the farm actually speaks.
Why a fluent answer is not a field-specific answer
Fluency and field-specificity are different properties, and buyers routinely test the first while believing they are testing the second.
A 2024 study in Scientific Reports by Ibrahim, Senthilkumar and Saito of AfricaRice put 32 real farmer questions from Kano State, Nigeria, to ChatGPT and to six local extension agents. Four local rice experts, working blind, preferred the chatbot's answers in 78% of cases and rated them "good" on average, against "acceptable" for the agents.
The same study shows where that fluency broke down. The chatbot's seed rate was too high, its dry-season planting time was wrong, and its fertilizer rate and timing scored below the human agents. Those are precisely the answers a farmer spends money on.
So the useful signal is not quality. It is variance. If a tool returns the same number for two fields with different soils, it is describing a crop, not a field.
Test 1: does the AI agronomy tool use local weather and soil?
A tool passes this test when its recommendation changes as the field changes, and when it can tell you which local inputs produced the change. Ask the same question for two fields with different soil texture, rainfall or previous crop. Identical answers mean generic advice.
Site-specific advice is worth the trouble. Across 1,594 side-by-side on-farm trials in India's Indo-Gangetic Plains, Sapkota and colleagues reported in Scientific Reports (2021) that more than 80% of farmers using CIMMYT's Nutrient Expert tool raised both yield and income, and that the size of the gain depended on crop, agro-ecology, soil properties and the farmer's existing practice. In the Philippines, IRRI and DA-PhilRice reported in 2021 that Rice Crop Manager recommendations added about 0.4 t/ha per season, worth roughly US$100/ha.
Local data also has a resolution limit worth knowing. ISRIC's SoilGrids 2.0, documented by Poggio and colleagues in SOIL (2021), maps soil properties globally at 250 m — one cell covers 6.25 ha. FAO economists Lowder, Sánchez and Bertini found in World Development (2021) that 70% of the world's farms are under 1 ha. For most farms on earth, a global soil layer is a neighbourhood estimate. A good tool says so; a weak one presents the cell value as a measurement.
Test 2: can you check the source?
A tool passes when it names a source specific enough that a trained person can open it and find the same number. An extension bulletin, a national recommendation, a published trial. "Research suggests" is not a source, and a link to a general website that does not contain the figure is worse, because it looks like one.
Ask directly: where does this number come from? Then verify one number per answer. Writing in Nature Food in 2023, Tzachor and colleagues argued that language models in agricultural extension need human experts inside the design loop. A checkable citation is the mechanism that keeps you, the agronomist, inside it after the tool ships.
Test 3: does it say when it does not know?
A reliable tool states uncertainty and names the missing information that would reduce it. This matters most when it is working from regional estimates instead of a soil test.
Test it deliberately. Remove a detail the answer depends on — the previous crop, the irrigation method — or ask about a minor crop in your district. A tool that knows its own limits asks a follow-up question or flags the gap. A tool that returns the same confident number either way has told you something important about every other answer it gave you.
Test 4: does it work in the language the farm actually speaks?
Here the thesis inverts. Across languages the number should not move. Ask the same question in English and in the language your farmers use, and compare the rates and dates, not the tone.
The gap is measurable. IrokoBench, presented by Adelani and colleagues at NAACL in 2025, evaluated language models across 17 African languages and found a wide gap against English and French — the strongest open model reached only 63% of the leading proprietary model's score. If a rate shifts between languages, the tool is reliable for you and not for the farms you advise. Where farmers send voice notes, test voice too.
The same question, two answers
Both tools are asked: Corn after soybean, central Iowa, planting next week. How much nitrogen? The answers below are illustrative.
Generic: "Corn generally needs 150–200 kg N/ha, applied in split doses. Because soybeans fix nitrogen, reduce your rate by 30–50 kg. Consult a local expert."
Field-aware: "For corn after soybean in the main Iowa region, Iowa State University's Corn Nitrogen Rate Calculator gives about 139 kg N/ha (124 lb N/acre) at a nitrogen-to-corn price ratio of 0.10 — US$0.35/lb N and US$3.50/bushel. Do not subtract a soybean credit; the rotation effect is already inside that figure. Heavy rain is forecast this week, so consider side-dressing part of the rate. I have no soil nitrate test for this field; a late-spring nitrate test would confirm the adjustment."
| Check | Generic | Field-aware |
| Local data | None | Region, previous crop, forecast, prices |
| Checkable source | None | Named extension calculator |
| States uncertainty | No | Flags the missing soil test |
| Number you can act on | A 50 kg range, then a wrong deduction | One rate, one adjustment, one next step |
the generic answer reads as competent and contains a verifiable error. Iowa State Extension states plainly that no soybean credit should be subtracted from these rates, because the trials behind them are already corn-after-soybean trials.
What the difference costs
A farmer who takes the middle of the generic range applies 175 kg N/ha instead of 139 — an extra 36 kg N/ha.
- Nitrogen at US$0.35/lb is about US$0.77/kg N
- Extra cost: 36 × 0.77 ≈ US$28/ha (US$11/acre)
- Over 200 ha (494 acres): about US$5,600 per season
- As urea at 46% N, that is an extra 78 kg urea/ha
The previous crop matters more than the price. The same calculator gives about 203 kg N/ha (181 lb N/acre) for corn after corn. A tool that ignores rotation can be wrong by 64 kg N/ha before weather enters the calculation. Substitute your own prices; keep the method.
Running the test
Pick three questions whose correct answer depends on location: a fertilizer rate, a planting date, a pest threshold. Ask each for two contrasting fields, in two languages. Score the twelve answers against the four checks, then verify one number per answer against its named source. A tool that passes is one you can put in front of the farmers who trust you.
Valora Earth is built for this test: it assembles each farm's own soil, weather and crop history behind every recommendation, and answers on WhatsApp in more than 30 languages.
Frequently asked questions
Can a general AI chatbot give reliable fertilizer advice?
It can explain principles well but should not set rates. The 2024 AfricaRice study in Nigeria found experts preferred ChatGPT's answers overall, yet it performed worse than human agents specifically on seed rate, planting time, and fertilizer rate and timing. Verify any rate against a local recommendation before it reaches a field.
How accurate are free global soil maps for a single farm?
They are regional estimates, not measurements. SoilGrids 2.0 predicts soil properties at 250 m resolution, so one cell spans 6.25 ha — larger than most farms worldwide. Use them to plan sampling and flag likely constraints, then replace the values with laboratory results once you have them.
What should an agronomist check before recommending an AI crop advisory app?
Four things: the recommendation changes when the field changes, every number traces to a source you can open, the tool states what it does not know, and answers hold steady across languages. Test with location-dependent questions and verify at least one figure per answer yourself.
Use Valora On-the-Go!
