AI in Agriculture: When Advice Does Not Travel

AI in Agriculture: Advice That Does Not Cross Borders

AI in agriculture works remarkably well inside the conditions it learned. Computer vision on drones and phone cameras can spot pest pressure, fungal infection, and nutrient stress at a speed and coverage no scouting team matches, and advisory tools now put that capability in front of growers directly rather than through an agronomist. The failure mode is geographic rather than technical. Models trained on one region, one soil profile, and one set of varieties produce equally fluent answers outside that envelope, and nothing in the output distinguishes a well-grounded recommendation from an extrapolated one.

What Transfers Across Regions and What Does Not

Some agronomic knowledge is close to universal. Some is local in ways that look universal until it fails.

Why AI in Agriculture Advice Breaks at the Border

Training data follows commercial agriculture. The largest labelled image sets, the deepest soil records, and the longest yield histories come from a handful of intensively farmed regions, so models inherit those conditions as their implicit default. When a question arrives from outside that envelope, the model does not decline. It maps the unfamiliar input onto the nearest familiar pattern and answers with the same fluency it would use on home ground.

This is ordinary distribution shift, but agriculture makes it unusually costly. A wrong retail recommendation costs a return. A wrong agronomic recommendation costs a treatment window, and treatment windows do not reopen. Add that many growers receiving this advice have no agronomist to sanity check it, and the error path runs directly from a confident sentence to a field.

The Seasonal Irreversibility Problem in AI in Agriculture

Most software advice can be corrected on a short loop. You try something, see the result, and adjust. AI in agriculture operates on a loop measured in months, and often in a single annual attempt. A variety chosen in error is not discovered until harvest. A missed spray window cannot be revisited the following week with the same outcome. A soil amendment applied on wrong assumptions influences the next cycle as well as this one. That structure means the usual argument for tolerating occasional model error, that mistakes are cheap and correctable, simply does not hold here.

Five Questions Where Geography Changes the Answer

These are the questions where an answer trained elsewhere is most likely to be fluent and wrong.

1. What pest or disease is this. Visually similar symptoms map to different organisms in different regions, and the treatment follows the organism.
2. Which product and at what dose. Registrations and legal maximums are set nationally, so a correct answer in one country can be an offence in another.
3. When should I sow or harvest. Timing depends on local climate patterns and day length that a general model has no reliable access to.
4. Which variety suits my plot. Cultivar availability and performance are intensely local, and recommendations often name varieties that are not sold in the region at all.
5. How much water or fertiliser. Rates depend on soil baselines and water regulation, both of which vary at a scale below the national level.

Pros and Cons for Advisors and Growers

Advisory tools are genuinely valuable. The value is uneven across question types.

• Pro: coverage no scouting team can match. Image-based detection across a whole field catches pressure earlier than sampling does.
• Pro: access where agronomists are scarce. For many growers the realistic alternative is no advice at all, not expert advice.
• Pro: strong on explanation. Models are good at teaching the reasoning behind a practice, which builds capability rather than dependence.
• Con: no signal when out of distribution. The answer looks identical whether the model is on familiar ground or extrapolating.
• Con: regulatory advice is unreliable. Product registrations and limits change by country and by season, and models are frequently out of date.
• Con: errors are discovered late. Feedback arrives at harvest, long after the decision could have been changed.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how advisory errors play out in practice rather than presented as verified case studies.

A grower photographs leaf damage and receives a confident identification of a pest common in temperate cereal systems. The actual organism is a regional pest with a different life cycle. The recommended treatment is applied at the wrong point in that cycle and achieves very little, and the window for an effective intervention passes.

An advisory chatbot recommends a fungicide by trade name at a specific rate. The product is registered in the region the model learned from and not in the grower's country. Following the advice would breach residue rules for an export crop, which nobody would discover until testing at the buyer.

A cooperative asks several models the same variety selection question and finds that answers diverge sharply once the question includes local sowing dates. The disagreement itself is the useful output. It tells the agronomist exactly which part of the recommendation needs local verification rather than which model to trust.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

How to Localise Before You Trust an Answer

Three habits remove most of the risk without removing the benefit. Put the location, crop, variety, growth stage, and soil context into the question explicitly, because a model given no geography will default to the geography it learned. Treat any answer naming a specific product, dose, or legal limit as unverified until checked against the national registration list, since that category ages fastest and carries legal consequences. And route irreversible decisions, meaning anything tied to a season or a treatment window, through a person who knows the region before acting.

Where a local extension service or agronomist exists, the model is best used to prepare the conversation rather than replace it. Arriving with a shortlist of possibilities and specific questions makes scarce expert time go considerably further than arriving with a photograph.

Why Talkory Wins

The hardest thing to detect in agronomic advice is whether a model is on familiar ground. A single answer never tells you. Talkory runs the question across GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 at once, and the spread between them is a usable proxy for that uncertainty. Broad agreement on a pest identification or a general practice suggests the question sits inside shared, well-established knowledge. Sharp divergence on a variety, a dose, or a sowing window usually means the question is regional, and regional questions are exactly the ones that need a local source rather than a confident sentence.

Final Verdict

AI in agriculture is one of the more genuinely useful applications of the technology, particularly where expert advice was never available in the first place. The limitation to respect is that agronomic knowledge is local in ways the models rarely signal, and that the feedback loop runs a whole season rather than a few minutes. Give every question its geography, verify anything involving a registered product against the national list, run the high-stakes questions across several models and treat disagreement as a flag, and keep a human with local knowledge on the irreversible calls.