Anthropic's Economic Index looked at millions of real conversations with Claude and mapped them onto the work people actually do, using the U.S. Department of Labor's occupational categories. It is one of the clearest pictures we have of where AI is genuinely being used, not where people say it will be used someday.
Agriculture came in dead last. Occupations involving a high degree of physical labor, such as farming, fishing, and forestry, were the least represented in the data, at 0.1% of queries. For comparison, computer and mathematical work made up 37.2% of all queries. Software work shows up several hundred times more often than farming.
The interesting question is not whether agriculture is behind. It clearly is. The question is why, and the answer is not the one most people reach for. It is not that farmers are stubborn or backward. Several real reasons stack on top of each other, and they are worth separating out.
Reason one: today's AI was built for screens, not fields
The biggest reason is baked into what this generation of AI actually does well. The whole Economic Index is built by matching conversations to text-and-screen tasks, writing, coding, analysis, and the heaviest users reflect that. The lightest users were jobs involving a large degree of manual dexterity, from shampooers to obstetricians.
Farming lives entirely in that manual, physical world. A language model is superb at drafting a memo and useless at walking a field, feeling the soil, or judging whether a plant looks stressed. Most of a farmer's actual work is not a document problem, so most of it is simply outside what current AI can touch. The low number is, first and foremost, a statement about the tool, not the farmer.
Reason two: the work is judgment in context, which AI is weakest at
Even the decision-making parts of farming resist automation. Knowing when to irrigate, timing a treatment against weather that shifts by the hour, applying years of accumulated knowledge about one specific piece of land, this is judgment exercised in a messy, changing physical environment. That is precisely the kind of reasoning current AI handles worst. A recent socio-technical study of Midwestern U.S. farmers points to a problem unique to AI: the "black-box" issue, where farmers cannot verify how a recommendation was reached, which makes trusting it much harder than trusting a simpler tool.
Reason three: even where AI could help, the payback is unclear
For the narrow slices where AI genuinely can help, economics kicks in. In McKinsey's Global Farmer Insights survey, 49% of European farmers cited implementation costs as prohibitive and 45% said they could not justify the investment. Adoption also tracks farm size closely: large farms over 2,500 acres are 45% more likely to adopt agtech than small farms under 100 acres, because of the scale needed for a positive return. On a thin-margin smallholding, "this might help" is not enough. If a farmer cannot see the money it saves this season, the tool does not get adopted, however clever it is.
Reason four: the information and trust gap
There is also a quieter barrier. In a two-year Brock University study of agricultural automation in Ontario, Professor Charles Conteh found that even when tools were technically sound and commercially available, adoption stalled. He describes what he calls "information gap syndrome": many farmers simply do not know which AI tools exist or which are relevant to their operation. His conclusion is blunt: the problem is not a lack of sophisticated tools, but a lack of systems that help farmers understand, integrate and trust these technologies.
What farmers actually reward: a lesson from the field
Spend time with farmers and the pattern behind these numbers becomes concrete. On a research trip our team took to the rice belt of Akita, in northern Japan, the farmers we met were not remotely anti-technology, they used GPS-guided tractors, drones, and precision seeding readily. But they judged every tool by a single standard: it had to prove its result first. As they put it, the technology did not increase their harvest, it reduced their workload and cost. The return was in labor saved, not magic.
One rice farmer described how he handles pest pressure in a line that has stuck with us since: "I need to check and see if that pest exists in my field. And if I find it, I will do something about it. But if I don't find it, I'm not going to do anything." No spraying on assumption. The cooperative sends a regional alert, the farmer walks their own field, and they act only on what they actually find. That is the real template for AI in agriculture: not a system that overrides the farmer's judgment, but one that sharpens it, and that reaches the farmer through the cooperative they already trust.
The real explanation: two hard problems at once
Here is the synthesis. Most industries face one of two challenges: either the technology cannot do the job yet, or it can but adoption lags. Agriculture is rare in facing both at the same time. It is one of the hardest domains for current AI to be useful in, because the work is physical and judgment-heavy, and it is one of the hardest environments to adopt into, because of cost, trust, and access.
That combination is why agriculture sits at the very bottom of the chart. Not one barrier, but the two hardest barriers reinforcing each other.
And that is also why the opportunity here looks different from everywhere else. The Anthropic data offers a clue about the right approach: across the whole economy, AI use leaned toward augmentation (57%), where it collaborates with and enhances people, rather than automation (43%), where it performs tasks directly. The winning move in agriculture is not to replace the farmer, it is to augment them: point AI at the narrow, high-value tasks where it genuinely fits, spotting a pest pattern in field imagery, processing sensor and weather data at scale, and deliver it in a way that makes the payback obvious and earns back trust. The low number is not a reason to look away. It is a fairly precise map of the work that still has to be done.