Three years and 165 interviews across 16 orchards led UC Davis researchers to a conclusion the automation debate mostly skips: these tools need the knowledge they were meant to replace.

The argument about farm automation is usually conducted as a forecast. Machines will replace workers, the question being how many and how soon, with one side counting the jobs at risk and the other promising the labour shortage will be solved.

A study in New Technology, Work and Employment from the University of California, Davis suggests both sides have the mechanism wrong. Researchers spent three years on 16 fruit tree operations in California and Washington, conducting 165 in-depth interviews with workers and growers, and found automation changing who does what rather than removing the need for people.

“Automation isn’t something that just happens to a farm or happens to a labor force,” said M. Anne Visser, professor of community and regional development in the UC Davis Department of Human Ecology and the study’s lead author. “People have to actively make it happen.”

Fourteen per cent

The study’s sharpest illustration involves a yield forecast. One operation used a camera-based system that counts buds and fruit to predict how much a block will produce — the kind of tool that is supposed to replace a slow manual process.

The manual process it replaces is genuinely crude: a worker counts the buds on every sixth or seventh tree and the result is extrapolated. A camera that images everything should be strictly better.

It was not. The system tended to overestimate yield, and its output needed interpretation. A farm manager with more than a decade on that farm routinely lowered the camera’s counts by 14 per cent, drawing on how those particular blocks had performed historically and what the recent weather had been. His adjusted estimates were more accurate than the cameras’.

That is a precise description of what the paper calls the last mile of agricultural automation. The camera does the counting, which is the part that was tedious. What it cannot do is know that this block overperforms on paper, or that the bloom came during a cold snap, or that a fraction of what it has counted will not set fruit. Someone has to supply that, and the only people who have it are the ones who have watched the same ground for years.

The number matters commercially, incidentally. Yield forecasts drive decisions about labour contracts, packing capacity, storage and forward sales, and a systematic 14 per cent overestimate propagated into those decisions would be expensive.

Why tree fruit is harder

The paper’s title refers to the US tree-fruit industry specifically, and the choice of sector is not incidental.

Fresh-market specialty crops are harder to automate than row crops like corn and soybeans, because quality matters as well as yield. A combine harvesting soybeans needs to get the beans off the field; a crew picking apples for the fresh market has to judge ripeness, colour and size, avoid bruising, and leave the immature fruit on the tree for a later pass.

Many autonomous harvester prototypes cannot match the knowledge needed to pick fruit, the study reports — and the economics of trying are unforgiving in a way they are not for row crops. Orchards are multiyear investments, so growers cannot switch crops within a single season in response to a technology that works or does not. A planting decision commits the land for a decade or more.

This is the structural reason tree fruit has resisted a transition that transformed grain farming generations ago, and why the sector remains dependent on hand labour while being the one with the loudest labour shortages.

Replacing a task is not replacing a job

“Replacing a task on a farm is not the same as replacing a job,” Visser said. “These tools do not simply remove jobs on farms. They often change who does what and reshape the work around technology.”

The distinction has consequences that the displacement framing obscures. A camera system does not eliminate the yield estimator; it converts that person into someone who runs, calibrates and corrects a camera system. Whether that is a better job — more skilled, better paid, more secure — or a worse one is an open question that depends on who gets it and on what terms, and it is a question nobody asks if the only outcome being counted is the headcount.

The study’s authors see upside alongside this. Automation can still improve efficiency, occupational health and planning around labour shortages as workers move into other farm roles. Occupational health is not a small consideration in an industry where the physical work involves ladders, repetitive lifting and heat exposure.

The flipside, which the study’s framing implies rather than states, is that a system dependent on a manager with a decade of local knowledge has a vulnerability. That expertise is embodied, it is specific to particular blocks of particular orchards, and it leaves when the person does.

Why the framing matters

Visser’s stated motive for the work is about the shape of the public argument. She says the hope is that the research helps advance the conversation surrounding agricultural automation, expressing concern that a framing built on displacement divides people quickly.

That concern is reasonable on the evidence. A debate conducted as workers-versus-machines sorts growers and labour advocates into opposed camps before anyone examines what the technology actually does on a farm — and what it does, on these 16 operations, is neither replace people nor leave the work unchanged.

The study’s method is its main limitation as well as its strength. Sixteen operations and 165 interviews across two states give a detailed picture of how these tools are actually used, which no survey would capture; they do not establish how typical any of it is, and qualitative fieldwork cannot tell you what will happen when the next generation of harvesters arrives.

What it does establish is that the knowledge in question exists and is load-bearing. Any forecast that treats farm work as a set of tasks to be mechanised one by one is leaving out the part that corrects the machine.

The study’s co-author is Alex Bladimir Hernandez of UC Davis. The work was funded by the U.S. National Science Foundation and the University of California Office of the President, and the authors declare no conflicts of interest.

---

Notes

1. Visser, M. A., Hernandez, A. B. “Epistemological Friction: Embodied Expertise and the ‘Last Mile’ of Agricultural Automation in the US Tree-Fruit Industry.” New Technology, Work and Employment, 2026. DOI: 10.1111/ntwe.70047.
2. University of California – Davis. “The assumption about farm automation that may be wrong.” EurekAlert!, 22 October 2026. https://www.eurekalert.org/news-releases/1148942.