The camera fitted to an orchard vehicle has a simple assignment: look at the trees and count buds and fruit. A farm manager may have a less glamorous task, but sometimes the more consequential one. In research described by the University of California, Davis, a manager familiar with the history of a particular block adjusted the camera-based estimate downward by 14% after considering previous yields and recent weather. The resulting estimate proved more accurate. A machine had collected useful observations; a person supplied information the camera did not have.[1]
The example comes from research at 16 fruit-tree operations in California and Washington, reported by UC Davis on 22 October 2026. The original paper, published on 27 September in New Technology, Work and Employment, specifies an 18-month data-collection period between 2023 and 2025 and 165 formal interviews. It examines what happens after automated tools reach a farm and must be adapted to variable crops, weather and working arrangements.[1][2]
Counting fruit is only one part of forecasting a harvest
Orchard automation covers different activities: traversing rows, photographing flowers or fruit, estimating harvest volume and attempting more demanding jobs such as picking. Some tasks require repeated counting or predictable movement. A camera may be better suited to collecting consistent images than a worker walking the orchard and sampling every few trees. Yet a harvest forecast also depends on whether blossoms survived weather events, how particular blocks have performed in earlier seasons and what happens after the image was taken.[1]
Three technologies, three kinds of work
| Technology examined | Role on the farm | Finding in the study |
|---|---|---|
| Automated fruit counting | Generate observations for yield estimates | Widely integrated, with human interpretation and correctionsSource 2 |
| Motorised orchard platforms | Reorganise how crews work around trees | Active or partial use on more than half of the operationsSource 2 |
| Automated harvesters | Attempt to automate picking | Largely experimental, under trials rather than routine useSource 2 |
This distinction changes the economics of automation. If a machine eliminates one manual counting exercise but requires calibration, data review and interpretation, then the effect on a workforce is a redistribution of activities rather than necessarily the elimination of a whole position. For managers, the immediate question is whether the new workflow improves decisions after labour, equipment, maintenance and supervision are counted—not whether a demonstration looks technologically advanced. That is a different test from measuring how many objects the camera recognises.[1]
Agricultural workers can be sources of operational knowledge that is difficult to represent in a dataset. A person may know which row often underperforms after particular weather, where uneven ground changes a machine’s route or whether the visible fruit count overstates what will reach saleable quality. Such expertise is local, accumulated and partly tacit. The study indicates that operators and managers adapt technology to these conditions, rather than simply accepting a manufacturer's default workflow.[1]
Who performs and receives credit for the extra work?
The paper argues that workers take on additional cognitive labour when they interpret, correct and adapt the systems. That responsibility does not necessarily bring corresponding recognition or compensation. This makes implementation a question about workplace authority and skills as well as equipment performance.[2]
The qualitative design explains how work changed in these operations rather than forecasting employment across agriculture. The manager's 14% correction is a documented example of local judgement, not a standard adjustment for another orchard. Its value lies in showing what the camera's count omitted: the block's history and the effect of recent conditions.[1][2]
A tool that accepts local corrections and preserves the reasoning behind them can combine consistent observations with accumulated knowledge. In these orchards, that knowledge was part of the technology's working infrastructure: people had to supply it for the automated observations to become reliable decisions.[1][2]
Sources & further reading
- The Assumption About Farm Automation That May Be Wrong University of California, Davis · 22 Oct 2026 · Accessed 27 Oct 2026
- Epistemological Friction: Embodied Expertise and the “Last Mile” of Agricultural Automation in the US Tree-Fruit Industry New Technology, Work and Employment · 27 Sep 2026 · Accessed 27 Oct 2026