AI, jobs and agrifood: theory meets practice

The debate on the impact of artificial intelligence on labour markets is often marked by polarised views, ranging from expectations of rapid transformation to concerns about widespread job displacement. Recent empirical evidence, however, points to a more gradual and complex trajectory. 
A relevant contribution comes from a recent study by Anthropic, which introduces a new metric, “observed exposure”, to estimate the actual extent to which occupations are affected by AI. The approach combines theoretical capability, real-world usage data, and the nature of job tasks, assigning greater weight to automated and work-related uses. 
The main finding is the existence of a significant gap between capability and adoption. While a large share of tasks is theoretically compatible with AI, real-world implementation remains partial. In highly exposed sectors such as computer and mathematical occupations, theoretical exposure exceeds 90% of tasks, while actual coverage is around 33%. 
This gap is not trivial. It highlights that AI diffusion depends not only on technological readiness, but also on regulatory constraints, integration requirements, human oversight, and organisational adoption dynamics. 
From a labour market perspective, current data does not show systematic effects on unemployment. Time series analysis reveals no significant increase in unemployment rates for highly exposed occupations, even after the widespread adoption of generative AI tools. However, more subtle signals are emerging, including an estimated 14% decline in entry rates for younger workers (aged 22-25) into highly exposed roles. 
At the same time, the study finds that more exposed occupations tend to be associated with higher levels of education and income, and a higher share of female workers. This reshapes the profile of the groups most likely to be affected by ongoing changes. 
Overall, the evidence suggests a transition that is underway but not yet disruptive at scale. At this stage, AI appears to augment human work more than replace it, with impacts unfolding gradually and unevenly across sectors.

The agrifood sector: constraints and systemic opportunities 

Within this broader context, the agrifood sector displays specific structural characteristics. A significant share of agricultural work involves physical tasks, environmental variability, and adaptive capabilities that are difficult to automate. These features make the sector less immediately compatible with AI-driven automation based on language models. 
At the same time, this complexity creates different opportunities. Rather than focusing on direct task substitution, AI can contribute by improving coordination and efficiency across the value chain, from production planning to logistics, waste reduction, and better alignment between supply and demand. 
This points to a systemic dimension of innovation, where value is generated not only through task automation, but through the integration of technologies, data, and processes. 
Important constraints remain. High upfront investment costs, the seasonal nature of agricultural work, and the need for flexible, multi-purpose systems may slow adoption. In this context, emerging models such as robot-as-a-service, enabling temporary access to technology, may help reduce barriers, despite challenges related to fluctuating demand.

Making AI work in real agrifood systems

In light of these elements, the gap between theoretical capability and actual use should not be seen as a limitation, but as a field of action, where experimentation, real-world validation, and adaptation to production environments become essential. 
It is precisely within this gap, between technological capability and real-world application, that initiatives such as agrifoodTEF can take on a truly strategic role. This is not only about testing solutions, but about reducing the uncertainty that currently slows down adoption: understanding what works, under which operational conditions, and with what economic returns. 
In a context where AI does not diffuse automatically, the ability to validate technologies in real-world environments becomes critical. This includes, for example, assessing the actual impact of decision-support systems, evaluating how robotic components integrate with agricultural processes, or testing how digital tools can improve coordination across different stages of the value chain. 
At the same time, a less visible but equally important need emerges: bridging levels that remain structurally fragmented (from primary production to processing and distribution) moving beyond isolated, segment-level optimisation. In this perspective, experimentation is not only about technology, but also about organisational models and modes of collaboration across the value chain. 
In other words, moving from theoretical innovation to widespread adoption requires infrastructure capable of supporting this transition, reducing adoption costs and making tangible benefits visible where they would otherwise remain latent. 
Based on current evidence, the issue is not to anticipate large-scale job displacement, something the data does not yet support, but to understand and guide a gradual transformation. One driven by incremental adoption, skill reallocation, and the adaptation of production models. 

This transformation is unlikely to unfold in a linear or immediately visible way. Precisely for this reason, it requires robust analytical frameworks and real-world experimentation environments. This is where a significant part of AI’s impact on the agrifood sector will be shaped in the years ahead.