Before LLMs, Machine Learning Models Were Already Changing Work

Part 1: The shift that already happened 

It is easy to think AI suddenly arrived at work when LLMs became visible, as if the shift began only when machines started writing emails or answering questions in fluent language, but that is not really true. 

Long before conversational AI entered daily workflows, machine learning models were already changing how work was done, quietly, unevenly, and often without the kind of public attention we see today, even though their impact was already shaping decisions across industries. 

They were scoring credit risk, detecting fraud, forecasting demand, optimising routes, predicting equipment failure, identifying crop disease, and supporting fisheries and environmental planning, not by replacing entire professions overnight but by absorbing specific forms of work such as prediction, classification, ranking, anomaly detection, and optimisation, which tend to scale better in systems than through manual effort. 

That matters because much of today’s AI discussion treats LLMs as the beginning of workforce disruption, even though many industries have already lived through earlier waves of model-driven change; what is new now is not that AI entered work, but that it became visible, conversational, and much closer to everyday business interaction. 

The shift started earlier than most people admit 

For years, machine learning lived inside systems that businesses did not always describe as AI, appearing instead as scoring engines, recommendation systems, optimisation tools, and forecasting pipelines, which made their impact easier to normalise and harder to isolate. 

A credit analyst who once reviewed every case manually now focuses on exceptions; a fraud team works on model-flagged signals rather than scanning everything; and a logistics planner no longer builds routes from scratch because systems generate optimised options that humans evaluate and adjust, rather than making decisions entirely on their own. 

The same pattern appeared in agriculture and fisheries, where models supported disease detection, yield prediction, and resource planning, gradually shifting human effort from observation toward interpretation, intervention, and edge-case handling. 

Models rarely replaced entire roles in one clean step, but they compressed tasks, reduced manual review, narrowed decision surfaces, and reshaped where human effort was actually required. 

Why this change went relatively unnoticed 

Earlier model-driven change created less public anxiety because it was domain-specific, embedded in systems, described as analytics or automation, and introduced gradually rather than through visible interfaces. 

It lived inside lending, manufacturing, insurance, agriculture, logistics, and operations, and it entered workflows one decision point at a time, making it easier to normalise even as the underlying structure changed. 

But quiet change is still change, and in many organisations, the workflow had already shifted long before the public conversation caught up. 

What models actually changed 

The better question is not whether models replaced people, because that framing is too blunt and misses the structural nature of the shift. 

A better question is which work models were adopted, because that is where the real transformation occurred. 

Machine learning proved especially strong at prediction, classification, ranking, anomaly detection, optimisation, and recommendation, which map directly to operational work such as predicting defaults, detecting fraud, ranking leads, or optimising routes before a human even engages. 

As these capabilities improved, workflows reorganised around them, with humans reviewing less volume, handling more exceptions, defining thresholds, and owning outcomes without making every first-level judgment themselves. 

The debate often swings between extremes — AI replaces everything or changes nothing, but the reality is structural: AI changes how work is distributed and where judgment sits. 

Why LLMs feel different 

LLMs did not create this shift, but they made it visible. 

Because they operate through language, they feel closer to human intelligence, making the change more direct and more personal, as a hidden scoring engine becomes something that can explain, summarise, or respond in real time. 

This visibility makes the impact appear broader, but earlier systems had already reshaped workflows; LLMs simply exposed it. 

The enterprise mistake 

Once AI becomes visible, discussions collapse into model-level questions: which model is better, which benchmark matters, whether to fine-tune, whether to use RAG, or which framework to adopt. 

These are valid questions, but they are not the whole system, because a better model does not define accountability, a more accurate predictor does not define authority, and a stronger assistant does not explain who owns the outcome when workflows begin to change around it. 

The deeper issue is what happens to the surrounding system when capability changes, and that is where the discussion must move next. 

Part 2: When models change structure, not just tasks 

Enterprise structure changed before LLMs, too 

The shift was not only in tools but also in enterprise structure, as models changed how work was distributed across the execution, monitoring, and decision-making layers, gradually moving organisations away from labour-heavy execution toward system-guided coordination. 

In agriculture, this structural change is visible over decades of productivity and labour transformation. 

These figures reflect long-term productivity growth combined with declining labour intensity, driven first by mechanisation and later by digital and precision agriculture systems. 

Before model-driven operations, work relied more on manual inspection, localised judgment, and experience-led planning, with decisions emerging closer to the field. 

After models and digital systems, decisions begin earlier in data, systems narrow where humans need to focus, and effort shifts toward exceptions and interventions, effectively moving organisations toward decision-heavy coordination rather than raw execution. 

Fisheries show a similar pattern 

Earlier operations depended heavily on manual observation, fragmented reporting, and localised experience, which limited the speed and consistency of decisions. 

With digital monitoring and model-driven insights, work shifts toward continuous monitoring, earlier risk detection, centralised oversight, and stronger dependence on analysis, compliance, and system-driven planning. 

The workforce does not simply shrink; it gets recomposed, with some roles reducing, others narrowing, and new roles emerging around monitoring, interpretation, and intervention. 

Uneven adoption matters 

These shifts do not happen evenly across countries or sectors. 

Where infrastructure, capital, digital access, and technical capability are strong, adoption moves faster; where these are weaker, adoption becomes slower, more selective, and often hybrid, with older and newer systems coexisting rather than replacing each other. 

This creates an uncomfortable reality in which the parts of the economy that could benefit most from better prediction, earlier detection, and stronger planning are often the ones with the highest structural constraints, leaving lower-income populations concentrated in less productive systems. 

In fields like software engineering, adoption can move faster because deployment is lighter and less dependent on physical infrastructure, while in robotics, automation, and field operations, adoption is slower due to capital intensity and integration complexity, with industrial robot deployment still heavily concentrated in a small number of regions. 

So the story is not only about the availability of technology, but also about whether the system around it is ready to carry it. 

Conclusion

Before LLMs, models were already changing work, not by replacing entire enterprises in one dramatic step but by quietly reshaping tasks, narrowing decisions, and redistributing human effort across industries. 

What has changed now is visibility. 

And once we see it clearly, the question is no longer just what models can do, but how enterprises reorganise around them. 

Because the real impact of AI is not only in the model. It is in the structure that forms around it. 

References 

1. USDA Economic Research Service — Increases in Labour Quality Contributed to Growth in U.S. Agricultural Output 

2. USDA Economic Research Service — Farm Labour, Human Capital, and Agricultural Productivity (ERR-302) 

3. USDA Economic Research Service — Precision Agriculture Adoption Data (EIB-248) 

4. FAO — The State of World Fisheries and Aquaculture (SOFIA) 

5. World Bank — Digital Progress and Trends Report 2023 

6. UNDP / FAO — Precision Agriculture & Digital Adoption in Developing Economies 

7. International Federation of Robotics — World Robotics Report