Soquimich's fertilizer-recommendation system turned agronomy advice into sales

For a fertilizer producer, telling each farmer what blend to apply is the core service: farmers expect fast, reliable, low-price recommendations, and they take product quality for granted. But Soquimich S.A., the Chilean fertilizer producer, found the advice process complicated and slow — its experts described the relationships among fertilizers, soils, weather and previous or target yields only vaguely.

Soquimich initiated development of a system that would determine optimal fertilizer blends automatically. The team built a hybrid: expert-system technology to capture the specialists' knowledge, mixed-integer linear-programming models to pick the blend that meets the agronomic constraints at least cost.

The first, highly sophisticated version failed — not on the math, but on the humans. Farmers could not always provide the inputs the system demanded, and users did not know how to operate and maintain it. The team stripped the design down to a simplified version that real users could feed, and that one succeeded.

The simplified system increased fertilizer sales and was being extended to other regions of Chile and to other countries. The instructive part is the shape of the arc: the modeling was never the bottleneck; matching the tool's inputs and operation to its weakest user was.

The first version was too clever; the simple one sold

The dumb alternative keeps agronomy advice as a bespoke expert service: accurate but slow, unscalable, and it dies with the experts' availability.

Splitting the problem fits the knowledge: expert systems hold the qualitative agronomy, MIP handles the cost-minimizing blend — neither alone does both.

Version one failed because input demands exceeded what farmers knew or could measure; a model is only as good as its worst-feasible input.

Simplification made maintenance possible in the field, which is what let the system spread to new regions instead of dying in pilot.

When quality is taken for granted, service differentiates — a recommendation engine is sales infrastructure. A model nobody can feed is worth nothing; build version one for the weakest user.

The system paper appeared in Interfaces in 2003. By then the final simplified version was running in Soquimich's business, credited with increasing fertilizer sales, and being extended to other Chilean regions and other countries. Soquimich (SQM) went on to become one of the world's largest fertilizer and lithium producers.