How a wines and spirits house made every vineyard visible

The Challenge

The quality of a great wine or spirit is decided in the vineyard, months before the cellar. Sugar, acidity and aromatic ripeness reach their balance over a narrow window, and a harvest called a few days too early or too late leaves its mark on the whole vintage.

For a leading wines and spirits house, that window has to be read across a supply base it does not farm itself. The house sources from tens of thousands of hectares worked by more than a thousand independent growers, each with its own plots, practices and microclimate. The house carries the quality risk of decisions taken parcel by parcel, far from its own cellars.

A team of a dozen people monitored this network on foot and on paper. With that capacity, only one parcel in ten could be tracked through the season. The other nine were judged by extrapolation, and harvest dates were called on experience: the know-how of seasoned field staff, applied to a sample far too small for the size of the supply base.

The cost showed up every season. Every wrong call cost quality, yield or both, with fruit picked before it reached its potential or left on the vine past its peak. The house needed to see its whole supply base continuously, and to put its scarce field expertise where it would change the outcome.

The Solution

Our team built a harvest monitoring engine on high-frequency satellite imagery, visual and infrared, read at parcel level. The engine ingests imagery at a high revisit rate and reads it plot by plot. Infrared bands capture vine vigour and stress that cannot be seen from the road, and the parcel-level read turns tens of thousands of hectares into a single map the team can act on, refreshed through the season. 

We then trained machine-learning models to track ripeness and predict the right harvest window per parcel. The models follow how each parcel ripens over the season and translate the signal into a predicted window, so every grower's plot gets its own harvest date instead of a regional average.

To support the vineyard team, we put the output into their hands, so visits go where the models flag a risk or a decision. The predictions were built into the team's working routine. Field experts keep the final call and spend their time on the parcels where their judgement changes the outcome.

Finally, we designed the method to extend to the group's other vineyards without starting again. The imagery pipeline, the models and the operating routine were built as reusable components from the start, so the first deployment becomes a capability the wider group can roll out across its other sourcing regions.

The Impact

Coverage. 100% of the vineyard footprint is now monitored, up from 10%. Every parcel supplied by more than a thousand growers is read through the season, and the house no longer extrapolates from a one-in-ten sample.

Timing. Harvest windows are predicted 5 to 8 days earlier. That lead time gives the house and its growers room to organise picking and reception around the fruit, and to call each parcel at its optimum.

Productivity. The field team is eight times more productive. The same dozen experts now cover the whole supply base and spend their time on the parcels that need it, with evidence behind each decision. Their experience, once spread thin across a sample, is now concentrated on the calls where it makes the difference.

Beyond. The AI that moves P&L in an incumbent sits on the process that makes the product. Here, machine learning reads the raw material at source and acts on the quality, yield and timing of every harvest, a lever that a generative assistant on a support function cannot reach. The same approach transfers to any business whose product quality is set upstream, in fields, forests or supplier networks it does not directly control.