Days to minutes: machine learning that changed the decision

We build models to change decisions, not to sit in a report. Every product starts with the person making the decision and the moment they make it, and works backwards to the data. The measure of success is not model accuracy alone; it's whether the decision got faster, better, or both.

A market commercial team expected to spend 60+ days manually selecting the range for a new store format. The range selection product, applying advanced analytics to identify the popularity, portability and profitability of a range fit for purpose, did it in minutes, using models that ran in a tenth of the time of traditional methods. Rolled out across new store formats, city centre stores and redesigns.

Countries typically spent 14+ days per forecasting submission. A driver-based rolling forecast, applying machine learning to internal and external datasets, generated it in minutes, replaced manual goal setting, and fed the same forecast into the real-time sales dashboard and market sales report. Live in the first market, then rolled out.

Safety and security teams spent hours browsing ledgers for abnormal transactions. Automated evaluation of fraud anomalies across co-worker discount, voiding behaviour and refunds surfaced the outliers in minutes, with exception-based reporting to direct attention where it was needed.

Machine learning applied to internal and external datasets to evaluate new sites with confidence, assessing potential impact on existing and planned channels. The site selection model went on to drive value inside the wider new business metrics product. Related asset optimisation work delivered up to 15% capital expenditure savings.

For the energy business: modelling the transition from combustion to electric vehicle infrastructure saved 2,000 hours a year of manual analysis, and EV planning built on the single customer view delivered a 50% reduction in load cost.

Most questions never become a model. In our delivery model, the Explore stage targets 24 hours to assess and answer ad hoc questions, and around 80% stop there. The models above are the minority that earned further investment, and because they were built as modular components on shared foundations, each could be reused inside the next.