Sevilla FC lets recruiters question 300,000 scouting reports with a Llama-based Scout Advisor

The problem

Sevilla FC scouts write detailed reports on players, noting traits such as attitude, tenacity and leadership. The club holds more than 300,000 of these reports. Its other data tools handled numbers well but could not use this free text. Meta says that before the tool, working through one shortlist of candidates took recruiters 200 to 300 hours, and its case study describes the manual research as taking days.

What they built

The club's data department worked with IBM to build Scout Advisor. A recruiter types a question about the kind of player wanted. The tool returns matching players with short summaries drawn from the scouting reports. It runs Meta's Llama 3.1 70B Instruct, an open-weight model (its files are published, so a company can run it itself), on IBM's watsonx platform. Answers are in Spanish.

How they did it

Football language is full of slang, so the team added a step that rewrites each question with football context before searching. A request for "talented wings" is expanded into a description of what a good winger does, which stops the system from returning unrelated results. The enriched question drives a similarity search over the reports, and the matching passages are passed to Llama to write the summaries. This pattern is called retrieval-augmented generation: the model answers from the club's own documents rather than from memory. The club's chief data officer says the model was chosen for text enrichment and summarization, especially in Spanish. He also says running an open model inside the club's own environment avoids sending data to third-party services.

Results

Meta reports that evaluation of multiple scouting reports on one player dropped from hours to seconds. The Sporting Director says he no longer has to read dozens of reports and can get what he needs in perhaps two minutes. The club also reports better talent identification and new consulting work with other teams that want to use AI. No accuracy measure, cost or team size is published.