Premier League: an AI fan companion built on Microsoft Foundry
Premier League · Media and entertainment · United Kingdom (global audience)
The challenge
The league holds 30 seasons of match data, about 300,000 editorial articles and 9,000 videos, spread across many systems. To answer fan questions it needed this content in one queryable store that refreshes during live matches and can be personalized, while staying within GDPR (no personal data processed without consent) and with guardrails on every AI response.1
What was implemented
In July 2025 the league and Microsoft announced a five-year partnership that includes moving core infrastructure to Microsoft Azure. The Premier League Companion, presented under the Copilot brand, runs in the league's app and website.
Microsoft describes a multi-agent system built on Microsoft Foundry using Azure OpenAI in Foundry Models. Azure AI Search provides agentic retrieval over the content, Semantic Kernel coordinates the agents and connects them to partner systems, and Azure Managed Redis caches answers so that live interactions stay under 500 milliseconds. Data arrives from more than 10 APIs through Azure Data Factory and Azure Databricks, with Azure Cosmos DB and Azure Machine Learning also in the stack. Partners include WSC Sports and Adobe.
Multilingual answers, vertical video and fantasy football features were described as planned or in development. Dynamics 365 Finance and Operations, Microsoft 365 and Power Platform are part of the plan for internal operations.12
Reported outcomes
up approximately 20%
Consumption of the Premier League app and website, year over year, early in the first season after launch (Microsoft narrative; the period is not defined and the rise is not attributed to the Companion alone)12
Measured60 million users
Users across website and app so far this season, according to the league's Chief Commercial Officer (cumulative count, not Companion users)1
Measuredmore than 8,000 transactions per second
Throughput the platform was designed for (design capacity, not a measured load)1
Projected, not measured
Each outcome links to its source; the label there says whether the company, Microsoft or a third party reported it.
What others can learn, and the limits
In our view, the reusable part for publishers and broadcasters is the architecture: grounded retrieval over an owned archive, a single index that refreshes during live events, and caching for traffic peaks. The engagement figure is not good evidence for the assistant itself, because it covers the whole app and website in a season when a new app also launched. Smaller publishers can apply the same pattern with a single search index and a managed model rather than a multi-agent design.
Sources
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.