Bayer Fine-Tunes Microsoft Phi on Crop Protection Labels to Build E.L.Y. Crop Protection (Mini) in Azure AI Foundry

Bayer's agronomic advisors have to give precise crop protection guidance that depends on region, crop and season: a product that works in one state can be restricted in another. Frontline teams were spending hours or days searching crop protection labels that can exceed 100 pages, and escalated questions they could not answer to a technical team, which added delay and risk. Bayer tested general-purpose models but found they did not consistently deliver the precision needed for regulated agricultural use. With Microsoft's Cloud for Industry engineering team, Bayer built E.L.Y. Crop Protection (Mini), a small language model based on the Microsoft Phi architecture and fine-tuned only on Bayer's proprietary product label data, regulatory rules and expert-authored question and answer pairs. Microsoft's team helped shape the model architecture and get it hosted, deployed and published in Azure AI Foundry Models, which gives Bayer IP ownership and control over updates, access, versioning and traceability. Every response is source-linked and audit-logged, all data sources are versioned and documented for audits, and label updates and product changes can be incorporated without retraining from scratch. The model is delivered through E.L.Y. Copilot, Bayer's existing internal digital agronomy assistant, so advisors do not need a new tool. It is also offered to strategic retail partners through secure APIs and access controls, so external users get answers without direct access to sensitive documents. Principal data science architect Balathasan Giritharan said: "Crop protection isn't just about answering a question, it's about answering it precisely, in the right context, under the right conditions. That's why general-purpose models weren't cutting it." Giritharan said Bayer estimates 5 to 10% productivity gains among early users, and that complex questions which once took days or weeks to resolve now take under 30 seconds. Bayer reports positive feedback from pilot users and plans to expand deployment to more retail partners in the United States. Planned enhancements include fine-tuning for seed selection, biologicals and carbon products, multilingual support for advisors in Latin America and Europe, and integration into partner agronomic platforms.