Bayer brings agronomic expertise to the edge with Azure AI Foundry

Agriculture is full of nuance. What works in one region, crop, or season can be completely wrong in another. A weedkiller that’s effective in Nebraska might be restricted in Indiana. A fungicide that works wonders on soybeans in May could damage tomatoes in August. And when it comes to making those decisions, speed and accuracy aren’t just nice to have—they’re critical. For Bayer, this imperative meant rethinking how to scale agronomic expertise, especially at the frontlines of decision making.

“Crop protection isn’t just about answering a question—it’s about answering it precisely, in the right context, under the right conditions,” says Balathasan “Giri” Giritharan, Principal Data Science Architect at Bayer. “That’s why general-purpose models weren’t cutting it.” So, Bayer built something better. And more importantly—something safer. With help from Microsoft and its Cloud for Industry engineering team, Bayer developed E.L.Y. Crop Protection (Mini): a fine-tuned small language model built on the Microsoft Phi framework, trained using Bayer’s proprietary product label data and regulatory rules. The model is hosted and iterated through Azure AI Foundry, surfaced to internal advisors via E.L.Y. Copilot, and made accessible in a safeguarded environment to strategic retail partners.

We sat down with Bayer’s Giritharan and Sachi Desai (VP, AI Go-to-Market) to understand how the solution works—and why it matters for IT decision-makers seeking speed, safety, and domain-specific intelligence at scale.

A model trained to speak “Bayer”

Q: What makes this solution different from other models? 

Sachi Desai, VP, AI Go-to-Market at Bayer: E.L.Y. Crop Protection (Mini) is a specialized AI that understands Bayer’s regulatory and product context. It was fine-tuned using our own data and domain-specific Q&A, so it delivers precise, label-accurate answers that advisors can trust in the field. And that’s critical when you’re dealing with pesticides and herbicides. A wrong answer isn’t just a missed opportunity—it could cause real harm to a farming operation. This model removes the guesswork. It also surfaces through the existing E.L.Y. Copilot tool, so advisors don’t have to learn something new.

Problem first, not product 

Q: What prompted you to build this in the first place? 

Giritharan Balathasan, Principal Data Science Architect at Bayer: Our frontline teams were spending hours—or days—digging through crop protection labels that can exceed 100 pages. If they had questions, they’d escalate to the tech team, which introduced delays and risk. We needed something faster, smarter, and safer. 

Sachi: What we had wasn’t scalable. It created friction. We didn’t want to build tech for tech’s sake—we wanted something that made our advisors more effective and protected growers in the process. 

Why Microsoft, why Azure AI Foundry 

Q: Why not use a commercial GenAI model out of the box? 

Sachi: We explored general-purpose models, but they didn’t consistently deliver the precision we needed for regulated agricultural use cases. We required a solution grounded in our proprietary data and domain-specific context.

Giri: With Microsoft’s help, we fine-tuned Microsoft Phi using our internal datasets. We worked closely with Microsoft’s Cloud for Industry team to shape the model’s architecture and ensure it aligned with our compliance and operational needs. The Cloud for Industry team helped us get the model hosted, deployed, and published in Azure AI Foundry. Azure AI Foundry also gives us IP ownership and lets us control updates, access, and traceability. 

Q: What would have happened if you didn’t use Azure AI Foundry? 

Sachi: We would’ve had to invest heavily in standing up infrastructure. Azure AI Foundry gave us full IP ownership and let us manage the entire model lifecycle, from compliance logging to updates and access control. It also allowed us to go straight to value by publishing the model in the Azure AI Foundry Models catalog, so strategic partners like agricultural resellers can use it—or even build on it—through trusted, secure channels.

Delivering measurable results 

Q: What kind of impact are you seeing so far? 

Giri: We estimate 5-10% productivity gains among early users. Complex questions that once took days or weeks to resolve now take under 10 seconds. That includes surfacing, reviewing, and interpreting documents. 

Sachi: Bayer’s E.L.Y. Crop Protection (Mini) is like giving a new advisor 20 years of agronomic experience in their back pocket. That kind of speed and confidence is a game changer, and our partners love it. The solution reduces dependence on technical support while staying fully aligned to compliance.  

Under the hood: Inside Bayer’s purpose-built AI stack 

Model foundation 
Bayer’s E.L.Y. Crop Protection (Mini) is a domain-specific small language model based on Phi architecture. Unlike general-purpose language models, it’s fine-tuned exclusively on Bayer’s proprietary product label data, regulatory rules, and expert-authored Q&A. 

Secure, enterprise-grade hosting 
The model is hosted in Azure AI Foundry Models, allowing Bayer to retain full IP control while enabling rapid iteration, versioning, and traceability. Every response is source-linked, audit-logged, and compliant with Bayer’s internal standards. 

Field-ready delivery 
The solution is deployed via Bayer’s E.L.Y. Copilot interface—originally used for internal digital agronomy—now extended to strategic retail partners through highly secure APIs and access controls. External users get fast, contextual answers without direct access to sensitive documents. 

Compliance by design
Every layer is built with regulatory alignment in mind. Label updates, product changes, and policy shifts can be rapidly incorporated without retraining from scratch. All data sources are versioned, authenticated, and documented to support internal and external audits. 

What’s next 

Bayer is already seeing measurable productivity gains and positive feedback from pilot users of the model, and plans are underway to expand deployment to additional retail partners across the United States. 

The solution architecture aligns with Microsoft’s top-tier solution accelerator for regulated AI deployments, ensuring scalability and compliance.

Planned future model enhancements include: 

• Expanded domain coverage through fine-tuning for seed selection, biologicals, and carbon products. 

• Multilingual support to support advisors across LATAM and Europe. 

• Integration into partner platforms, allowing field users to access the model within existing agronomic tools. 

By continuing to scale its trusted AI foundation, Bayer aims to set a new industry benchmark—not just for agricultural AI, but for responsible, high-impact innovation in regulated environments. 

Learn more about Azure AI Foundry and Microsoft Phi architecture to explore how to build domain-specific AI models: Microsoft introduces new adapted AI models for industry - The Official Microsoft Blog

This work builds on Bayer’s earlier exploration of GenAI through Azure AI Foundry and OpenAI. Read more about that foundation: Bayer and EY help farmers use AI to unearth critical data with Azure OpenAI Service | Microsoft Customer Stories

Discover more about Bayer and its Crop Science division.