Bayer Crop Science has deployed a natural-language layer over its governed analytics platform, allowing breeders and researchers to ask about model runs, results and failures without searching separate APIs, dashboards, documentation and logs.
The Agentic Decision Engine uses a supervisor to route each request among 22 specialist agents. A troubleshooting agent can combine workflow status with telemetry, while other agents handle domain knowledge, operations and visualizations. Bayer says a diagnosis that previously took more than an hour of log review can now take under 10 minutes, and the first phase resolves about 80 percent of user questions without an engineer.
The system keeps durable workflow checkpoints outside the agent runtime, separates employee identity from service and tool authorization, and limits each specialist to relevant, mostly read-only tools. Monitoring records task failures rather than treating every successful HTTP response as a successful answer. Prompt caching reduced input-token cost by 20 to 30 percent in one repeated workload.
More than 100 Bayer Crop Science users have accessed the gated production release. The performance figures come from Bayer and AWS, not an independent audit. The implementation’s broader lesson is that production agents depend less on one prompt than on identity boundaries, constrained tools, recoverable state, task-level observability and evidence tied to the underlying analytics system.