Alphabet’s Gemini Flash helps a Michigan dairy farmer automate daily analytics with local AI agents

Paul Windemuller, a Michigan dairy farmer, says he has been using Gemini 3.6 Flash to automate parts of his farm management that used to take hours each morning. In a Google blog post published July 28, Windemuller describes building a multi-agent AI setup that pulls together otherwise separate streams of farm data, calculates key efficiency indicates, and then turns the results into an actionable daily briefing.

The operation Windemuller runs, Dream Winds Dairy, began in 2014 with 30 leased cows and has grown over 12 years into a highly automated facility milking 260 Holsteins. Like many modern farms, it generates continuous data: sensor collars track each cow, a local weather station records climate conditions, and an online portal logs milk quality and shipments.

Before adopting the AI agents, the blog says Windemuller spent mornings downloading files, merging spreadsheets, and calculating performance. That routine, he says, kept him away from physically caring for the herd during the most time-sensitive part of the day.

To reclaim time, Windemuller built a local AI system that runs in a “directory-based file-interface” rather than relying on APIs or web scraping. In the described workflow, when CSV exports and other documents such as photos, receipts, and invoices are saved to a monitored folder, Gemini’s multimodal capabilities extract figures from both visual and numeric inputs and combine them into a single operational view.

The system is portrayed as a specialized, orchestrated set of agents that replace long prompts with role-specific steps. The first stage converts raw inputs into a cohesive business overview. Another agent focuses on reporting, translating the analysis into a “Farm CEO Briefing” designed to isolate daily margin drivers so the farmer can act quickly.

A central concept in the post is that the agents are optimized around “Daily Static Variable Margin” (SVM), a metric intended to reduce the noise created by market price changes. The blog contrasts SVM with more traditional measures such as income over feed cost, which it says fluctuate as milk and feed prices move. By holding market prices constant, the approach is meant to better separate biological and operational efficiency from external pricing volatility.

The post also links the shift in cost to the specific model chosen. Gemini 3.6 Flash is described as being built for advanced reasoning, tool use, and coding, with a 1 million token context window and a 64,000 token maximum output. Google’s blog further claims benchmark evaluations show Gemini 3.6 Flash can reduce output tokens by roughly 17 percent compared with Gemini 3.5 Flash, and that it does so at a lower cost per output token, aimed at making repeated “agentic loops” more affordable for daily use.

Because the system runs on local file exports, the blog emphasizes that sensitive data stays on the farm rather than being centralized through external data feeds. Windemuller’s goal, according to the post, is broader adoption by independent farmers who may find it difficult to build or pay for the kinds of continuous, reasoning-heavy workflows that large businesses can support.

Still, the blog does not provide a quantified before-and-after estimate of time saved, cost reductions, or changes in farm performance. It also does not specify the exact configuration of the multi-agent architecture beyond the described role-based workflow, nor does it outline how the “Farm CEO Briefing” recommendations are validated against outcomes over longer periods.

Looking ahead, the case highlights a direction for AI deployments in agriculture: instead of trying to digitize everything into a single platform, the post describes an approach that integrates existing file exports and documents, then focuses on a stable operational metric to support faster decisions. For Google, the story serves as a demonstration of Gemini 3.6 Flash in a real-world, small-business-like workflow where daily automation cost is a key constraint, and where “local first” handling may matter to operators.