- AI
- Agri-food
Building intelligence around data: lessons from a GPT assistant project in the agricultural sector
July 7 — 2026
Organizations aren't short on AI tools. ChatGPT, Claude, Gemini: language models are widely available, and their capabilities speak for themselves. Yet many companies still struggle to turn their artificial intelligence initiatives into reality. The reason? It's rarely a technology problem.
This is exactly the kind of challenge faced by Farm Management Canada, a national organization that supports Canadian farm operations in managing their businesses. A challenge that revealed where true AI expertise actually lives: not necessarily in mastering the technology itself, but in the ability to design intelligent systems that know how to use data.
An AI project for Canada's agricultural sector
Farm Management Canada and Mirego have been working together for several years, most notably on the development of AgriShield, an agricultural risk management platform launched in 2018. That longstanding relationship made it a natural fit when Farm Management Canada began exploring how artificial intelligence could enhance its offering. The organization had accumulated a substantial library of educational resources on generational transition and agricultural risk management, but that content was scattered across its website, accessible only through traditional document searches. The idea: build conversational assistants capable of answering member questions in an interactive way, drawing on those existing resources while maintaining a precise tone and positioning. Two distinct tools for two specific needs — one assistant to support discussions around succession planning, and another to guide risk assessment.
“We had initially approached Mirego to develop a custom AI tool. Given strict budget constraints and the need to integrate within an ecosystem already familiar to farmers, Mirego's team proposed an alternative that would allow us to bring the idea to life — while sparing us the ongoing costs of maintaining the tool down the road.”
— Mathieu Lipari, Program Manager at Farm Management Canada
The solution: two custom GPTs, deployed on OpenAI's GPT Store. This approach came with several key advantages: no servers to manage, no infrastructure to maintain, rapid deployment, and publicly accessible tools with no barriers to entry.
Two AI tools for farm operations
This assistant supports farm families as they think through succession planning. Its positioning is key: it acts as a guide, not a consultant. The distinction matters. A consultant provides answers; a guide asks questions that help people find their own. This approach encourages users to structure their thinking rather than passively receive information. It also significantly reduces legal exposure by framing the tool as educational rather than advisory. The assistant helps facilitate family conversations, clarify the vision for the transition, and prepare organized materials before meeting with professionals such as lawyers or accountants.
This tool draws on the AgriShield knowledge base to guide users through the assessment and management of agricultural risks. It helps complete risk analysis grids by breaking down complex theoretical and academic concepts into plain language, making methodologies accessible that would otherwise be buried in PDF documents. The tool also points users to relevant resources based on their specific context.
A wealth of documentation to make sense of
Farm Management Canada supplied a substantial body of reference material: transition guides, provincial resources, risk management methodologies, educational content, and information on available programs. The expertise was all there — just locked in static documentation.
The technical challenge surfaced quickly. OpenAI enforces an 8,000-character limit on a GPT's instructions. There was simply no room to fit all the content, guardrails, positioning, and contextual nuances into that space. Prioritization became essential: deciding what the model needed to know at all times (core instructions, tone, ethical and legal boundaries) versus what could live in knowledge files, retrieved on demand based on the questions being asked.
Architecting intelligence around data
An AI model doesn't read information the way a human does. It looks for patterns, semantic relationships, and precise contextual signals.
A common assumption is that you need perfectly clean, well-organized data before launching an AI project. In practice, as long as the data is reliable, there are ways of working that make it possible to build effective intelligent systems. The traditional approach — centered on data transformation and organization — still holds value in certain contexts. But with the rise of powerful language models, a new paradigm has emerged: AI engineering. Rather than transforming data, you layer intelligence on top of what already exists. You design how the model will navigate information, what instructions will shape its behavior, and where to draw the boundaries.
These two approaches are not in opposition. They serve different needs and can even be combined within a single project, depending on data volume, stability, and intended goals.
Two assistants, two complementary approaches
In the Farm Management Canada project, both approaches were combined based on the specific needs of each assistant.
The Farm Transition Navigator relied on a limited set of stable documents, so the approach focused on compressing and restructuring the information while adding intelligence layers to guide the model's behavior.
The Farm Risk Navigator, on the other hand, needed access to a broader knowledge base tied to AgriShield. A Retrieval-Augmented Generation (RAG) architecture was chosen, allowing the model to dynamically retrieve relevant information based on the questions asked.
The expertise work focused on several key areas:
Prompt engineering: Defining the tone, positioning (guide vs. consultant), and ethical and legal guardrails within the system instructions.
Defining clear boundaries: Declining legal advice, acknowledging uncertainty rather than fabricating answers, and directing users to qualified human resources.
Retrieval-Augmented Generation (RAG) architecture: Determining what needed to be immediately accessible versus what could be retrieved on demand. Legal boundaries? In the instructions. The full list of risk categories, subcategories, and best practices (600+ items in total)? In the knowledge files.
Compression and structuring: Working within OpenAI's 8,000-character instruction limit by distilling content to its essence while preserving critical nuances, then chunking, annotating, and organizing documents so the model knows where to look depending on the question asked.
Rigorous user testing: Domain experts, both internal and external to Farm Management Canada, including specialists in risk management and agricultural transition, tested both assistants using predefined questions. These tests drove multiple improvement iterations to validate proper assistant behavior and the accuracy of generated responses.
“This kind of project shows that deep expertise in AI engineering makes all the difference — and it goes beyond tools or technology that anyone can access. It requires a thorough understanding of how language models actually work, knowing when to combine approaches, and mastering intelligent system architecture. That expertise is what makes it possible to achieve results this effective, this fast.”
— Pascal Hamel, VP of Artificial Intelligence at Mirego
Fast deployment and full autonomy
The project was completed in under 50 hours, delivering two fully functional assistants — available publicly in both French and English — with no complex maintenance infrastructure required.
“Mirego's team also supported us by training all of our staff in a one-hour session on how to maintain and manage the GPTs to keep information current. As a result, our team is now fully autonomous in managing this tool, which also requires minimal ongoing effort.”
—Mathieu Lipari, Program Manager at Farm Management Canada
This project speaks to a reality that extends well beyond the agricultural sector: organizations don't need to wait for perfect data to launch effective AI initiatives. As long as the data is reliable, AI development expertise can be used to build intelligent systems that address real needs — quickly and sustainably.
