AI-assisted decision support for agricultural research and public systems

The problem

Public-sector decision-making in agriculture is constrained by fragmented data ecosystems and the growing volume of scientific literature. Translating complex research into timely, actionable insights for policymakers, extension services, and advisory systems remains a persistent bottleneck.

The solution

This project, developed by Employment and Social Development Canada (ESDC), explored the use of Artificial Intelligence (AI)-enabled workflows to support rapid research synthesis and decision support in agricultural contexts.

The approach focused on systematically analysing large volumes of scientific literature using AI tools, converting unstructured research into structured, decision-relevant outputs, bridging the gap between domain complexity and practical usability, and embedding domain expertise in plant sciences to validate and refine outputs.

Rather than full automation, the system was designed as an augmentation layer — meaning AI handled the initial processing and structuring of information, while human experts retained responsibility for interpreting results and making decisions.

This approach was chosen to address the challenge of synthesising large volumes of complex agricultural research in a timely and scalable way. Traditional methods are resource-intensive and difficult to scale in public-sector contexts. AI-assisted workflows enable faster analysis while maintaining flexibility. A human-in-the-loop design — where human experts review and validate AI outputs at each stage — was intentionally used to ensure accuracy, transparency, and alignment with domain expertise.

The impact

The project reduced the time required for literature synthesis by an estimated 40–60% and improved the accessibility of complex scientific insights for non-specialist stakeholders. It also demonstrated the practical feasibility of AI-assisted decision support in an agricultural context.

Lessons learned

• AI-assisted workflows can significantly speed up research synthesis — with the right safeguards. AI-assisted workflows significantly improved the speed and efficiency of synthesising large volumes of agricultural research. Structured prompting — where instructions to the AI are carefully designed to produce consistent, targeted outputs — and iterative refinement produced consistently high-quality summaries, while domain expertise ensured outputs remained accurate and contextually relevant. The low-cost, modular setup also made the approach highly adaptable to different problem areas.
• Quality depends on prompt design and validation. Initial outputs were inconsistent without carefully designed prompts and validation steps. AI-generated summaries occasionally lacked nuance or introduced inaccuracies, particularly when dealing with highly specialised or context-dependent research. Additionally, translating outputs into formats directly usable by policymakers required more effort than anticipated.
• Design for the end user from the start. Future iterations would prioritise the development of standardised prompt frameworks and validation protocols earlier in the process. Integrating structured datasets alongside unstructured literature would improve reliability. Greater focus would also be placed on designing outputs specifically for end users — such as policymakers or advisory systems — rather than adapting research outputs retrospectively.

While this project was conducted as an independent, early-stage exploration, it was designed with public-sector regulatory contexts in mind.

Key considerations included data protection and privacy (ensuring no sensitive or personal data was used), transparency of AI-assisted outputs, and the need for human-in-the-loop validation — where human experts review and approve outputs before they are acted on — to mitigate risks such as bias and hallucination. Hallucination refers to instances where an AI system generates information that sounds plausible but is factually incorrect.

The approach is also aligned with emerging public-sector AI governance principles, including accountability, explainability, and responsible use of AI in decision-support contexts. Particular attention was given to ensuring that outputs could be clearly interpreted and validated by domain experts before any application in policy or advisory settings.

Although formal procurement or regulatory approval processes were not directly applicable at this stage, the workflow was intentionally designed to be compatible with typical public-sector requirements for auditability, reliability, and ethical AI use.