Can AI help advise everyday agronomy decisions?

The problem

The UK produces a great deal of high-quality, independent agricultural research. The AHDB archive alone runs to thousands of documents, and there is more again from NIAB, ADAS, Rothamsted, HSE, breeders and the agronomy firms. Very little of it is easy to use on farm. Finding an answer means searching several websites, opening long PDFs and reading complex tables, so farmers often fall back on habit or product-led information instead. Decision support tools have been tried and mostly see low uptake.

What we are testing

An AI assistant on WhatsApp, an app farmers already use, that answers questions about varieties, nitrogen, fungicides, soils and rotations using only a defined body of trusted UK sources, and cites the exact source for every answer. It is used alongside normal agronomy advice, not instead of it. The farmer always decides.

Underneath sits the Cosecha knowledge layer: structured variety and product data, keyword and meaning-based search in one database, and a record of the rights and source for every item. Several AI models (Anthropic, Google and OpenAI) are run over the same knowledge so we can see how much answer quality depends on the model and how much on what we give it.

How the knowledge layer is built

The asset is not the chatbot. It is an organised, rights-cleared store of UK farming knowledge that any AI model can search. Five rules govern it:

• Rights firstNothing is stored without its licence and provenance recorded, and rights are filtered before results are ranked.
• Never overwriteOriginals are kept. A changed document becomes a new version.
• Always citeEvery answer links to the exact source section it came from.
• Model-agnosticThe layer sits under any AI model through a single search interface.
• Learn from gapsA weak-retrieval report shows what farmers asked that we answered badly, and decides what to add next.

Sources so far include 24 editions of the AHDB Recommended Lists (2004/05 to 2026/27), the HSE pesticide product register including expired products, Defra and Rothamsted open data, and transcripts from several hundred UK arable videos and podcasts. More is added as permissions are confirmed.