Harnessing AI and the farm feedback loops it needs
Farming has been running feedback loops for centuries. Here is why that matters, and how you can capture that value moving forward.
Robert Bakewell's New Leicester sheep, developed through systematic selection in the 1760s and 1770s, reached twice the carcass weight of the old Leicester breed within a few generations of recorded, deliberate breeding. Bakewell had no sensors, no spreadsheet, and no deep understanding of genetics. What he had was a structured record of which sire was selected, why, and what his offspring produced. The same principles apply to train modern AI.
UK National Milk Records (NMR) - the perfect modern example
UK NMR herds, since 2010, have shown how careful record keeping that lifetime milk per cow per day has increased by 25% from 10.5kg to 12.7kg in 2023, whilst simultaneously decreases , increasing milk fat and protein and decreasing mastisis incidence and somatic cell counts (SCC) [1]. NMR does this through their data management system Herd Companion, they record monthly milk yields (volumes, protein, fat, cell counts) updated automatically from diary management software, combine that with regular health and disease testing regimes, and make human entry easy and meaningful for vet and farmer observations. This allows the farmer to step back, look at their herd management and select breeding that maximising their herd characteristics. The next generation are born and managed, allowing feedback into the system of choices made and incremental gains that stack over time.
Bringing profit to pasture
Teagasc, the Irish farming body, have been improving pasture land through systematic reseeding and evaluation, their 2023 data shows their best varieties make an astonishing 253 euros/ha more than the national average. They monitor seasonal growth performance of the sward through regular farmer measurements, DM production, digestibility, silage production and persistency to systematically increase the swards production levels. [2]
What the loop requires
Three things must be present: what decision was made, why it was made, and what followed. Looking back at our previous examples, these would be:
National Milk Records:-
Action: Select the most productive animals for breeding
Reason: Recorded monthly milk yields, health checks and farmer insights.
Outcome: The next generation is produced and the cycle begins again.
Teagasc Pasture Profit Program:-
Action: Sward varietal selection and management choices (Herd movements, grazing patterns)
Reason: Spring, Summer, Autumn and Winter production, DM yield, digestibility, silage production and persistency measured.
Outcome: Varietal and grazing patterns improvement, with sward reseeded for the cycle to be repeated.
Both examples can be recordable in a notebook, a modern spreadsheet, or management software but if you want to make this understandable by a machine, our AI needs us to ‘label’ the data
This blog piece accompanies Moats Ai Readiness Quiz, designed to help you understand if your farm is AI ready, and what you could improve to capture the value of AI. If you want to try it out for yourself, click here.
How to keep records that a machine can learn from
A yield figure, recorded without the inputs and management decisions that produced it, is called ‘unlabelled’ by machine learning researchers. To let a machine learn from our loops and decisions, we need to keep records of yields, and the major management practices that caused it, think fertiliser sprays and their timings, pesticide sprays and soil/foliar tests, and what we did with those results the next season. If just the yield outcome is recorded, but the cause or what drives that yield is absent, a model can’t distinguish a good decision from a good season!
Data vs Insight
A field diary that records growth stage at each visit, notes why a spray decision was made, and logs yield at harvest is generating structured, labelled data. A farm running five sensors producing continuous observations, but with no record of why management decisions were made or what outcomes followed, is producing unlabelled data at high volume.
Many farms already hold the raw material: spray records, nutrient plans, agronomist reports, and grain buyer yield figures are all records of decisions and outcomes. The gap is frequently in their structure, whether the decision and its reason are linked, at field level or sub-field zones, to the result it was intended to produce. Uploading and connecting existing records is often the highest-value first step, and it requires no new equipment.
Where to start?
The Moat AI Readiness Quiz asks where your farm's data practice currently sits across five areas: yield and crop performance, crop protection, nutrition and soil health, field history, and operations and work records. It returns a result by category, with a practical guide to where the highest-value improvements are. You can complete it in around ten minutes here. The question worth asking beforehand is the same one Bakewell was answering at Dishley Grange in the 1770s: can you look back at what you decided, understand why, and see what followed?
Sources
[1] https://www.nmr.co.uk/news/nmr-500-herd-kpi-2023-report-released
[2] https://teagasc.ie/wp-content/uploads/2025/05/Updates-to-the-pasture-profit-index-for-2023.pdf