What AI Crop Planning Tools Actually Do With Your Field Data

Your farm software has just ranked second wheat as the most profitable option for Top Field next harvest. It says so with a green tick and a margin figure to the pound. Your agronomist wants spring barley. Before you decide who to believe, it’s worth knowing what’s going on behind that number. Most of the “AI” in crop planning is less clever than the marketing suggests, and more useful than the sceptics admit, as long as you know which bits to lean on.

This piece looks inside the black box. It covers what goes in, what comes out, and where the model’s answer is a starting point rather than the final word.

The four inputs, and how good each one really is

Every crop planning tool on the UK market works from roughly the same four ingredients. Gatekeeper, Omnia from Hutchinsons, xarvio FIELD MANAGER, Climate FieldView, Agworld and the benchmarking side of Yagro all package them differently. What separates a sound recommendation from a confident guess is the quality of each input.

Soil is where many tools quietly cut corners. If you haven’t uploaded your own sampling, the software will often fall back on Soilscapes. That map is fine for a county-level picture and not much use on a farm where one field runs from sandy loam at the top to heavy clay by the brook. Ask your tool which soil source a recommendation used. If it can’t tell you, assume the coarse one.

History is the input you control completely, and it carries more weight than most people realise. Rotation logic (take-all risk in second wheats, the clubroot interval for oilseed rape, the blackgrass seedbank) depends on the tool knowing what was in that field for the last four or five years. A missing year gets treated as “unknown”. Some tools handle that well, while others quietly assume the best case.

Prices are the input the model can’t actually know. Any margin comparison for harvest 2027 is built on a price you or a futures feed supplied. The software has no special insight into November 2027 wheat. Whatever certainty it shows is the certainty of the number you put in.

Weather data drives disease risk models (xarvio’s septoria and yellow rust alerts are the best-known example) and drilling and spraying windows. It does that job well over the next five to seven days. Beyond that it is climatology dressed up as a forecast.

A worked example: Top Field, 22 hectares

Here’s a realistic case for a heavy-land arable farm in the East Midlands. The field is clay loam with P and K at index 2 and pH 7.4. It grew oilseed rape in 2024 and first wheat in 2025, which yielded 8.9 t/ha. The blackgrass burden is moderate and has been getting worse.

A typical planning tool sets out three options like this:

The tool ranks second wheat first, at £25/ha ahead of spring barley. Over 22 ha that’s £550. It looks decisive, and it isn’t.

The 7.8 t/ha second wheat figure is a regional average penalty applied to your first wheat yield. The arithmetic is correct. The assumption underneath it is the weak point. Run the same calculation with a slightly worse take-all year:

Second wheat, yield sensitivity (price £180/t, variable costs £560/ha)

Yield t/ha   Output £/ha   Gross margin £/ha   vs spring barley (£819)
   7.8          1,404             844                 +25
   7.5          1,350             790                 -29
   7.2          1,296             736                 -83
   7.0          1,260             700                -119

A drop of 0.3 t/ha flips the ranking. On a field where the previous first wheat was drilled late into a wet seedbed, 0.3 t/ha is well within the normal noise. The tool’s “best option” is really a statement that the options are close, and the model can’t separate them.

None of that accounts for blackgrass. Spring barley gives you an autumn of stale seedbeds and glyphosate passes, and a drilling date that dodges the main blackgrass germination flush. If that saves you a pre-emergence stack of £80 to £100/ha in the following wheat, and keeps the field out of the cycle that ends in whole-field spraying off, the “£25/ha loss” turns into a clear win across the rotation. Most planning tools score one year at a time. Your agronomist scores five.

Outputs you can trust

Some of what these tools produce is solid, and you should use it without second-guessing:

• Margin arithmetic. Once the inputs are right, the sums are right. A side-by-side gross margin table across 30 fields, rebuilt in seconds when you change the wheat price, is a real time saver over a spreadsheet.
• Rotation rule checks. Flags for a short OSR interval, a third cereal, or a field that breaks an SFI or Environmental Land Management commitment are reliable because they are rules, not predictions.
• Nitrogen planning within RB209. Tools that follow AHDB’s Nutrient Management Guide (RB209) will give you a defensible starting figure, typically around 220 kg N/ha for a feed wheat on medium soils at a moderate SNS index. It’s an auditable figure you can show an inspector.
• Short-range spray and disease windows. Five-day risk alerts using local weather are good at telling you when to walk a field.
• Satellite crop variation maps. Sentinel-2 imagery at 10 m resolution reliably shows you where a field differs. It cannot tell you why, and in a cloudy British spring you might go three weeks without a clear pass.

Outputs to treat as a hypothesis

These outputs look just as polished on screen and deserve much less faith:

1. Field-level yield forecasts before drilling. The error band on a single field’s yield a year out is easily ±1 t/ha. That band is wider than the gap between most of the options you’re choosing between.
2. “Optimal rotation” recommendations. These are usually margin rankings repeated year by year, with simple penalties bolted on. They don’t model your blackgrass seedbank, your drainage, or the fact that your contractor can’t drill 60 ha of spring barley in a wet March.
3. Price-driven switching. If a tool pushes you towards a crop because its price ticked up last week, it’s doing what you could do with a futures screen. It just looks more scientific.
4. Disease risk scores beyond a week. Anything longer is a seasonal average.

Where your agronomist overrides the model

The override isn’t a matter of gut feel against data. Your agronomist holds information the model has no input for:

• the blackgrass head count from last July’s field walk, and whether it’s resistant
• clubroot, verticillium or cabbage stem flea beetle history in that particular block
• which fields really drain and which only drain on paper
• grain store, drier and labour capacity at harvest
• what the merchant will take at what spec, including malting premiums that never reach a national price feed

When the software and your agronomist disagree, ask one question: which input would have to change for the model to agree with the agronomist? In the Top Field case the answer is “a blackgrass cost carried into next year and a yield 0.3 t/ha lower”. That makes the disagreement concrete. Usually it settles it too.

The data question nobody reads

Everything above depends on you feeding these tools years of yield maps, input records and soil results. Read the data terms before you connect a combine or a Gatekeeper account to a new platform. Check three things: whether your data is shared in aggregate with the provider’s parent company (several are owned by input manufacturers), whether you can export everything in a usable format if you leave, and whether the data is deleted when you cancel. A tool that holds ten years of your field history is very hard to walk away from, so that negotiation is worth having on day one.

For how the different platforms compare on rotation modelling specifically, our guide to AI for crop planning and rotations walks through each of them field by field.

Next time the green tick appears on Top Field, change the yield by half a tonne and watch whether the ranking survives. If it doesn’t, the tool hasn’t made a decision for you, and you’re paying your agronomist for the judgement it can’t make.