AI for Crop Planning and Rotations

Most UK farms already plan rotations with AI to some degree. When Gatekeeper flags a field that’s had wheat three years running, or xarvio FIELD MANAGER sends a septoria risk alert, a model is doing part of the work. What has changed in the last two years is that general-purpose assistants (ChatGPT, Claude, Microsoft Copilot) can now read your own field records and do the arithmetic and comparisons that used to take a wet February afternoon with a spreadsheet.

This guide is for people who already know their rotation and want sharper answers from these tools. It covers what each type of tool is good at, what it costs, and a full worked example of a rotation change on a blackgrass farm. It also covers where AI goes wrong, because when it’s wrong about a pesticide approval or a scheme rule, the bill is real.

Three kinds of tool, three different jobs

“AI crop planning UK” covers three quite different categories of product. Mixing them up is the most common reason farmers feel let down.

Farm management software with planning modules. Gatekeeper, Farmplan, Muddy Boots, Agworld and Hutchinsons’ Omnia hold your field history, inputs and yields. Their planning features are mostly rule-based: they check rotational constraints, carry forward nitrogen calculations and produce the records for assurance schemes. Some now include AI features for drafting plans or answering questions. Their strength is that the data is already structured and tied to real fields.

Imagery and agronomy decision tools. This group includes xarvio FIELD MANAGER (BASF), Climate FieldView, John Deere Operations Center, Syngenta’s Cropwise and Omnia’s mapping side. They use satellite imagery (mostly Sentinel-2, which is free and revisits roughly every five days, cloud permitting), weather data and disease models to create zones, variable-rate maps and spray-timing alerts. They help within a season. They do much less for deciding what goes in Long Meadow in 2028.

General AI assistants. ChatGPT, Claude, Copilot and Gemini know nothing about your farm until you tell them. But they’re very good at reasoning across messy information: a CSV export of five years of yields, a soil analysis PDF, your contract prices and an AHDB Recommended List table. This is where most of the new value in rotation planning sits, and also most of the risk.

A sensible setup uses one from each group. Your records live in the farm software, in-season decisions come from imagery, and the “what if” questions go to a general assistant.

What it costs

The prices below are approximate for 2026 and exclude VAT. Software pricing for farms changes often and is frequently negotiated or bundled through an agronomy company, so check with the vendor before budgeting.

For a 400 ha arable farm, a working stack of one assistant subscription, free xarvio and your existing records package adds roughly £240 a year. That’s well under the cost of a single tonne of lost wheat yield per field.

Getting your data into a usable state

An assistant can only reason about what you give it. The most useful thing you can do before trying any of this is export a clean field history. You need one row per field per year, with these columns:

field_name, hectares, harvest_year, crop, variety, drilling_date, yield_t_ha, soil_type, blackgrass_score_0_5, notes
Long Meadow, 18.4, 2022, Winter wheat, KWS Extase, 2021-10-08, 9.1, Hanslope clay, 2, first wheat after beans
Long Meadow, 18.4, 2023, Winter wheat, KWS Dawsum, 2022-10-14, 7.2, Hanslope clay, 3, second wheat, take-all patches
Long Meadow, 18.4, 2024, Winter OSR, Aurelia, 2023-08-22, 2.6, Hanslope clay, 3, CSFB larvae high

Gatekeeper and Farmplan will export most of these columns. The blackgrass score is the one you’ll usually have to add by hand from walking notes or your agronomist’s reports. A simple 0 to 5 scale is fine, and it does a surprising amount of work later on.

Also gather your most recent soil analyses (P, K, Mg and pH index, plus organic matter if you have it) and a realistic price sheet for each crop. Use your own contract or forward prices, not whatever the assistant thinks wheat is worth. Left to guess, most AI models will use a price from their training data, which may be two harvests out of date.

Before you upload, take out anything you wouldn’t want stored on someone else’s server: bank details, SBI numbers, staff names. On ChatGPT, turn off “Improve the model for everyone” in data controls. Claude’s consumer plans have a similar training setting. Team and Enterprise plans don’t train on your data by default.

Worked example: a heavy-land rotation under blackgrass pressure

Take a 420 ha arable farm on Hanslope clay in north Bedfordshire. It’s been running a six-year rotation for a decade:

Current rotation: first wheat, second wheat, winter oilseed rape, first wheat, second wheat, winter beans.

Blackgrass is getting worse, with herbicide resistance confirmed by a Rothamsted resistance test, and oilseed rape has failed twice to cabbage stem flea beetle. The farm manager wants to know whether a rotation with more spring cropping would pay.

They uploaded the field history CSV to Claude and gave it this prompt:

You are helping me compare two arable rotations on heavy clay in
Bedfordshire. Use ONLY the prices and costs I give you. Do not invent
figures. Where you are making an assumption, list it separately.

Prices (ex-farm, harvest 2027 forward): feed wheat £185/t, feed barley
£165/t, spring oats £160/t, winter beans £220/t, OSR £390/t.

Variable costs per ha (seed + fert + sprays) from our Yagro report:
WW1 £645, WW2 £680, OSR £520, winter beans £330, spring barley £400,
spring oats £330.

Yields: use my 5-year field averages from the attached CSV, split by
first/second wheat. Second wheats on fields scoring 3+ for blackgrass
should use the actual figure from those fields, not the farm average.

Rotation A (current): WW1, WW2, OSR, WW1, WW2, winter beans.
Rotation B (proposed): winter beans, WW1, spring oats, WW (after oats),
spring barley.

Give me average gross margin per ha per year for each rotation, then
a sensitivity table showing what happens if second-wheat yield
recovers by 0.5 and 1.0 t/ha.

The assistant came back with this gross margin table (output minus variable costs):

Rotation A averages £735/ha per year: (1,020 + 541 + 650 + 1,020 + 541 + 638) ÷ 6. Rotation B averages £742/ha: (638 + 1,020 + 550 + 946 + 557) ÷ 5.

The difference is £7/ha, or about £2,900 a year across 420 ha. On gross margin alone it’s a draw.

What makes the exercise worthwhile is what the gross margin leaves out, and a good prompt makes the assistant list those things:

• Blackgrass trajectory. AHDB-funded research puts the reduction in blackgrass heads from spring cropping at around 80 to 90% compared with September-drilled winter wheat. Rotation B has two spring crops in five years. Rotation A has none. On this farm’s figures, if blackgrass in Rotation A pushes second-wheat yields down another 0.5 t/ha over three seasons, Rotation A falls to about £720/ha.
• Fixed costs and workload. Dropping OSR removes the August pinch between harvest and rape drilling. Spring crops move cultivation and drilling work into March and April, which may let the farm keep one fewer tractor or delay a drill replacement. On a 420 ha unit, that’s worth more than £7/ha.
• Risk. The OSR row assumes 3.0 t/ha, but this farm’s real five-year average includes two crops written off to flea beetle. Using the actual OSR figure of 2.4 t/ha brings Rotation A down to roughly £696/ha.

The useful output here wasn’t “Rotation B is better”. It was a set of numbers built on the farm’s own data, with every assumption written out where the manager and their agronomist could argue with it. This took about 40 minutes. Doing it by hand in Excel would have taken most of a day, and the sensitivity table probably wouldn’t have been built at all.

For a closer look at dedicated software that runs this kind of comparison across every field automatically, see AI Crop Rotation Planning Software for UK Arable Farms.

Prompts that get better answers

The worked example worked because the prompt was strict. These are the habits that make the biggest difference.

Give it the numbers and forbid it from inventing others. Write “Use ONLY the figures I provide” and “list assumptions separately”. Without that instruction, models will quietly fill gaps with plausible numbers.

Ask for the table before the opinion. If you ask “which rotation is better?”, you get a confident paragraph. If you ask for the gross margin table, then the sensitivity, then a recommendation, you can see where the reasoning goes wrong.

Name the constraint that actually binds. On most farms, one thing limits the plan more than anything else. It might be blackgrass, the drill’s capacity in a wet October, a malting contract that needs 120 ha of Laureate, or the grain store. Say what it is.

Make it check against field rules. For example:

Check every field in this plan against these rules and list breaches:
- No OSR closer than 1 year in 5 on any field (clubroot history on
  Top Barn and Dovecote: 1 in 6).
- No more than two consecutive cereals, counting oats as a break for
  take-all but not for blackgrass.
- Sugar beet no closer than 1 in 4 (BCN-tolerant varieties only on
  fields with BCN results above threshold).
- Spring crops only on fields with blackgrass score 3 or above.

This is the part of the work that farm software already does well, so treat the assistant as a cross-check rather than the record of truth. Where the two disagree, find out why. Usually one of them has a field boundary or a year wrong.

Mixed, dairy and livestock farms

Rotation planning isn’t just an arable problem. On mixed and livestock farms, the questions are about forage, ley length and grazing, and the same approach applies.

Consider a 180-cow dairy herd in Cheshire with 40 ha of maize and 110 ha of grass. The manager wants to bring maize fields back into grass and break up some tired leys. Assuming about 7.3 t DM per cow per year of total intake (20 kg DM a day), the herd needs roughly 1,314 t DM before youngstock are counted. At 16 t DM/ha, the 40 ha of maize supplies about 640 t DM. A good assistant prompt here asks for a five-year sequence that keeps total forage dry matter within 5% of demand every year while rotating maize onto older leys, with the ley yield per year taken from the farm’s own silage clamp measurements.

The plan it comes back with should show where the lean year falls. That’s usually the year a large reseed goes in and produces only a light first cut. Knowing that 18 months in advance means you can sow some Italian ryegrass or wholecrop to cover the gap, rather than buying in silage in March at whatever the market wants.

On sheep and beef units, assistants handle the timing arithmetic for catch crops well. If the winter barley comes off on 20 July and stubble turnips need around 12 weeks to bulk up, you can check whether finishing lambs on that field before the following spring drilling is realistic. Add the actual field area and stocking rate and it will estimate grazing days. Then check the estimate against what the field actually did last year.

Herbal leys deserve their own mention. They can earn scheme payments and fit naturally into mixed rotations as a fertility-building break. Ask the assistant to model a three or four-year herbal ley phase within an arable rotation, using your own figure for the first-wheat yield after the ley. Don’t let it assume the benefit. The yield boost from a ley varies a lot with soil type and how the ley was grazed.

Schemes, rules and the things AI gets wrong

This is where to slow down. General assistants are poor at UK regulation and scheme detail, for three reasons: the rules change often, they differ between England, Wales, Scotland and Northern Ireland, and the models are trained on whatever version was published when they were built.

What goes wrong in practice:

• Scheme actions and payment rates. Ask about SFI actions and you may get 2023 codes, 2024 rates, or actions that closed to new applicants. Take every action code and rate from the current published offer on GOV.UK (or the equivalent devolved scheme) and paste it into the conversation. Don’t let the model supply them.
• Pesticide approvals. Assistants regularly suggest active ingredients that have lost approval, or quote label rates and latest timings that are wrong for the product you’ll actually buy. Check every product against the HSE CRD pesticides register, and take the decision through your BASIS-qualified agronomist. That isn’t optional.
• Nutrient figures. Models often mix up RB209 editions or quote nitrogen recommendations without the soil nitrogen supply index that makes them meaningful. Use the AHDB Nutrient Management Guide (RB209) directly, or the calculator in your farm software, and give the assistant the result.
• Farming Rules for Water and NVZ closed periods. These depend on soil type, manure type and location. An assistant’s general answer isn’t good enough to plan a spreading programme on. Use it to draft the plan, then check it against the rules for your holding.

In short, the assistant can arrange, calculate and compare information you’ve given it. It shouldn’t be the source of any fact that a cross-compliance inspector or a Red Tractor assessor might ask you about.

Three setups by farm type

400 ha combinable crops, owner-operator with part-time office help. Keep Gatekeeper or Farmplan as the record. Use free xarvio for in-season disease and variable-rate nitrogen. Pay for one ChatGPT Plus or Claude Pro seat (about £240 a year) for rotation scenarios, contract comparisons and drafting the cropping plan your agronomist will review. Get a full field-history export done before the end of November so scenarios are ready ahead of seed orders.

1,200 ha contract-farmed estate with a farm manager and an administrator. At this size the money is in modelling every field each year, not in one-off prompts. Look at dedicated rotation software that pulls from your records system (compared in AI Crop Rotation Planning Software for UK Arable Farms). Add Yagro so the cost lines in your scenarios are benchmarked rather than guessed. Two assistant seats, one for the manager and one for the office, cost around £480 a year, and the administrator will probably use theirs more, mostly on scheme paperwork and contract summaries.

Mixed dairy and arable, 250 ha. The rotation question here is mostly forage, so the critical data is clamp and cut-yield records, not a yield map. Microsoft 365 Copilot works well if the herd and forage records already live in Excel, since it can work directly on the existing workbook. Otherwise a standard assistant subscription with CSV exports does the same job for less.

Before the next seed order

The best time to run this is between harvest and the autumn seed order, while the yield figures are fresh and before any variety decisions are fixed. Export the field history. Put together the price sheet with real forward prices. Write the field rules down as a list, so the clubroot rule on Top Barn stops living only in your head. Then run the current rotation and one alternative through an assistant with a strict prompt like the one above. Take the resulting table to your agronomist and get them to argue with every assumption on it, field by field. The fields where you disagree with the assistant usually show you something about the farm that isn’t in the data yet, so add a column for it before next year’s run.