Get Your Field Records Straight Before Any AI Touches Them

Every month another farm software company announces an AI assistant. Gatekeeper, Omnia, Fieldmargin and the John Deere Operations Center are all adding chat boxes, rotation planners or “smart” recommendations, and ChatGPT, Copilot and Claude will happily build you a five-year rotation in half a minute if you paste in a list of fields. It’s easy to spend a winter comparing them.

Don’t. The tool matters far less than what you feed it. An AI planner working from clean digital field records will beat a cleverer tool working from a muddle every time, because none of these tools can tell that “Top Ground”, “Top Field” and “TF2” are the same 14.2 hectares. If your records are a spreadsheet from 2019, a notebook in the pickup and whatever the agronomist emailed last spring, sort that out first. The rest of this piece explains why, and gives you a weekend plan to do it.

Why messy records produce confident nonsense

AI tools don’t flag gaps the way an experienced agronomist would. If a year of cropping is missing, a large language model won’t stop and ask. It fills the gap with whatever looks most likely, and it tells you the result in the same calm, tidy sentences it uses when it’s right.

Here’s a real-world kind of failure. A 300-hectare arable farm in Lincolnshire keeps cropping history in three places: Gatekeeper from 2023 onwards, an Excel sheet before that, and the RPA land parcel map for SFI. In 2023 a field was renamed after a hedge came out and two parcels were merged. Here’s what the farm manager pasted into ChatGPT:

Field,2022,2023,2024,2025,2026
Long Meadow,W Wheat,OSR,W Wheat,W Barley,Spring Beans
Church Piece,OSR,W Wheat,W Wheat,Spring Barley,OSR
Hall 12,,,W Wheat,W Barley,
Hall Field,W Wheat,OSR,,,
Bottom Hall,Spring Barley,W Wheat,,,

“Hall 12” is the merged field. It was “Hall Field” and “Bottom Hall” until 2023. The AI treated it as a field with two years of history and no oilseed rape, then put OSR on it for harvest 2027. Part of that field had OSR in 2023, so that’s four years, which looks fine at first. The other half, though, sits next to Church Piece, where clubroot was confirmed in 2025, and that note lived only in the agronomist’s email. The AI never saw it.

The model didn’t fail there. It answered exactly what it was asked. The records were the problem.

The three things that actually matter

Forget yield maps, satellite NDVI and soil scan layers for now. They’re useful, but they sit on top of a foundation most farms don’t have yet. That foundation has three parts.

One name per field, tied to a permanent ID. The RPA land parcel ID (sheet reference plus parcel number, like SK2847 6153) doesn’t change when you rename a field, so use it as the key. Keep the common name your team uses alongside it, but the ID is what joins your records across software, SFI applications and anything you paste into an AI.

Boundaries that match the ground. If your mapping software says Long Meadow is 11.8 ha and the RPA says 11.31 ha eligible, every per-hectare figure an AI works out (seed rates, nitrogen, margins) carries that 4% error. On a field growing 9 t/ha wheat at £180/t, a 0.5 ha error is about £810 of output put on the wrong field, every year.

Five full years of cropping history. Five years lines up with what you already have to keep: NVZ records must be held for five years, and most rotation questions (OSR spacing, take-all risk in second wheats, blackgrass build-up after a run of winter cereals) need at least that much history to answer properly. Crop, variety, drilling date and yield per field per year is enough. Anything more is a bonus.

Get those three right and any decent tool becomes useful. Get them wrong and the best tool on the market turns out polished guesswork.

The weekend plan

This assumes a mixed or arable holding of 20 to 60 fields. A smaller farm will finish by Saturday lunchtime. A big estate should split it across two weekends or give Sunday’s job to the farm secretary.

Friday evening (1 hour): gather everything

Pull every source into one folder on your laptop. That means:

• your land parcel list and maps from the Rural Payments service
• exports from whatever farm software you use (Gatekeeper, Farmplan, Muddy Boots, Fieldmargin all export to CSV or Excel)
• old spreadsheets, scanned spray records, the agronomist’s recommendations
• combine yield summaries if you have them, even if they’re only the totals per field off the cab display

Don’t clean anything yet. Just get it all in one place.

Saturday morning (3 hours): build the master field register

Start a new spreadsheet. One row per field, and these columns:

The “Former names” column does more work than any other. Every time you find a field called something different in an old record, write it down here. Merged and split fields get a line in “Known issues” with the year it happened, for example “Merged from Hall Field + Bottom Hall, autumn 2023”.

Saturday afternoon (3 hours): fill in five harvest years

Add columns for harvest 2022 to 2026. For each year, record crop, variety and yield in t/ha. Work through one year at a time across all fields rather than one field across all years. You’ll spot gaps faster that way, and you’re reading one source document at a time.

Where a field was merged, split the history honestly. If half of Hall 12 was OSR in 2023 and half was winter wheat, write “OSR (6.1 ha) / W Wheat (5.2 ha)”. Don’t average it or pick the bigger half.

Expect gaps. On most farms it’s 2022, the year before someone bought proper software. Ring the agronomist, check seed invoices, look at the old spray records. A seed invoice for 180 kg of Aurelia delivered in August 2021 tells you something went into the ground for harvest 2022. If you really can’t find a year, write “UNKNOWN” in capitals. An AI can work around a gap it’s been told about. It can’t work around one that’s hidden.

Sunday morning (2 hours): check the boundaries

Open your farm mapping software next to the RPA map. For every field where the areas differ by more than 2%, find out why. Usually it’s one of three things: a hedge or track that the RPA has taken out of the eligible area, an old boundary from before a merge, or someone drawing the field by hand in 2018 and never fixing it.

Correct your software to match the RPA eligible area unless you know the RPA is wrong, in which case raise it with them (that’s worth doing anyway for SFI). Fieldmargin’s free tier is good enough to redraw boundaries if your main software makes it painful.

Sunday afternoon (1 hour): test it

Now try an AI tool, and only now. Save the register as a CSV and paste it into ChatGPT, Claude or Copilot with a prompt like this:

Here is my field register for a 310 ha arable farm in Lincolnshire on
mostly clay loam. Using the cropping history, flag any field where:
- oilseed rape would be grown closer than one year in four in 2027
- a third consecutive cereal would occur
- the known issues column conflicts with a likely 2027 crop
Do not assume any crop for cells marked UNKNOWN. List those separately.

Then check its answers against what you know. If it flags something you disagree with, the reason is usually in the data: a missing alias, a wrong year, a known issue you never wrote down. Fix the register, not the prompt.

What it costs

The weekend is roughly 10 hours of your time. A general AI subscription runs about £18 to £20 a month for ChatGPT Plus or Claude Pro, and the free tiers will handle a 40-row register. Fieldmargin’s basic mapping is free. If you already pay for Gatekeeper or Farmplan you’ve got the export tools.

Compare that with one mistake the register would have caught. Putting OSR on a clubroot-infected field can cost you most of the crop. At 3.5 t/ha and £400/t, that’s £1,400 a hectare, and on a 12-hectare field it’s close to £17,000 gone because a note sat in someone’s inbox.

Keeping it clean

A register you build once and leave alone will be out of date by Christmas. Two habits keep it current. Update the crop columns the week you finish drilling, while it’s fresh. And every time a field changes shape, whether a merge, a split, a new margin or a woodland strip for SFI, add the change to “Known issues” with the date before you touch anything in the software.

Once the register is solid, you can ask the bigger questions: break crops, spring cropping for blackgrass, how SFI options fit around the rotation. Our guide to AI for crop planning and rotations picks up from here, and it assumes you have exactly this register ready to paste in.

The tool you end up using will probably change twice before 2030. A field register keyed to RPA parcel IDs, with five honest years of history, will work with every one of them.