Where AI Can and Can't Help With SFI

Most of the SFI application help on offer right now falls into two camps. One says AI will do the whole thing for you. The other says don’t let a chatbot near anything with the Rural Payments Agency’s name on it. Neither is much use when you’re sitting at the kitchen table at 9pm with a field map, a soil lab report and an application window that closes sooner than you’d like.

There’s a more useful way to look at it. An SFI application is a chain of separate jobs. AI tools like ChatGPT, Microsoft Copilot, Claude and Google Gemini are genuinely good at some of those jobs. For others they’re a liability. Below, each stage gets a verdict: helps, partly helps, or keep it away.

First, the rule that runs through all of it. Payment rates, action codes and eligibility rules come from gov.uk and the Rural Payments service, and nowhere else. SFI has changed shape several times since 2022. The 2024 offer closed to new applications in March 2025, and the scheme that replaced it has a different action list, different caps and different entry windows. A chatbot trained on last year’s internet will quite happily give you last year’s scheme with total confidence. It won’t warn you that it might be out of date.

The stage-by-stage map

Eligibility and land data: keep it away

Ask ChatGPT “am I eligible for SFI?” and you’ll get a tidy, plausible answer. It might mention the minimum area, land management control, or whether you have an existing agreement. Some of that may even be right for the current offer. You have no way of knowing which parts are, short of checking gov.uk anyway, so you may as well start there.

Land data is worse. The tool has no access to your parcels in the Rural Payments service. It doesn’t know a parcel has been reclassified, or that a field you rent on a 364-day grazing licence may not count as land you control for the full agreement. When you ask it about your land, it’s guessing. Do this stage in RPS itself, or in the Land App (which is free for SFI mapping and pulls from RPA data), and fix any wrong land cover with the RPA before you go further.

Choosing actions: partly helps

AI earns its keep here, as long as you give it the action list rather than asking it to remember one.

Here’s a worked example. Say you farm 220 ha of arable near Grantham: winter wheat, oilseed rape, spring barley and a couple of grass margins. Download the current SFI action list from gov.uk, copy the actions you’re interested in into the chat, then add your rotation:

Here is my rotation for 2026-2028 by field, with hectares.
Here is the list of SFI actions I'm considering, copied from gov.uk today.
For each field, tell me which actions could physically fit the
rotation and what management it would commit me to. Do not suggest
any action that isn't in the list I've pasted. Do not give payment rates.

That last pair of instructions matters. They make it work only from what you gave it. What you get back is a working sheet showing, say, that a cover crop action fits ahead of your spring barley but not ahead of the wheat that follows oilseed rape, because you won’t have the window. That’s real thinking time saved. It took one Lincolnshire grower I spoke to about 40 minutes instead of an evening with a highlighter.

The weakness is agronomy at the margins. AI will tell you a herbal ley is a good fit for a heavy clay field it knows nothing about. It doesn’t know your drainage, your blackgrass history or that the neighbour’s sheep are the only reason the ley idea works. Treat the output as a list to argue with, not a decision.

Payments: keep it away from the rates

This is the one that catches people. Ask any of the big chatbots what an SFI action pays and there’s a decent chance you’ll get the 2024 figures. For example, the 2024 offer paid £129/ha for a multi-species winter cover crop (CSAM2), £739/ha for a pollen and nectar flower mix (AHL1), and £6/ha plus £97 per agreement for the soil assessment action (CSAM1). Those numbers were real. They aren’t a basis for budgeting under a different offer with a different action list and an agreement-level cap.

Even with the right rates, chatbots still slip on arithmetic over long tables. They multiply 14.37 ha by a rate and round somewhere they shouldn’t, or drop a row. Put the figures from gov.uk in a spreadsheet and let Excel or Google Sheets do the sums. If you want AI involved, ask it to build the spreadsheet formula, then check one row by hand:

Field     Action   Hectares   Rate (£/ha, from gov.uk)   Annual £
Top Road  [code]   14.37      [paste current rate]       =C2*D2

Here, the AI writes the structure, and the numbers come from you.

Plans and written evidence: this is where it helps

Several SFI actions have required a written plan: soil management, integrated pest management, nutrient management. This is drafting work, and drafting is what large language models are actually good at.

Feed it your soil lab results (pH, P, K, Mg, organic matter by field), your notes on compaction from digging a few test pits, and the plan requirements copied from the action guidance. Ask for a soil management plan in the structure the guidance asks for. A job that used to mean two or three hours of writing becomes 30 minutes of correcting a draft. On ChatGPT Plus or Claude Pro, at roughly £18 to £20 a month, that alone can pay for the subscription in one season. Free tiers will do it too if you’re patient with the limits.

Two cautions. Check whether the action needs the plan produced or reviewed by a qualified adviser (some nutrient management actions have specified a FACTS-qualified adviser), because a chatbot doesn’t count. And read every figure in the draft against your lab report. Models sometimes “tidy” a pH of 6.4 into 6.5 or swap two field names. If your name is going on the plan, you have to have checked it line by line.

Records during the agreement: helps

Once the agreement is live, the boring work is evidence: dated photos, drilling and destruction dates, invoices for seed mixes, notes of when the hedges were cut. Microsoft Copilot on your phone can read a photographed seed invoice and pull out the supplier, date, mix and quantity into a line you paste into your records sheet. Ask Gemini or ChatGPT to turn a voice note recorded in the cab (“drilled cover crop Long Meadow 14th September, 25 kilos a hectare”) into a formatted log entry. Nobody gets excited about this, but it’s the part that saves you at inspection.

Keep the originals, though. The AI-tidied log is your index. The invoice and the timestamped photo are the evidence.

Replying to the RPA: partly helps

When an RPA letter asks you to explain a discrepancy, AI can help you write a clear, polite, well-ordered reply, especially if letters aren’t your strong suit. Give it the RPA’s letter and your bullet-point version of what happened, and ask it to write the reply using only those facts. Then strike out anything it has added. Chatbots like to smooth things over with phrases such as “as per our agreement terms”, which can commit you to something you didn’t mean.

Submitting: keep it away

Never give your Defra ID or Government Gateway login to an AI tool, a browser extension that promises to “fill the form for you”, or an AI agent that offers to drive the Rural Payments site on your behalf. Submission is your declaration. If something goes wrong, the liability sits with you, and you won’t be able to show the RPA what the tool actually entered.

One setting is worth changing before you paste anything about your business into a chatbot. In ChatGPT, go to Settings, then Data Controls, and turn off “Improve the model for everyone”. Claude and Gemini have equivalent controls. Your SBI number and field-level yields don’t need to end up in anyone’s training data.

The same pattern applies to Capital Grants, Countryside Stewardship Higher Tier and assurance paperwork. We’ve mapped those in the wider guide to AI for SFI, grants and farm assurance. The questions are always the same: is the AI working from documents you’ve given it, or from memory? And can you check every number it hands back against gov.uk before it goes anywhere official?