Using AI to Prepare Your SFI Application Step by Step

This page covers one job: using a general-purpose AI assistant to get a Sustainable Farming Incentive application ready, from pulling your land data together to checking the final numbers before you click submit in the Rural Payments service. If you want the wider picture first, including Countryside Stewardship, capital grants and how AI fits into Red Tractor prep, start with our overview of AI for SFI, grants and farm assurance and come back here.

Before you start, two things to know. The AI won’t apply for you, and it shouldn’t. What it does well is the slow middle part: matching actions to fields, adding up payments, spotting rotation clashes and turning guidance written by civil servants into a checklist you can use. The application itself goes through your own Rural Payments login.

The second thing is that SFI changes. The offer, the actions on it, the payment rates and any cap per business are set by Defra and can change between application windows. Every rate on this page is there to show how the method works. Check every figure against the current SFI guidance on GOV.UK before you rely on it.

What you need before you open an AI tool

Half the time saved comes from having your inputs ready. Gather these into one folder on your phone or laptop:

• Your land parcel list from the Rural Payments service: parcel IDs (for example SK4521 3307), areas in hectares and land cover. Check it against your maps first, because mistakes in this list flow into everything after it.
• Your cropping plan for the next three years. A spreadsheet export from Gatekeeper, Omnia or Farmplan is fine, and so is a typed list.
• Hedgerow lengths if you plan to use hedgerow actions. Measure them from the Rural Payments map or your own records.
• Any existing agreements: Countryside Stewardship, older SFI agreements or woodland schemes. Overlaps are where applications go wrong.
• The current SFI action guidance, downloaded as PDFs from GOV.UK.

Leave out your SBI number, bank details and Government Gateway credentials. None of them help the AI, and none of them belong in a chat window.

Choosing a tool (and what it costs)

Any of the main assistants will do this job. The difference is mostly how well they handle uploaded documents and spreadsheets.

Free tiers will manage a small holding, but they limit uploads and conversation length. Paying for one month while you do the application costs about the same as an hour of an agent’s time. Cancel afterwards if you don’t need it.

Step 1: Give the AI the rules before you give it your farm

The most common mistake is pasting your fields in and asking “what SFI actions can I do?” The model then answers from whatever it picked up in training, which may be the 2023 offer, the 2024 offer or something it has made up.

Upload the current guidance PDF first and tie the AI to it:

I'm preparing an SFI application for a farm in England. I've attached
the current SFI action guidance from GOV.UK. Use ONLY this document
for eligibility rules and payment rates. If something isn't covered
in the document, say "not in the guidance" rather than guessing.
First, list every action in the document with its code, payment
rate, and the land types it applies to, as a table.

Go through that table carefully. If an action code you expected is missing, or a rate looks wrong next to the PDF, sort it out now. Every later step depends on this table.

Step 2: Match actions to parcels

Now add the land. Here’s a cut-down example from a 240 ha mixed arable holding in Lincolnshire: winter wheat, oilseed rape, spring barley and a block of permanent grass for a small suckler herd.

Here is my land parcel list and 3-year cropping plan (attached CSV).
For each parcel, suggest which SFI actions from the guidance table
could apply, given its land cover and cropping. Flag any parcel
where two suggested actions can't be combined on the same land.
Show results as a table: parcel ID, hectares, crop 2027,
suggested actions, conflicts.

A typical reply looks like this:

| Parcel      | Ha    | Crop 2027      | Suggested actions        | Conflicts                     |
|-------------|-------|----------------|--------------------------|-------------------------------|
| SK4521 3307 | 18.4  | Spring barley  | CSAM2 (winter cover)     | None                          |
| SK4521 5512 | 6.2   | Winter wheat   | AHL1 on 1.5 ha headland  | Check AHL1 % cap per holding  |
| SK4522 0918 | 22.0  | Oilseed rape   | None suggested           | OSR drilled Aug: no cover gap |
| SK4522 4401 | 11.7  | Permanent grass| CSAM1, grassland actions | None                          |
| SK4523 1176 | 4.8   | Legume fallow  | NUM3                     | Can't also enter CSAM2        |

Two things to look at. First, the AI caught that a parcel in legume fallow can’t also take a winter cover crop action, which is exactly the overlap you want caught. Second, it flagged that a percentage cap might apply to AHL1. Go and check that cap yourself in the guidance, and don’t take the AI’s word for the number.

Step 3: Run the payment sums

Agents earn their fee here, and so does the AI. Ask it to cost each scenario, showing its working:

Using the rates from the guidance table, calculate the annual
payment for these actions. Show each line: action, quantity,
rate, total. Then give a grand total.

Using illustrative rates from the 2024 offer, the Lincolnshire farm’s draft came out like this:

Check the arithmetic on a calculator or in a spreadsheet. Language models still make adding-up mistakes, especially across long tables. On this farm the AI’s first total was £8,855, which was right, but a second run with more rows came out £400 short because one line had been left out. Once the rows are agreed, putting the final table into Excel and using SUM is safer than trusting the chat.

Then look at what each payment costs to earn. A winter cover mix at £129/ha might cost £60–£80/ha for seed and drilling, which leaves a smaller margin than the payment suggests. Ask the AI to add a cost column using your own seed quotes, not its guesses.

Step 4: Stress-test the rotation

SFI actions run on fixed timings. An action that pays well in year one can clash with a crop you’ve planned for year two. Give the AI your full three-year plan and push on it:

Look at my 3-year rotation against the actions we've chosen.
For each year, tell me if any action's timing requirements
(e.g. cover in place by a date, no cultivation until a date)
clash with when I'd normally drill or harvest the planned crop.

On our example farm this turned up one real problem: a field down for CSAM2 in 2027 was planned for an early-drilled winter barley in autumn 2028. That would have meant destroying the cover early. The fix was to move CSAM2 to a parcel going into spring beans. It took two minutes to find, against an awkward conversation with the RPA a year later.

Step 5: Build your evidence checklist

Every SFI action comes with record-keeping. The guidance spreads it across many pages. Ask for it in one place:

For each action we've chosen, list exactly what records or
evidence I must keep, how long I must keep them, and any
dates by which something must be done. Format as a checklist
grouped by month of the year.

The output becomes your working calendar: soil sampling done and the plan written within the first year for CSAM1, seed invoices and dated photos of established cover for CSAM2, the IPM plan completed with an agronomist. Print it, stick it on the office wall, or paste it into the task list in your farm software.

Step 6: Check it like an inspector would

Before you submit, turn the AI round and ask it to find the faults:

Act as an RPA compliance inspector. Review my proposed action
list, parcel allocation and rotation. Identify anything that
would be ineligible, double-funded with my existing Countryside
Stewardship agreement (details attached), or hard to evidence.

This often throws up things you missed, such as a parcel already under a Countryside Stewardship option paying for similar work. Double funding is a rejection, and sometimes a clawback.

Step 7: Submit it yourself

Log into the Rural Payments service, pick your actions and allocate them parcel by parcel, using your checked table as the reference. The service validates as you go, so if it rejects a combination the AI said was fine, the service is right. Note the discrepancy and move on.

Allow about an hour for this part on a farm the size of our example. The steps before it took roughly three hours of AI-assisted work in total, compared with the eight to ten hours the same farmer had spent by hand on their previous application.

Where AI gets SFI wrong

It’s worth knowing the common failures so you can catch them:

• Out-of-date rates. Without the PDF uploaded, models confidently quote figures from earlier offers.
• Invented action codes. We’ve seen “CSAM4” and “HRW5” turn up in answers. Neither existed in the guidance we gave it.
• Rounded hectares. Some tools round 18.43 ha to 18.4 or 18, which throws out totals and can stop the area matching your RPA land data.
• Missed caps. Per-holding percentage limits on some actions get forgotten partway through a long conversation. Restate them if the chat runs long.

If your holding is complicated (tenancies, shared grazing, common land or several businesses), use the AI to prepare your figures and questions, then book a session with a land agent or adviser. You’ll pay for an hour of their time rather than a day, because the groundwork is already done.