Preparing for a Red Tractor Audit With AI

Most farmers don’t lose sleep over the assessment itself. What gets them is the fortnight before it, going through ring binders and laptop folders to check that the sprayer test certificate really is in date and that every bottle of Metacam in the fridge is in the medicine book. AI can’t do that fortnight for you. It can make it much shorter, though, and this post shows how.

The method has two jobs. First, you get an AI tool to turn the Red Tractor standards that apply to your farm into a checklist built around your farm. Second, you get it to compare that checklist with the records you already keep and list what looks missing. Before any of that, keep one rule in mind for the whole process.

The AI cannot tell you that you are compliant. Only the current Red Tractor standards document and your assessor on the day can do that. The AI reads, sorts and flags. Every item it produces has to be checked against the PDF before you rely on it.

Why a generic red tractor audit checklist isn’t enough

Red Tractor already publishes its standards and self-assessment checklists for each scheme (Dairy, Beef & Lamb, Combinable Crops & Sugar Beet, Fresh Produce and so on) on its website. They’re free and they’re the authority. The trouble is that they’re written for every farm in the scheme. A mixed holding with 180 dairy cows and 120 hectares of winter wheat and oilseed rape ends up working through two or three long documents. A good share of the lines don’t apply, and the ones that do are scattered across different sections.

A red tractor audit checklist that’s actually useful has three columns that the official documents can’t fill in for you: who on this farm is responsible, where the evidence lives, and when it was last checked. You can build that table by hand. An AI tool will draft it in about ten minutes.

What you need before you start

Settings matter here. In ChatGPT, open Settings, then Data Controls, and switch off “Improve the model for everyone” before you upload anything that includes staff names or your holding number. Claude and Gemini have equivalent controls, so check them. Better still, delete names and CPH numbers from your exports first. The AI doesn’t need either of them to find gaps.

Download the PDF fresh rather than letting the AI “remember” the standards. These tools learned from older material on the internet, and Red Tractor revises its standards every few years. If you ask about Red Tractor without uploading the document, you may get a confident answer based on a superseded version. Upload the current file and tell the tool to use nothing else.

Step one: build the farm-specific checklist

Here’s the prompt one of our readers used. They run the mixed farm described above, in Shropshire, with three staff and a self-employed relief milker:

I have attached the current Red Tractor Dairy standards and the
Combinable Crops & Sugar Beet standards. Use ONLY these documents.
Do not use any other knowledge of Red Tractor.

My farm: 180 Holstein cows, all-year-round calving, 140 youngstock
reared on farm. 120 ha arable: 70 ha winter wheat, 50 ha oilseed
rape. On-farm grain store (flat floor, 900 t). Own sprayer, 24 m
trailed, bought 2019. One diesel tank, bunded. Staff: me, herdsman,
tractor driver, relief milker.

Build a checklist table with these columns:
Standard ref | What it requires (short) | Applies to us? | Who is
responsible | What evidence the assessor would expect | Where we
keep it

Quote the standard reference number exactly as it appears in the
document. If you are unsure whether something applies, mark it
"CHECK" rather than guessing. Leave "Where we keep it" blank.

The instruction to quote reference numbers is the important bit. A reference number lets you find each line in the PDF within seconds. When the AI can’t give one, treat the line as suspect.

Claude Pro produced 94 rows across both schemes. An extract looked like this (we’ve replaced the reference numbers with placeholders so nobody copies them; yours will come from your document):

The farmer then spent about 40 minutes checking each row against the two PDFs. Two rows were wrong. One stated a requirement more strictly than the standard actually words it. The other merged two separate standards into one line, so a training record requirement for the relief milker got lost. Two errors in 94 rows is a decent rate. It also shows why you can’t skip the check: the merged row was exactly the sort of thing an assessor would pick up on.

Step two: find the gaps in your existing records

This is where the time saving really shows. Export what you have and give it to the same conversation, so the checklist is still in front of the AI.

Attached are:
1. Medicine records export from Herdwatch, 1 Oct 2025 to 30 Sep 2026
2. Spray records export from Gatekeeper, harvest 2026
3. A list of the certificates and documents in our farm office folder,
   with dates

Compare these against the checklist you built. For each checklist row,
tell me: evidence found / evidence missing / evidence out of date /
cannot tell from these files. Then list every individual record that
looks incomplete, e.g. a medicine entry with no withdrawal date or no
batch number. Do not tell me whether we are compliant. Only report
what is present or absent in the files.

The last line of that prompt makes it harder for the tool to start reassuring you. Ask “am I compliant?” and some tools will say yes in a cheerful tone. You don’t want that.

On this farm the medicine export had 612 lines. The gap report found:

MEDICINE RECORDS (612 entries checked)
- 9 entries with no batch number (all dry cow tubes, Feb-Mar 2026)
- 3 entries with no withdrawal period end date
- 1 entry where administered by = blank (14 Jul 2026, foot spray)
- Medicine purchase records: not in files provided, cannot tell

SPRAY RECORDS (41 applications)
- 2 applications with no wind speed/direction noted
- 1 application to field "Top Barn" with no product rate recorded

DOCUMENTS FOLDER
- NSTS sprayer certificate: dated Aug 2025, check validity period
  in the standard
- Herd health plan: last signed Sept 2025, review due
- Relief milker: no training record found
- Pre-harvest grain store check: no 2026 entry found

All of it was useful, and none of it was a verdict. The farmer looked at each item. The nine missing batch numbers came from one week when the herdsman was off sick and the relief milker didn’t know the system. The batch numbers were still on the empty tube boxes in the bin store, so they could be recorded properly. The grain store check had been done but written on the back of a delivery note. Twelve gaps were fixed in an afternoon, before any assessor came near the yard.

What it actually costs and saves

The farmer kept rough timings. Preparing the year before, without AI, took about a day and a half spread over two weeks, mostly spent reading the standards and hunting through folders. This year:

• Building and checking the checklist took 1 hour 40 minutes
• Exporting records and running the gap check took 50 minutes
• Fixing the gaps took around 3 hours, which had to be done whatever method was used

That’s about 2.5 hours of preparation against roughly 10, from one month of a £18 subscription that they also use for other things. Next year should be quicker again, because the checklist only needs updating, not rebuilding. One caution: if Red Tractor publishes a revised version, start from scratch with the new PDF rather than patching the old table.

The assessment came back with one minor non-conformity, a COSHH assessment for a new teat dip that nobody had thought to look for. Neither the AI nor the farmer found it, because the dip arrived after the document list was exported. Records you don’t give the tool can’t be checked.

Where it goes wrong

Three failure modes come up again and again.

Answering from memory. If a response mentions a requirement and you can’t find it in the PDF, it probably came from the tool’s general training data or an older version of the standards. Ask “quote the exact wording from the attached document”. If it can’t, delete the row.

Overconfident summaries. A line like “your medicine records appear fully compliant” means nothing. The tool has no idea what the assessor will ask to see, and it can’t spot a missing physical item like an unlabelled container or a broken bund. Keep your prompts asking for present-or-absent findings only.

Large files. A full year’s milk recording export plus three scheme PDFs can be more than a free-tier account will handle in one go. You might find some rows quietly skipped. Split the job by scheme, and when the report comes back, check the row count it says it processed (612 in, 612 checked) against the count in your export.

If you’re also using AI for SFI agreements, Countryside Stewardship claims or other assurance schemes, our guide to AI for SFI, Grants and Farm Assurance covers how to keep one set of records that serves all of them, so you aren’t exporting the same spray data three different ways.

Before the assessor arrives

Print the final checklist with the “Where we keep it” column filled in by hand. Put it at the front of the folder. On the day, the assessor works from the Red Tractor standards, not your table, so have the official document open too. If they ask about a standard, take them to the evidence. Don’t read them anything the AI wrote.

The checklist gets you ready. Only the assessment decides whether you pass, and any non-conformity they raise has to be closed within the time Red Tractor allows, which is usually 28 days, using evidence that you collected and checked yourself.