ChatGPT and AI for Farm Admin

Most farm offices run on the same few things: a spray record that needs filling in, a pile of RPA letters, a milk recording report that nobody has time to read properly, and a VAT return due on Friday. ChatGPT and similar tools will not do any of that for you. They can cut the time each job takes by a third to a half, as long as you know what to give them, what to check, and where they are likely to get it wrong.

This guide is for people who have already opened ChatGPT a few times, asked it something, and got a reply that was either surprisingly useful or confidently wrong. The aim is to make the useful replies more frequent and the wrong ones easier to catch.

What it does well, and where it falls over

ChatGPT is a language model. It is very good at reshaping text, so it can turn a rambling voice note into a tidy table, a three-page tenancy clause into five bullet points, or a list of bank transactions into categorised lines. Give it a spreadsheet and ask a specific question, and it will write and run a small piece of Python in the background to answer it, which is much more reliable than asking it to do sums “in its head”.

The weak spots are facts it has to remember rather than read. Ask it for current SFI payment rates, the withdrawal period for a particular wormer, or the latest closed period for slurry spreading on your soil type, and it may give you an answer that sounds authoritative but is out of date or simply made up. The rule that saves most trouble is this: give it the source, then ask the question. Paste in the scheme guidance, upload the product label, attach the export from your herd software. Don’t make it guess.

Here’s a rough split from a season of use on mixed holdings:

Which tool, and what it costs

You do not need to pay for anything to start. The free tiers are capable enough to test whether this is worth your time. Paying gets you better models, file uploads without tight caps, longer conversations and, on business plans, firmer guarantees about your data.

Approximate UK prices as of autumn 2026 (they shift, so check the billing page before you commit):

If you only pay for one, ChatGPT Plus is the sensible default for a farm office. It handles spreadsheets well, its phone app has a good voice mode for dictating records while you’re still in the yard, and the Projects feature lets you keep a folder for each enterprise (arable, dairy, sheep, grants) with its own files and instructions.

A mixed farm with a manager, a herdsman and a part-time secretary is the point where ChatGPT Business starts to make sense. Three seats at roughly £25 each is £75 a month, and you get shared projects and an admin console. It also means staff aren’t putting farm data through personal accounts.

Set it up before you paste anything in

Ten minutes of setup makes every later conversation better.

Data controls first. In ChatGPT go to Settings, then Data Controls, and turn off “Improve the model for everyone” if you don’t want your conversations used for training. On Business plans this is off by default. Either way, avoid pasting employees’ National Insurance numbers, bank details or medical information. Use initials or staff numbers instead. If you have to handle something sensitive, Temporary Chat doesn’t save to your history.

Then tell it who you are. Under Settings, Personalisation, there is a box for custom instructions. Whatever you put there gets read before every conversation. Something like this works:

I run a 420 ha mixed farm in Shropshire: 310 ha arable (winter wheat,
winter barley, oilseed rape, spring beans), 180-cow block-calving
spring dairy herd, 350 Mule ewes. Farm is in an NVZ. We are Red Tractor
assured for crops and dairy, and milk goes to Arla.

Use UK units: hectares, tonnes, litres, £. Use UK regulations
(RPA, APHA, VMD, Environment Agency, HMRC) not US ones.

If you don't know a figure (product rates, MAPP numbers, withdrawal
periods, payment rates), say so and leave a blank. Never estimate
regulatory figures.

Keep answers short. Use tables where data has more than three columns.

The line about never estimating regulatory figures matters most. It won’t stop every invented number, but it makes the model far more likely to write “[check label]” than to guess.

Worked example: a voice note into a spray record

Red Tractor wants a complete application record: field, crop, product, MAPP number, rate, area, date, operator, reason, and conditions. On a busy day those details end up scribbled on the back of a delivery note, if at all.

With the ChatGPT app open in voice mode, the operator says something like this on the way back to the yard:

“Sprayed Top Ground and Long Close today, winter wheat, both T2. Ascra Xpro at 1.2 litres, plus a bit of manganese. 100 litres water. Started about twenty to ten, finished half twelve. Wind southwest, maybe 8 mile an hour, dry. Me, Tom.”

Then, back in the office, paste the transcript into a project called “Arable records” with this prompt:

Turn this into a spray record row for each field. Columns: Date,
Field, Crop, Growth stage/timing, Product, MAPP no., Rate (/ha),
Water volume (l/ha), Area (ha), Total product used, Operator,
Start, Finish, Wind, Reason. Field areas: Top Ground 18.4 ha,
Long Close 12.95 ha. Leave MAPP numbers blank. Flag anything missing.

The output looks like this:

| Date     | Field       | Crop   | Timing | Product      | MAPP | Rate     | Water | Area  | Total   | Operator | Start | Finish | Wind       | Reason          |
|----------|-------------|--------|--------|--------------|------|----------|-------|-------|---------|----------|-------|--------|------------|-----------------|
| 09/05/26 | Top Ground  | W wheat| T2     | Ascra Xpro   |      | 1.2 l    | 100   | 18.40 | 22.08 l | Tom      | 09:40 | 12:30  | SW 8 mph   | Disease control |
| 09/05/26 | Long Close  | W wheat| T2     | Ascra Xpro   |      | 1.2 l    | 100   | 12.95 | 15.54 l | Tom      | 09:40 | 12:30  | SW 8 mph   | Disease control |

Missing or unclear:
- Manganese product name, rate and MAPP/registration not given.
- Single start/finish time covers both fields; record separately if required.
- Date assumed as today; confirm.
- MAPP numbers left blank as instructed. Check the Ascra Xpro label.

Three things are worth noticing. It worked out totals (22.08 l and 15.54 l) that you can check against what went into the sprayer. It flagged the vague “bit of manganese” instead of inventing a rate. And Long Close had been measured in acres in the old records (32 acres), but because the prompt gave it in hectares there was no conversion error. Units are where most mistakes creep in, so always give areas in the unit you want back.

Copy the table into Gatekeeper, Muddy Boots, or whatever spreadsheet you use for records. Five minutes of dictation and tidy-up replaces twenty of writing out.

Worked example: field gross margins from a messy spreadsheet

Upload last season’s cropping spreadsheet, even if half the fields are in acres and the costs are on three separate tabs. Then be specific:

Calculate gross margin per hectare by crop for harvest 2026.
Convert any acre figures to hectares (1 ac = 0.4047 ha).
Output = yield (t/ha) x price (£/t). Variable costs = seed +
fertiliser + sprays only. Show your working as a table and tell me
which rows you had to convert or exclude.

A typical answer:

Check one row by hand every time. For wheat, 9.0 × 185 = 1,665, less 545 = 1,120. If that row is right, the method is right and the others will follow. If it’s wrong, it usually means a column was misread (seed cost per tonne instead of per hectare is a common one), and asking “show me the raw values you used for wheat seed” will expose it in seconds.

Where this pays is in the follow-up questions. “What wheat price would oilseed rape need to match on margin?” or “If fertiliser goes up 15%, what happens to each crop?” take seconds to answer and give you something concrete to take to the agronomist or the bank manager. They are not a cropping decision in themselves. Rotation, blackgrass pressure and drilling capacity still need judgement that no spreadsheet captures.

Worked example: reading milk recording data for herd health

Most dairy farms get a monthly milk recording report from NMR or CIS and look at the bulk SCC figure and not much else. The individual cow data is where the herd health information actually is.

Export the last three recordings as CSV from your herd software (InterHerd+, Uniform-Agri and Herdwatch all export) and upload to ChatGPT:

This is milk recording data for three recordings. SCC is in
'000 cells/ml. List:
1. Chronic high: SCC over 200 at both of the last two recordings.
2. New infections: under 100 at previous recording, over 200 at latest.
3. Count of cows in each group and as % of cows recorded.
Sort by latest SCC, highest first. Tell me how many cows you
excluded for missing data.

The output:

Cows recorded at latest test: 176 (4 excluded: no SCC at one or more tests)

Chronic high (2 consecutive > 200): 14 cows (8.0%)
| Cow  | Lact | DIM | SCC -2 | SCC -1 | SCC latest |
|------|------|-----|--------|--------|------------|
| 1423 | 4    | 212 | 310    | 388    | 412        |
| 0987 | 6    | 245 | 260    | 295    | 371        |
| 1561 | 3    | 198 | 180    | 240    | 330        |
...

New infections (<100 then >200): 9 cows (5.1%)
| Cow  | Lact | DIM | SCC -1 | SCC latest |
|------|------|-----|--------|------------|
| 1702 | 2    | 64  | 45     | 520        |
| 1655 | 2    | 71  | 82     | 290        |
...

Two of the nine new infections are second-lactation cows under 75 days in milk, which suggests looking at the dry period and calving environment. Look at that pattern together with your vet. Neither you nor ChatGPT should be deciding treatment off this table. What it does do is turn a 30-minute squint at a PDF into a two-minute list you can take to the next herd health visit, and it makes the conversation with the vet about the AHDB QuarterPro dry-cow and environmental patterns much sharper.

The same approach works for beef and sheep weights. Upload a Breedr or AgriWebb export of weaning weights and ask for daily liveweight gain by sire, or for a list of lambs below 0.2 kg/day since marking. Always ask for the exclusion count. If 40 lambs out of 600 had no second weight, you need to know that before trusting the averages.

Paperwork: letters, grants, tenancies and tax

This is where most farm offices save the most time, and where the “give it the source” rule matters most.

Official letters. Paste the full text of an RPA letter or an APHA notice and ask: “What does this require me to do, by what date, and what happens if I don’t? Quote the sentences you’re relying on.” Asking for quotes is the trick. If it can’t point to a sentence, it’s padding.

Scheme agreements. If you’re in an SFI agreement or a Countryside Stewardship Higher Tier agreement, upload your agreement document and the relevant scheme guidance from GOV.UK, then ask for a table of each action, its parcel references, what has to be done when, and what evidence you need to keep. Pin that table to the office wall. Do not ask ChatGPT what the current payment rates or eligibility rules are without giving it the published guidance; scheme rules have changed often enough since 2023 that its memory is unreliable.

Tenancies and contracts. Claude and ChatGPT both handle a 40-page Farm Business Tenancy well. Useful prompts are “List every obligation on the tenant with clause numbers” and “What does this say about rent review timing and the notice required?” Use the output to prepare for a conversation with your land agent, not to replace them.

Letters you’d rather not write. A contractor who baled late and charged full rate, a neighbour whose sheep keep getting through, a buyer who’s paying 60 days late. Give it the facts and the tone: “Firm, not rude, I want to keep using them.” Then cut the first and last paragraph, which are nearly always too polite.

Making Tax Digital. MTD for Income Tax came in from 6 April 2026 for sole traders and landlords with qualifying income over £50,000. The threshold drops to £30,000 from April 2027 and £20,000 from April 2028. That means quarterly updates for a lot of family farms. ChatGPT can’t submit anything to HMRC. You still need recognised software such as Xero, QuickBooks, Sage or Farmplan’s Cashbook. What it can do is take a bank export and propose categories:

Categorise these 312 bank lines into: Feed, Fertiliser, Seed,
Sprays, Vet & Med, Contractors, Fuel, Repairs, Rent, Wages,
Milk sales, Livestock sales, Grain sales, Subsidy/scheme,
Other. Mark anything you're unsure about as REVIEW rather than
guessing. Output as CSV with the original line plus a Category column.

Expect 80 to 90 per cent to come back right on the first pass. The REVIEW lines are usually the interesting ones, such as a payment to a merchant that covers both seed and sprays, or a transfer that’s actually a loan repayment. Hand the result to your bookkeeper as a starting point. That’s often an hour saved per quarter on a farm with a few hundred transactions.

Prompts that actually work

The prompts above share a pattern, and it’s worth making it a habit:

1. Context: what the data is and what units it uses.
2. Task: one specific output, with column names if it’s a table.
3. Constraints: what not to guess, what to flag, what to leave blank.
4. Check: ask it to show working, counts or quotes so you can verify.

“Help me with my spray records” gets you a generic essay about integrated pest management. “Turn this transcript into a spray record with these 15 columns and leave MAPP numbers blank” gets you something you can paste straight into Gatekeeper.

We keep a full set of tested prompts covering cropping, livestock, compliance, staff rotas and accounts at ChatGPT Prompts for Farm Admin: A Ready-to-Use Library. Each one shows the prompt, the kind of file to attach, and an example of the output, so you can copy it into your own projects and adjust the farm details.

Where it goes wrong, and the checks that catch it

A few failure modes come up again and again on farms.

Invented regulatory numbers. Asked for the meat withdrawal period of a common sheep wormer, ChatGPT gave a confident figure that was a week shorter than the one on the product’s SPC. That sort of mistake could put residues in the food chain. Withdrawal periods come from the label or the VMD Product Information Database, every time. Spreading closed periods, buffer zones and LERAP categories also come from the label and the Environment Agency guidance, not from the chatbot.

Unit drift. Acres and hectares, kg/ha of product versus kg/ha of nitrogen, litres versus kg for liquid fertiliser. State units in the prompt and check one row by hand.

Silent exclusions. When a file has blanks or odd formatting, the model can quietly drop rows. Always ask how many rows were read and how many were excluded.

American defaults. Without custom instructions it will happily talk about the USDA, bushels per acre and Fahrenheit. The setup step above mostly fixes this.

Out-of-date scheme knowledge. Anything to do with SFI, Countryside Stewardship, the Farming Equipment and Technology Fund or Capital Grants should be answered from documents you’ve uploaded, dated, from GOV.UK.

A useful habit is to finish any important conversation with: “What in your answer are you least sure of?” It’s surprisingly candid, and it often points straight at the figure you should go and check.

Is it worth the money?

Do the sums the way you would for any other input. ChatGPT Plus at about £20 a month is £240 a year. Say it saves two hours a week across spray and fertiliser records, monthly milk recording review, letters and the quarterly bank categorisation. Two hours at £15 an hour for office time is £30 a week, or around £1,560 a year. Even at half that saving it pays for itself three times over.

The bigger return is harder to cost. Records that are complete on the day rather than reconstructed the night before a Red Tractor audit, and the cow with a rising SCC that gets spotted at the second recording instead of the fourth. There’s also the grant evidence you can actually find when the RPA inspector asks for it.

Where it isn’t worth it: if your records are already tight in good farm software and someone enjoys doing them, the saving shrinks. And if nobody on the farm will take the time to check outputs, the risk of a confidently wrong number getting into a medicine book or a claim outweighs the time saved.

Your first fortnight

Pick one job only, ideally the one that’s most often late. For an arable farm that’s probably spray and fertiliser records. On a dairy unit, start with the monthly milk recording review. If you’re a sheep farm, try weight analysis or the flock health plan update.

Set up the custom instructions on day one. Make one Project for that job and upload any reference documents it needs: field list with areas in hectares, product labels, last year’s records. Use it every time that job comes up for two weeks, and keep a note of the two or three times it got something wrong and what the prompt was. By the end you’ll have a prompt you trust and a clear idea of which checks to keep doing by hand. Then you can move on to the second job.