What AI Actually Is (and Isn't) on a UK Farm

Most of what gets sold to farmers as “AI” falls into one of three groups. Some of it isn’t AI at all. Some of it is, but isn’t worth the money on a 300-acre mixed holding. A small part of it can save you real hours this winter.

Talk about AI in farming in the UK tends to lump everything together: robot weeders, chatbots, satellite maps, collars and “smart” weigh crates. That makes it hard to tell what to spend money on. A better approach is to sort every tool into one of three groups, then decide which farm jobs each group can actually do.

The three kinds of farm AI

That’s the whole map. If someone pitches you a product and you can’t put it in one of those rows, ask them which one it belongs to. If they can’t answer, it’s usually a normal piece of software with an AI label on it. An EID weigh crate that flags lambs under target weight is a good example. It’s useful, but it’s a spreadsheet rule, not AI.

Chat assistants: the office worker who never went to agricultural college

ChatGPT and similar tools are trained on huge amounts of text. They’re very good at reading, rewriting and structuring words. They’re poor at facts they can’t check, and they don’t know your farm unless you tell them.

On a farm, that means the paperwork.

Worked example: a Red Tractor procedure in ten minutes. Say your next audit is due and you need a written procedure for cleaning down the grain store. A farm manager on a 400 ha arable unit in Lincolnshire could type something like this:

I farm 400 ha of combinable crops in Lincolnshire. I need a written
grain store cleaning and pest control procedure for a Red Tractor
Combinable Crops audit. Our store is a 2,000 t flat store with
on-floor drying. We use a contractor for rodent control who visits
every 6 weeks. Write it as a one-page procedure with numbered steps,
a record sheet, and gaps marked [CHECK] where I need to fill in
details I haven't given you.

You’d get back a usable draft. It won’t be perfect. The [CHECK] instruction is what makes it safe to use, because it makes the tool show what it’s guessing instead of covering the gaps with confident made-up details. Checking and fixing the draft takes about ten minutes. Writing it from scratch takes about an hour.

Chat assistants do well at these jobs:

• Turning a voice note dictated in the cab into a tidy spray record or field diary entry
• Summarising a 40-page scheme guidance document into the six points that affect you
• Drafting letters to the land agent, the bank or a neighbour about a boundary hedge
• Rewriting a job advert for a herdsperson so it doesn’t read like 1994
• Explaining an unfamiliar clause in a tenancy or contract in plain English

What they can’t do: tell you the current SFI payment rate, the exact cut-off for an RPA claim, or whether a product is still approved for use on your crop. Scheme rules change and the tool may be months out of date. It will still answer, and the answer will sound sure of itself. Treat every number or date it gives you as unchecked until you’ve confirmed it on gov.uk, the CRD pesticides register or your agronomist’s notes.

There’s also an arithmetic trap. Ask a chat assistant to work out whether your 150-cow dairy herd stays under the 170 kg N/ha whole-farm organic manure limit for an NVZ, and it’ll lay out the method well. It can also slip a decimal point in the sum. Use it to set up the calculation, then put the numbers into your own spreadsheet or the AHDB tools.

If admin is where the hours go on your farm (and for most family farms it is), the ChatGPT and AI for farm admin guide covers prompts, data privacy and which paid plan is worth it.

Image and sensor analysis: eyes that don’t get tired

This group is the closest to what people picture when they hear “AI”. A camera or sensor collects data all the time, and a trained model picks out the pattern a person would spot if they had time to stand and watch all day.

Dairy lameness is the clearest case. CattleEye uses a standard 2D camera mounted over the parlour exit. It mobility-scores every cow at every milking and flags new cases before they’re obvious in the yard. AHDB has put the cost of a lameness case at roughly £330 once you add lost milk, treatment, fertility knock-on and culling. Say a 250-cow herd sits at 30% lameness, which is close to the UK average in several surveys. Catching even 20 cases a year earlier is worth several thousand pounds. That doesn’t make it a sure buy. It means you can do the sum against the quote.

Heat detection collars (Afimilk, Allflex SCR, Nedap and others) are a mix of sensor analysis and prediction. They track activity and rumination, then flag cows likely to be bulling. On a block-calving herd, a few extra cows in calf in the first three weeks of the breeding season often pays for the system within two or three seasons. Ask the dealer for references from herds the same size and system as yours. A 600-cow housed herd is a different case from 180 cows at grass.

For arable, the cheap end is very cheap. Plantix is a free phone app: photograph a diseased leaf and it suggests what the problem is. It’s useful as a second opinion, not a diagnosis. It was built mainly on crops grown in India and other developing countries, so it’s weaker on UK problems like yellow rust compared with brown rust in winter wheat. Skippy Scout sits further up. It flies a DJI drone you may already own on an automatic pattern over a field, takes close-up photos, and reports crop growth stage, weed patches and gaps. It’s a reasonable buy if you’re walking several hundred hectares of crops every week in spring.

What image AI isn’t good at: unusual situations. A lameness camera trained on Holsteins may struggle with a herd of Jersey crosses. A weed detection model trained in Germany may not know blackgrass at the two-leaf stage in a Cambridgeshire seedbed. Always ask the supplier where the model was trained and on what.

Prediction models: the AI you’re probably already using

This group is the least glamorous. It’s also the one most UK farmers already rely on without calling it AI.

The Rothamsted light leaf spot and phoma forecasts for oilseed rape use weather data and disease records from past seasons to estimate risk by region each autumn. BlightWatch, funded by AHDB, sends free alerts when your postcode hits a Hutton Criteria period: two consecutive days with a minimum temperature of at least 10°C and at least six hours each day of relative humidity at 90% or above. The SCOPS Nematodirus forecast uses soil temperature data from weather stations across the UK to predict when the hatch will peak, so sheep farmers can time drenching instead of guessing.

None of these will win a tech award, but they’re free, tested against UK conditions, and they change real spray and drench decisions.

The paid end includes yield forecasting, milk yield prediction and benchmarking tools that compare your cost of production with similar farms. Yagro and Farmbench (AHDB) both do some of this. How good they are depends on your own records. A model fed three years of patchy data will give you a confident-looking figure with very little behind it.

Before trusting any prediction tool, ask:

1. What data does it use? (Mine, regional, national?)
2. How far back does that data go?
3. Has it been tested on UK farms, and can I see the results?
4. When it's wrong, which way is it usually wrong?

Question four is the one sales reps rarely have a ready answer for. A blight model that cries wolf too often costs you sprays. One that misses periods costs you the crop.

Where to start this month

On a beef and sheep unit with no budget for kit, start with the free versions. Sign up for BlightWatch if you grow potatoes, check the SCOPS forecast every week from March, and use a free chat assistant for one paperwork job you’ve been putting off.

For a dairy herd over 150 cows, the strongest case for spending money is in sensors: lameness detection or heat collars, chosen by working the per-case sums above against a written quote.

Arable units over 300 ha should look at drone scouting before anything else, because it changes how you spend your own time in April and May. Paying for a robot weeder can wait until the price drops and there are more UK reference farms.

Try a chat assistant on the cheapest job first. Next time a scheme letter arrives, paste it in and ask it to list every date and action required, in order, with the page number each one comes from. Then check the page numbers. Doing that once shows you what the tool gets right and where it makes things up.