Planning Rotations Against Blackgrass With AI Help

On heavy clay, blackgrass decides the rotation for you if you let it. Agronomists have known the remedies for fifteen years: spring cropping, later drilling, competitive crops, ploughing now and then. The hard part is deciding which field gets which remedy, in which year, without wrecking the farm’s margin or loading 150 hectares of spring work onto one drill in a wet April.

AI tools are good at exactly that kind of question. They won’t spot blackgrass for you, and they shouldn’t be choosing your herbicides. What they will do is take your field records, your blackgrass scores and your costings, run three-year scenarios for every field in a few minutes, and lay out the trade-offs. This post works through how to do that on a real-shaped heavy-land farm, with the prompts and the numbers.

What you need before you open any AI tool

The output is only as good as the inputs, and with blackgrass that means field-level data you trust. Get these together first:

• A blackgrass score per field. Walk every field in June, when heads stand above the crop, and use the Rothamsted five-point density scale (absent, low, medium, high, very high). If you have patch maps from Hutchinsons Omnia, Patchwork Technology or a drone survey, use those as well.
• Last three years of cropping, drilling dates and yields. Export these from Gatekeeper, Muddy Boots, Agworld or John Deere Operations Center as a CSV. A tidy spreadsheet works just as well.
• Your variable costs by crop. Use your own figures rather than a handbook’s. Your heavy-land wheat spray bill will be well above the regional average if you’re stacking pre-emergence herbicides.
• Realistic prices. Take a forward price for each crop from your merchant or the futures screen. Don’t pick the number you’re hoping for.

A general-purpose assistant is enough for the modelling. ChatGPT Plus and Claude Pro both cost roughly £18–£20 a month and will read a CSV you upload. Copilot in Excel does a decent job if you’d rather keep everything in one spreadsheet. None of these needs to connect to your farm software. You’re uploading a file and asking questions about it.

The worked example: 420 hectares of Hanslope clay

Take a fictional farm in north Buckinghamshire: 420 ha of arable, mostly Hanslope and Denchworth series, a rotation that’s drifted towards winter wheat and oilseed rape, and blackgrass that has been getting worse since the rape went. One 6 m drill, one 4 m power harrow combination, two people at drilling time.

Six fields will show how the method works. Here’s the data that went into the AI:

For the costings, the farmer used £180/t for feed wheat, £170/t for feed barley with £20/t on top for malting, and £240/t for spring beans. Their own records put variable costs at £645/ha for early-drilled wheat with the full grassweed programme, £625/ha for October wheat, £360/ha for spring barley and £330/ha for spring beans.

The prompt that does the work

Most of the effort goes into writing a prompt that states your assumptions explicitly, so the AI doesn’t invent its own. Here’s the one used for this farm:

I farm 420 ha of heavy clay in north Buckinghamshire. I've uploaded a CSV
with field name, hectares, 2026 crop, drilling date, yield and a June
blackgrass score (Rothamsted 5-point scale).

For each field, model four three-year sequences for harvests 2027-2029:
1. Winter wheat drilled mid-September every year
2. Winter wheat drilled after 15 October every year
3. Spring barley, spring barley, then September wheat
4. Spring beans, October wheat, spring barley

Use these assumptions and nothing else:
- Blackgrass seedbank falls about 70% a year where no new seed is returned
- Delayed drilling cuts blackgrass plants by 50% compared with September
- Spring cropping cuts blackgrass plants by 85%
- Wheat yield loss from blackgrass: low 0%, medium 5%, high 12%, very high 25%
- Clean yield potential: Sep wheat 8.6, Oct wheat 8.0, spring barley 5.8,
  spring beans 3.6 t/ha
- Prices and variable costs as in the second tab

Give me a table of gross margin per hectare per year and the three-year
total for each sequence, field by field. Show the blackgrass score you've
assumed each year. Flag any field where spring work on heavy land looks
risky. Do not recommend specific herbicide products.

The final line is there on purpose. Pesticide recommendations in the UK should come from a BASIS-registered adviser, and a chatbot will happily suggest a stack it has no right to suggest. Use the AI for the arithmetic and the trade-offs, and leave product choices to your agronomist.

What came back for Long Clay

This is the output for the worst field, after checking the arithmetic by hand (more on why below):

LONG CLAY - 18 ha - starting score: VERY HIGH

Sequence                       2027    2028    2029    3-yr total
                               £/ha    £/ha    £/ha    £/ha
1. Sep wheat x3                 507     420     380     1,307
   (BG: VH -> VH -> VH)
2. Oct wheat x3                 671     650     640     1,961
   (BG: H -> H -> M/H)
3. SB, SB, Sep wheat            626     640     948     2,214
   (BG: L -> L -> L)
4. Beans, Oct wheat, SB         534     700     640     1,874
   (BG: L -> L/M -> L)

Field total, sequence 3 vs sequence 1: +£16,326 over three years

The first year looks like the case for staying in wheat. Spring barley at £626/ha comes out behind October wheat at £671/ha, which is the comparison most people make at the kitchen table. Then 2029 arrives. After two spring crops the seedbank has collapsed, the field can go back to early-drilled wheat at full yield, and that year’s margin is £948/ha. Over three years the spring barley route beats carrying on as before by £907/ha, which is a bit over £16,000 on an 18 ha field.

If the barley makes malting spec, year one rises to roughly £742/ha. That’s worth knowing, but the decision shouldn’t depend on it.

The whole farm, field by field

The same prompt, run across all six fields, gave this:

Brook Piece is the interesting one. On blackgrass alone the AI preferred spring barley, but it flagged that a field which floods in winter is a poor bet for March drilling. Asked to rerun the field with a one-in-three chance of missing the spring barley window, it moved the answer to spring beans, which tolerate a later April drilling. That kind of risk-weighting is tedious to do by hand and easy to do in a follow-up prompt.

The question to ask next: can you actually drill it?

Splitting the 2027 crops on these six fields gives 71 ha of spring cropping (Long Clay, Brook Piece, Hangings). Across the whole farm the plan put 138 ha into spring crops. The follow-up prompt was:

I have one 6 m drill doing about 30 ha a day on heavy land in good
conditions. Historically I get 6-9 workable days between 1 March and
15 April. Can I drill 138 ha of spring crops in that window? What
happens to the plan in a year with only 4 workable days?

It came back with this: on a normal spring the 138 ha needs five good days, which fits. With only four days, about 18 ha is left over, and the AI suggested putting the field most likely to be late (Brook Piece) into beans so it can wait until mid-April without losing much. That’s a sensible answer, and it came from your own drilling records rather than someone else’s averages.

The autumn side gets the same treatment. Delayed drilling on clay assumes October will be dry enough to travel, and in some years it isn’t. Ask the AI to model a year where half the October wheat doesn’t go in, and to say what happens to the margin if that land falls back to spring barley. If the answer is that you lose very little, delayed drilling is a low-risk bet on your farm.

Where the AI got it wrong, and how to catch it

On the first run the AI quietly applied the very-high yield penalty to spring barley, which made spring cropping look worse than it is. On another field it rounded 22 ha to 20 when totalling the farm. Neither error was obvious until someone added up a column with a calculator.

Three habits catch most of these mistakes:

1. Ask it to show the sum for one field in full. Yield × price, minus each cost line. If that’s right, the rest of the table probably is.
2. Change one assumption and see if the answer moves the right way. Raise the wheat price to £200/t and continuous October wheat should gain ground on spring cropping. If it doesn’t, something is wired up wrong.
3. Keep the assumptions in your own spreadsheet, not just in the chat. That way you can rerun next year with this June’s scores, without relying on the AI to remember what it told you last time.

Treat the blackgrass reduction figures as working estimates, not guarantees. The 50% from delayed drilling and the 85% from spring cropping are in the right range of published trial results, but your own fields may do better or worse. If you’ve scored the same fields for several years, give the AI your own history and ask it to fit the figures to that. You’ll get better numbers than any national average.

For the wider picture (break crop choice, oilseed rape’s place on heavy land, linking the rotation to your fertiliser and labour budgets), the AI for Crop Planning and Rotations guide covers the rest of the cropping plan.

With flufenacet already withdrawn in the EU and resistance to the post-emergence chemistry widespread, the margin in herbicide-only control on very-high fields is shrinking. Use the cultural options deliberately, field by field, and check the numbers before you commit. The work starts in June: walk every field, write down a score while the heads are showing, and send the AI a CSV in July while there’s still time to order spring seed.