AI for Agronomists: Crop Walk Write-Ups and Grower Advice Letters

Dictate your findings in the field, in the same order each walk, then give the transcript to an AI assistant with your report template and ask for a structured write-up and a draft advice letter. Check every observation against your photos, write or confirm each recommendation yourself, and sign the letter before it goes to the grower.

The aim is to take the evening paperwork out of your day, not to change who gives the advice. An independent agronomist sells judgement and accountability, and clients pay for the name at the bottom of the letter. AI is useful for the typing, tidying and consistency: the same headings every time, clear wording for the grower, photo captions that match the notes. It is not useful, and not safe, as a source of recommendations.

Follow me on Instagram@sagnikteaches

What the routine looks like in one week

An illustrative independent agronomist covers about 40 farms, with roughly 12 crop walks a week. Each walk takes an hour or two in the field, then about 40 minutes of writing in the evening, so around eight hours of weekly paperwork. The routine below aims to bring that to about 17 minutes a walk, most of it review.

Connect on LinkedInSagnik Bhattacharya

These times are illustrative, and the first few walks will run longer while you tune the template. The point is the order: your judgement comes before the letter is drafted, never after.

Subscribe on YouTube@sagnikteaches

Step 1: dictate in a fixed order

AI turns a rambling voice note into a good report only if the note has a shape. Choose an order and stick to it for every field. One that works: identification, stage, what you see, what you measured, what you suspect, what you want to check, what you recommend and by when.

Start each note by saying the field name, crop, variety if known and growth stage. Then speak the observations as you would to a colleague, including the measurements: tiller counts, plant populations, the percentage of leaves affected and where in the field you looked. Mention the weather and soil conditions that day. Say "recommendation" aloud before each recommendation so it is easy to find in the transcript. Here is an illustrative note, transcribed:

Field North Barn, winter wheat, variety [variety], growth stage 31 to 32. Walked a W across the field, ten stops. Septoria on leaf 4 and 5, about 15 percent leaf area affected on leaf 5 at the stops by the hedge, less elsewhere. Yellow rust, none seen. Some pale patches near the gate, looks like compaction, soil is wet there. Tiller count about 650 per square metre at stop six. Weeds, cleavers in the south headland, light. Recommendation: apply the first fungicide at stage 32 to 33 once leaf 3 has fully emerged, product and rate to follow after I check the label and resistance notes. Check the gate patch again in two weeks.

Note what is in the note: numbers, locations, a hedge-side pattern, an open item ("product and rate to follow"). The last point is important: it tells the AI not to fill the gap.

Where dictation goes wrong

Voice-to-text mishears crop and chemical names and drops decimal points. A realistic mistake: "6.5 tonnes a hectare" typed as "65 tonnes" because of a pause. Tell the assistant to leave any figure that looks implausible marked with a question mark, and read every figure in the final report against your memory of the field. Dictate the product name by spelling it the first time, or leave it out of the note entirely and add it by hand.

Step 2: turn the transcript into the field write-up

Give the assistant your template and your rules together. The rules matter more than the template, because they control what the model is allowed to do with gaps.

Below is a voice-note transcript from a crop walk. Write the field report using these headings only: Field and crop; Growth stage; Observations; Measurements; Possible causes (as I stated them); Recommendations; Items to recheck.

Rules:
- Use only what is in the transcript. Do not add diagnoses, products, rates, dates or advice that I did not say.
- Keep my wording for recommendations, tidying grammar only.
- If a figure looks implausible or ambiguous, keep it and add [CHECK].
- If something I said is incomplete (for example "product and rate to follow"), write [TO COMPLETE] there.
- List any photo reference numbers I mention.

Transcript:
{{paste}}

A realistic output (illustrative): "Observations: Septoria on leaves 4 and 5; around 15 per cent leaf area affected on leaf 5 at stops near the hedge, lower elsewhere. No yellow rust seen. Pale patches near the gate, likely compaction in wet soil. Light cleavers on the south headland. Recommendations: first fungicide at growth stage 32 to 33 once leaf 3 is fully emerged [TO COMPLETE: product and rate]. Items to recheck: gate patch in two weeks." Everything traces to the note, and the gap stays visible.

What you would fix in that output: the model wrote "likely compaction". You said "looks like compaction", which is softer. In professional advice the difference matters, so restore your wording or change it to what you believe. Mark every hedge in the transcript to guard against the model hardening an opinion into a fact.

Step 3: photo captions that describe, not diagnose

Photos are your evidence file. Name each one with the field and a number as you take it, and say the number into the voice note ("photo 4, leaf 5 at stop three"). Then ask the assistant to write captions from your words, with the image as a check.

For each numbered photo, write a one-line caption using my description in the transcript. Add what is visible in the photo only if it is plain description (colour, location on the plant, stage). Do not name a disease, deficiency or cause unless I named it in the transcript. If the photo and my note disagree, tell me instead of choosing.

A caption like "Photo 4, leaf 5, stop 3: lesions with dark specks, lower third of leaf, as noted for Septoria" is helpful. If the model writes "likely early nitrogen deficiency" about a photo and you never said that, strike it out. Vision models mix up many look-alike problems, such as nutrient deficiency, disease, frost and herbicide injury, so they are caption helpers, not diagnosticians.

Keep each original photo with its GPS location and date, unedited, in the client's folder. The caption is a convenience; the original is your evidence if advice is questioned later.

Step 4: the recommendations are yours

This is the step where nothing is delegated. A recommendation needs the right product for the target, a rate and timing within the product label, awareness of resistance management and the sequence of earlier applications, buffer and neighbouring crop issues, and the grower's constraints. Those come from your qualification, the label and the field's record.

A good habit: after you have the draft report, write each recommendation yourself in one line with its reason, in a standard form such as "Action; product and rate; timing; reason; check after". Only then ask AI to tidy the wording. If you want a second look, ask it to find omissions rather than to suggest products: "List any information a grower would need to act on this recommendation that is missing from the line." The answer might be "who applies it", "water volume" or "harvest interval", which you can then confirm from your records and the label.

Treat anything the model suggests about chemistry, rates or restrictions as unverified. Product approvals, label conditions and restrictions change and differ between places, and a model's memory of them may be out of date or simply wrong. Always read the current label yourself, and see checking AI quality, not just speed for a simple routine to apply before you sign.

Step 5: draft the grower advice letter

Once the report and your recommendations are confirmed, ask for a letter in the grower's terms. Growers want to know what you saw, what to do, when, and what you will check next. They do not need a lecture on pathology.

Write a letter to the grower from the confirmed report below. Tone: plain, direct, courteous, as a trusted adviser. Structure: one sentence on what I found; the actions in a short numbered list with timing; one paragraph on why, in non-technical words; what I will check next and when. Use only the confirmed content. Do not add products, rates or dates. Length under 300 words. Sign off with my name, [your name].

Confirmed report:
{{paste}}

A short illustrative extract of what comes back: "Dear [first name], I walked North Barn wheat today. Septoria is present on the lower leaves, heaviest by the hedge, and no yellow rust. 1. Apply the first fungicide when leaf 3 is fully out, around stage 32 to 33, using the product and rate I confirm below. 2. Leave the gate patch for now; I will look at it again in two weeks." The model has kept to the report.

Adjusting for the reader

Some growers want three lines and a table; others want the reasoning. Keep two or three letter styles in a project or saved instructions and name them: "brief", "full reasoning", "landlord copy". Choose per client. Over a few weeks, ask the assistant to learn your phrases by pasting in two of your own best letters as style examples, and tell it to match the sentence length and tone, not to copy content.

Step 6: a wording check before you sign

A last pass catches the errors that creep in between report and letter. Ask a fresh chat, so the check is not anchored on the drafting, to compare the two documents.

Compare the letter with the confirmed report. List: (1) anything in the letter that is not in the report; (2) anything in the report that is missing from the letter; (3) any word that is stronger or weaker than in the report (for example "will" versus "may"); (4) any number, date or product that differs. If there are none, say so. Do not rewrite the letter.

Then read the letter out loud once. If a sentence tells the grower to do something you would not say face to face, change it. Before sending, add the standard line you use for conditions, for example that advice depends on conditions at the time of application and on current label directions. Wording of disclaimers and professional terms is your business and your insurer's, so keep your usual text rather than letting AI invent one.

Four situations to handle on purpose

A walk with a problem you cannot name

If you see something you cannot diagnose, say so in the note ("cause unclear, sample sent / second opinion needed"). The report should state that the cause is not confirmed. Do not let the draft smooth the uncertainty into a likely cause; ask for [TO COMPLETE] markers instead.

Several fields on one farm

Dictate each field as a separate note, then ask for one farm summary at the top: the same issue across fields, what to prioritise first, and a table of fields against actions. Check that figures from different fields have not been mixed. A realistic error is a tiller count from field A appearing under field B.

A grower who disagrees

If a client questions your advice, your records are the answer. Because the report, photos and signed letter sit together with dates, you can show what was seen and when. Do not use AI to write the reply argument; write the reply yourself, and use AI only for a polite tidy-up.

Records for the grower's farm assurance

Many growers need written evidence that advice was given, by whom and when, for their own assurance schemes and audits. Keep the signed letter, the report and the photos in a consistent folder structure by farm, field and date, and tell the client where they are. Ask the assistant to produce a one-line index per visit if it helps, but confirm each line.

A filled-in report, start to finish

Here is the report an illustrative walk might produce after you have corrected the draft. It is the standard to aim for: short, traceable and honest about what is open.

Field report: North Barn, winter wheat, [variety], growth stage 31 to 32
Date: [date]  Conditions: dry, light wind, soil wet near gate
Observations: Septoria on leaves 4 and 5; about 15% leaf area on leaf 5 at stops 1 to 3 (hedge side), lower elsewhere. No yellow rust seen at 10 stops. Pale patches near the gate. Light cleavers on the south headland.
Measurements: tillers about 650 per m2 (stop 6, photo 7)
Possible causes (my view): pale patches look like compaction in wet soil
Recommendations: 1. First fungicide at GS 32 to 33 once leaf 3 is fully emerged; product and rate [TO COMPLETE]. 2. No herbicide action on cleavers this visit; reassess at next walk.
Recheck: gate patch in two weeks (photo 9)

The two [TO COMPLETE] markers are deliberate until you have read the label and the field's spray record. A report with an honest open item is better than one with a confident guess.

A before and after: tightening a letter

A first draft of the grower letter might read: "It was a pleasure to visit your farm and I am delighted to report that the wheat is looking generally healthy with some areas of concern that may warrant attention in due course." The grower learns nothing from it. After you ask the assistant to cut it to what the grower needs and add timing, you get: "Wheat on North Barn is in good order. Septoria is on the lower leaves, worst by the hedge. Spray timing is when leaf 3 is fully out, about two weeks away on current growth; I will confirm the product and rate by Friday." It is shorter, it commits to a time for the next contact, and every word traces to the report.

Several fields, one prompt, a quick table

Suppose a visit covers three fields. Dictate each separately, then give the three confirmed reports to the assistant and ask for a one-table summary the grower can read in thirty seconds. An illustrative result:

The risk in this step is mixing fields, so always read each row against its own report. Ask the assistant to flag any value it took from the wrong field; models sometimes carry a number across when two notes use similar wording.

A mistake that nearly went out

An illustrative case: an agronomist dictated "do not apply before the end of the month" about a product, and the transcription dropped "not", giving "apply before the end of the month". The draft letter repeated it. The Step 6 wording check could not catch it, because the report and the letter agreed with each other: both carried the transcription error. It was caught only when the agronomist read the letter aloud before signing and remembered saying the opposite. The routine fix is to listen back to any recommendation that contains a negative, a date or a number, and to keep the audio until the letter is signed.

What this saves, as a quick sum

Take the 12 walks a week from earlier. Writing by hand at 40 minutes each is 480 minutes, or eight hours. With the routine at about 17 minutes each it is 204 minutes, or 3.4 hours. The saving is about 4.6 hours a week, or roughly 200 hours across a 44-week season. At an illustrative charge-out value of $60 an hour, that is $12,000 of time, which can go to more visits or to evenings at home. The tools themselves cost an assistant subscription and perhaps a transcription plan; check the current prices on each vendor's page, since they change.

Privacy and data

Field notes, yields and input records are commercial information, and the grower relies on you to keep them confidential. Use a business-grade account and check whether the vendor trains on your content, which you can usually control in settings. Strip the grower's name and address from what you paste where practical, use field codes, and add personal details in the final letter yourself. Store transcripts for no longer than you need, and tell clients in your terms that you use software to help prepare reports. For a broader checklist, see is it safe to put customer data in ChatGPT.

Trying it: a three-week trial

Do not switch over all at once. Run the new routine in parallel for three weeks, in the spirit of shadow-mode piloting.

1. Week 1: write your reports and letters the usual way, but also run the AI drafts for three walks. Compare them. Note every difference that mattered.
2. Week 2: send letters from the AI route for the simpler walks, with full review, and keep your usual process for complex ones.
3. Week 3: time the whole thing. Record minutes per walk, the number of corrections you made, and any that would have reached a grower wrongly.

If corrections concern recommendations, revisit the prompt to stop the model adding content. If they concern figures, change how you dictate numbers. A realistic result (illustrative): the average falls from 40 to 17 minutes a walk, with about two corrections a report, mostly numbers and softened or hardened wording. Across 12 walks that saves roughly four and a half hours a week. For voice-note records on the livestock side of a mixed farm, this comparison of Herdwatch and AgriWebb voice notes shows how those tools handle dictation.

What stays the same

The agronomist's name is on every recommendation, and that does not change with tools. The routine above is one in which AI drafts and checks the form while you decide the content, and nothing reaches a grower that you have not read. If you want help setting up the templates, the prompts and the review steps for your practice, that is the kind of thing my AI implementation consultation covers; you can get in touch here. The same method will also work for a farm business's own paperwork, as described in this tutorial on farm records and paperwork.

Questions about AI for crop walk paperwork

Can AI make the agronomic recommendation?

It should not. Recommendations depend on field history, weather, label rules, resistance risk and your professional judgement, and you are responsible for them. Use AI to structure notes, tidy wording and check consistency. The product, rate, timing and reasoning must come from you, and every letter should carry your name and signature.

Is it safe to put grower and field details into a chatbot?

Treat field records as client confidential. Use a business plan or a tool where training on your inputs is off, check the data terms, and leave out anything you do not need, such as bank details. Where you can, replace the grower's name with a code in the notes and add it back in the final letter.

How accurate are AI photo captions for crop problems?

Use them for describing what is visible, such as leaf colour, stage and pattern in the field, not for diagnosis. Models confuse nutrient deficiency, disease, herbicide effects and weather damage. Write the diagnosis yourself, then let AI tighten the caption wording around it. Keep the original photos and location notes with each report.

Will growers accept AI-assisted letters?

Growers care about whether the advice is correct, timely and clear, and whether you stand behind it. A letter in your voice that you have read and signed meets that. If a client asks how you prepare reports, answer honestly: you dictate notes and use software to format them, and you check everything.

Further reads

• How to Set Up Human Review for AI Work Without Slowing Down — A light review step before anything is sent.
• AI Assistant for Farmers: ChatGPT, Claude or a Farm App? — Compare general assistants with farm-specific apps.
• AI Chatbot for Agriculture: A Farm Supply Business's Setup and Cost — How an agricultural supplier might use AI with customers.
• Free AI for Agriculture: What a Small Farm Gets Without Paying — What free AI tools can and cannot do on a farm.
• How to Scale AI From One Workflow to the Whole Business — Extend the routine once the first one is steady.
• Farmonaut Pricing vs FieldView Plus: What Satellite Monitoring Costs — Farmonaut plans by area and FieldView Plus from $649 a year, converted to cost per hectare and acre for five farm sizes.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general good practice for professional advice records; no vendor prices or plan limits are quoted. Product, rate and timing rules come from labels and your own qualifications, not from AI output.