Spray Records With AI: Voice Notes to Logs, Then Check Doses

Record what you did by voice as soon as you finish spraying, have AI turn the transcript into a log with every required field, and then check the product, rate and intervals against the printed label yourself. AI is for layout and spotting gaps. It must never be the source of a dose.

The reason this works is timing. Most incomplete spray records come from writing them up at the end of the week from memory, when nobody remembers the wind speed or whether the second tank was mixed at the same rate. A thirty-second voice note made in the cab, with the engine off, captures the facts while they are fresh, and the AI does the tedious part of arranging them.

Follow me on Instagram@sagnikteaches

The same step is also where mistakes creep in, so the method has two halves: capture and tidy, then verify. The first half saves time; the second is the one that protects you at an inspection or a buyer's audit. Whenever the AI's version of a product, rate or interval and the label disagree, the label wins.

Connect on LinkedInSagnik Bhattacharya

What a complete record contains

The exact list of required fields depends on where you farm, what you grow and which assurance or buyer scheme you sell into, so check your own regulator and scheme documents. Most lists contain the same core items, and a record that has all of them is a good base to adapt.

Subscribe on YouTube@sagnikteaches

Print this list, or put it in a sheet, and use it as the checklist the AI works to. A record is only as good as its most commonly missed field, and the table's third column is where I would look first.

How to dictate so the AI has something to work with

The quality of the log depends on the quality of the note. Speak in a fixed order, and say numbers with their units. A good habit is a fixed script that you say aloud every time, like a pilot's checklist.

Spray note. Date today. Field: Lower Meadow, block 4. Crop: winter wheat, growth stage thirty-nine. Product: [trade name]. Rate: one point five litres per hectare. Area sprayed: six point four hectares. Water: two hundred litres per hectare. Target: leaf blotch. Started eight fifteen, finished ten forty. Wind: about eight kilometres an hour from the west. Temperature twelve degrees. Operator: me. Sprayer: the trailed one, flat-fan nozzles, three bar. Buffer next to the stream respected. Rinsed in the field. Anything else: tank two mixed fresh, same rate.

Say what you did, not what you intended. If you changed the rate halfway, say so in your own words. Do not tidy the note in your head before speaking, because the AI will do the tidying.

Getting the words onto the page

You have several options, and none needs special equipment. On many recent phones the notes app can record and transcribe audio, or you can dictate straight into a note; some chatbots accept audio files too (Gemini's help pages, for example, list audio among the file types you can upload, up to 10 minutes in total on the free tier), and where yours does not, paste the transcript in. Test the route with a real note in the cab, because engine noise, wind and gloves change what works. Whichever you use, keep the original audio for a while and compare a few transcripts with what you said, because speech recognition struggles with trade names and fields with unusual names.

Keep numbers unambiguous. "One point five" is safer than "one-five", and saying "litres per hectare" every time stops the AI deciding the unit for you.

The prompt that builds the record

The prompt does three jobs: extract the fields, refuse to guess, and list what is missing. Keep it in a Project or a saved note so you do not retype it.

You turn a farmer's spoken spray note into a spray record. Rules:
1. Use ONLY information in the note. Never fill a missing value with a typical or likely one. Write "MISSING" instead.
2. Keep every number and unit exactly as spoken. If a unit is unclear, write the number and mark "UNIT UNCLEAR".
3. Do not look up, confirm or correct any product, rate or interval. I will check these against the label myself.
4. Output a table with these fields: date, start time, finish time, field/block, crop, growth stage, product (trade name), reason/target, rate per hectare, area treated, total product used (calculate = rate x area, show the sum), water volume per hectare, total water, operator, equipment, nozzle and pressure, wind speed and direction, temperature, buffer zones respected, rinse/disposal, notes.
5. After the table, list: (a) fields marked MISSING, (b) anything in the note that is ambiguous, (c) anything that looks inconsistent (for example times that do not fit the area).
6. Do not give agronomic advice.

NOTE: [paste transcript]

Rule 3 is the safety valve. Without it, many models will helpfully "correct" a product name or add the label rate, which blurs the line between what you did and what the AI thinks you should have done. That distinction is the whole point of the record.

What comes back, and what you fix

For the note above, an illustrative response:

| Field | Value |
| Date | MISSING (note says "today" - confirm date) |
| Start / finish | 08:15 / 10:40 |
| Field / block | Lower Meadow, block 4 |
| Crop / stage | Winter wheat / GS 39 |
| Product | [trade name] |
| Target | Leaf blotch |
| Rate | 1.5 litres per hectare |
| Area treated | 6.4 hectares |
| Total product | 1.5 x 6.4 = 9.6 litres |
| Water | 200 litres per hectare; total 1,280 litres |
| Operator | "me" - full name MISSING |
| Equipment | Trailed sprayer, flat-fan nozzles, 3 bar |
| Wind | about 8 km/h, west |
| Temperature | 12 degrees (unit assumed as Celsius - UNIT UNCLEAR) |
| Buffer | stream buffer respected |
| Rinse | rinsed in the field |

Missing: date, operator full name, approval number, re-entry and harvest intervals, nozzle size.
Ambiguous: "tank two mixed fresh" - was tank one a different rate?
Inconsistent: 6.4 ha at 200 L/ha is 1,280 L; check against tank capacity.

You still have work to do. Fill in the date and name, add the approval number from the label, check the "tank two" remark (it flags a possible rate difference), and answer the tank-capacity query. The AI did what it was told: it arranged the data and exposed the holes.

Checking the dose against the label: your job, every time

This step cannot be delegated. AI models can misquote rates, mix up products with similar names, or describe an older version of a label from memory. A label also differs by crop, by growth stage and by country or scheme, and what the AI knows may not match the bottle in your shed. So the rule is simple: the printed label, or the current official label for that product, is the only authority for the dose.

Do the check in this order, with the container or the current label in front of you.

1. Product identity. Is the trade name in the record exactly the one on the container? Look for look-alike names and different formulations.
2. Crop and use. Is the product approved for this crop, and for the target you wrote down?
3. Rate. Is the rate you applied at or below the label maximum for this crop and stage? Check the unit (litres or kilograms per hectare).
4. Number of applications. Does this application, with earlier ones this season, stay within the per-crop or per-season limit?
5. Timing. Is the growth stage inside the label's window?
6. Water volume and nozzle. Does the label specify a minimum volume or a drift-reducing nozzle?
7. Weather and buffers. Do wind and temperature fit the label's limits? Were buffer zones respected?
8. Intervals. Read the re-entry period and the pre-harvest interval, then calculate the earliest harvest date and write it into the record.

Add a "Label checked" column to the record: the date, your initials and the label version or edition if shown. If the AI suggests something that the label does not say, trust the label, and treat the suggestion as a reason to look harder.

Doing the sum yourself

Total product is rate times area. 1.5 L/ha across 6.4 ha is 9.6 L. If the note said the sprayer carried 1,000 L tanks, 1,280 L of spray means 1.28 tanks, so two loads, with the second part-filled. That matches "tank two mixed fresh". The arithmetic is simple, but AI sometimes slips on multi-step conversions, as why AI is bad at maths explains. Check each calculated line with a calculator, especially total product, because a misplaced decimal there leads to a wrong order quantity as well as a wrong record.

Five cases that show where the system helps and where it bites

The dose that was higher than the label

A note says "one point eight litres per hectare". The label maximum for that crop is 1.5. The AI, following rule 3, records 1.8 and does nothing else. The label check catches it. That is the system working: a record that shows the real figure is honest, and the discrepancy is a matter for you and your adviser. Never edit the record to match the label after the event. Record what happened, and note the corrective step.

Mixed units in one note

The farmer says "two hundred" for water volume without a unit and "five hundred" for product used, meaning millilitres per tank. The AI marks both "UNIT UNCLEAR". Without rule 2, it might have assumed litres per hectare for both and produced a record showing 500 L of product on 6 ha.

A product name the transcript mangled

The speech tool writes a trade name as a different, real product. The AI repeats it faithfully. Your product check against the container catches it. If your speech tool lets you add words to a personal dictionary or custom vocabulary, add your trade names; otherwise spell them out at the end of the note.

Two products in one tank

You mix a fungicide and a growth regulator. The prompt's table has one product field. Add a rule to produce one row per product, repeating date, field and tank, with the mixing order noted. Check each label for compatibility and rate when tank-mixed, because mixing can change what is allowed.

An old paper diary to digitise

You have three seasons of handwritten records. The photo-to-sheet approach in turning handwritten forms into spreadsheet data works, but read each line back against the page. Handwriting misreads are common, and a wrong digit in a rate is exactly the error an audit finds.

When the record was written three days late

A common failure is worth walking through. A grower sprays on a Monday, forgets to dictate, and writes the entry on Thursday from a mixture of memory and the sprayer's job sheet. The note says "wind light" and "about 6.4 hectares". The AI turns it into a clean table, and nothing in the table shows that the details are a reconstruction. Two weeks later a buyer's audit asks for the wind speed at the time. There is none, only the word "light".

The fix is a rule about honesty, not a better prompt. Add a "Recorded when" field with the options "at the time", "same day" and "reconstructed", and let the AI copy it from the note, for example "this is from memory, written on Thursday". A reconstructed record is not worthless, but it should say so. Where you can, fill gaps from objective sources, such as the sprayer's logged area and time, or a weather station's history, and cite them in the notes column.

Cab-side habits that make the notes better

Small habits at the point of dictation do more for record quality than any prompt tweak.

• Speak before you climb down. The best moment is the engine switched off and the sprayer still in the field, when the numbers on the monitor are still showing.
• Read from the monitor, not from memory. Area, volume and rate often appear on the sprayer's screen. Say them as shown.
• Name the field the way the record names it. If the sheet says "Lower Meadow, block 4", say that, not "the bottom one".
• Say the product from the container. Read the trade name off the label as you rinse, and spell anything unusual.
• Add the odd thing. A blocked nozzle, a stop for rain, a neighbour's gate left open. These are the notes an inspector or your own future self will ask about.
• Take one photo. A picture of the monitor and the label next to the product gives you something to check the transcript against.

A filled-in row and a season summary

Once the record is checked, the finished row should read like a complete, boring fact. An illustrative row, after the label check:

Date: 2026-06-09 | Field: Lower Meadow, block 4 | Crop: winter wheat, GS 39
Product: [trade name], approval no. [from label] | Target: leaf blotch
Rate: 1.5 L/ha | Area: 6.4 ha | Total product: 9.6 L | Water: 200 L/ha, 1,280 L
Operator: [full name], certificate [number] | Sprayer: trailed, flat-fan, 3 bar
Wind: 8 km/h W | Temp: 12 C | Buffer: stream, respected
Intervals: re-entry [from label]; earliest harvest 2026-07-14 (label interval 35 days)
Recorded: at the time | Label checked: 2026-06-09, initials, label edition [shown]

With a season of rows like this, the AI can do useful summarising from the sheet: total product used per field, number of applications per crop against a limit you entered, or a list of fields that have received the same active ingredient more than twice. Give it the sheet and ask for a table, then check two or three totals by hand. That is a safe use, since the summary describes data you already verified.

Setting up the log itself

You need somewhere for finished records to live, and it need not be complicated. A spreadsheet with one row per product per application, and columns matching the field table, is enough for most holdings.

• Locked headings. Use the same column names as the prompt's table, so you can paste the AI's output straight in.
• Status column. "Draft", "Label checked", "Complete". Only "Complete" rows count.
• Calculated columns. Earliest harvest date (spray date plus interval), total product (rate times area), and a flag if the rate exceeds a maximum you have entered for that product and crop.
• A products sheet. One row per product you use: trade name, approval number, label maximum rate by crop, intervals, and the date you last checked the label. This sheet is your own, filled from the printed label, never from the AI.
• Storage. Keep the sheet in an account with a strong sign-in and a second copy elsewhere. Keep the voice recordings or transcripts with the matching row, if you choose to retain them.

With a products sheet, the AI can help with a safe task: compare the record's rate with the maximum you entered. Ask it to flag any rate above that figure, and treat the flag as a prompt to check the label, since the sheet is only as right as you made it. Whether farmers can use AI for farm records and paperwork covers the wider picture.

A weekly routine, ten minutes

Records stay accurate when someone looks at them regularly. A short weekly pass catches gaps while the memory is still fresh.

1. Open the log and filter on "Draft". Fill each missing field, or write why it cannot be filled.
2. For each draft, do the label check and change the status.
3. Scan the earliest-harvest-date column for any field due to be harvested soon and confirm the interval has passed.
4. Check the products sheet for any label that changed or any new product bought.
5. Back up the sheet.

This routine is one of the few things I would not automate further. Putting a human check on the output before it counts as final is the idea behind human review for AI work, and records that may be inspected are exactly where it belongs.

Privacy and the limits of the tool

Spray records seldom contain personal data, but they do show your fields, crops and purchasing, and some include operators' names and certificate numbers. Before you paste a transcript into a chatbot, check that tool's data settings, and where you can, use a business plan or turn off training on your chats. Leave out anything you would not want stored, such as bank details or phone numbers mentioned in the audio. Keep your master record in your own file, not only in a chat history.

Know what AI cannot do here. It cannot read a product label it has not been shown, and even when shown a photo it may misread small print. It cannot tell whether a product is approved for your crop where you farm. It cannot judge drift risk from the conditions you describe. If you ask it any of these, treat the answer as a suggestion to check, not a fact.

Photographing the label next to the container is a good habit, and it helps you read it later. If you upload such a photo, ask the AI to transcribe the text only, and compare the transcription with the original. What breaks when ChatGPT reads PDFs, spreadsheets and photos lists the usual failures, including misread numbers in small type.

Finally, record keeping rules differ by place, crop and scheme, and inspectors or buyers may ask for a particular format. Treat this as a method to adapt, take your required field list from your own regulator or scheme, and put any doubt about a rate, an interval or a legal duty to a qualified adviser before you spray, not after.

Questions about AI and farm spray records

Can AI be trusted to check my spray dose?

No. Use it to lay out the record and to flag gaps, then compare the dose with the printed label for that product, crop and date yourself. AI can misread a rate, mix up units or recall an old label. The label is the legal document, and it changes between product versions.

What if I forget a detail in the voice note?

Have the AI list every required field it could not find, and keep the record as a draft until you fill them. Never let it guess a value. A missing wind speed or start time is better shown as blank than filled with a plausible number.

How long should I keep spray records?

It depends on where you farm and which assurance or buyer schemes you supply, so check your own regulator's guidance and your scheme rules. Keep the originals, the voice recordings if you choose to, and the final record together, and ask your adviser or scheme body for the retention period.

Is it fine to put farm records into a chatbot?

Spray records are mostly business information, though operator names and certificate numbers are personal data, and they can show your fields, your yields and your suppliers. Check the plan's training settings, leave out anything you would not want stored, and keep your own master copy outside the AI tool.

Should I use a template or a farm app instead?

Both are reasonable. A spreadsheet template with the AI filling it from dictation costs almost nothing and you control it. Dedicated farm software can lock required fields and remind you about intervals, which suits larger operations. Start with the sheet and move when it hurts.

Further reads

• How Surveyors Use AI to Turn Site Notes Into Reports — The same site-notes-to-report pattern in another trade.
• AI Compliance Checklist for Small Businesses: What Applies to You — Work out which rules apply to your business.
• How to Chase Missing Client Records With AI Before Deadlines — Chase gaps in records before an inspection or audit.
• How to Clean Up Customer Records Before You Add AI — Tidy a messy records sheet before you add AI.
• Outgrowing Spreadsheets: When to Replace Manual Excel With AI — Know when a sheet stops being enough for your records.
• Refrigerant Log: Record Leak Checks and Top-Ups on Every HVAC Job With AI — Per-job fields for leak checks and top-ups, a voice-note routine and an AI prompt that structures the entry, with a worked cold-room repair.
• Livestock Record Keeping App: Herdwatch vs AgriWebb AI Voice Notes — Which livestock record keeping app turns spoken notes into records: what AgriWebb's Farm Diary does, what Herdwatch's Herdi does, and a notes-app fallback.
• AI Chatbot for Farming: What It Can Answer and Where It Fails — What a chatbot can reliably do on a farm, with sample prompts, and the agronomy, livestock and safety questions it should never settle.
• AI Crop Monitoring on a Small Farm: Does It Work Without Drones? — How satellite-based crop monitoring works without drones, what it costs per hectare, and why fields under about a hectare need extra care.
• Do Farmers Use AI? Eight Uses of Artificial Intelligence, With Costs — How many farmers use AI, and eight small-farm uses with indicative costs, example prompts and the checks that keep each one honest.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: product labels and safety data sheets for the products you use, your regulator's and assurance scheme's record-keeping guidance (check the current version for your market), Gemini Apps Help on uploading files, Apple's iPhone User Guide on recording and transcribing audio in Notes, and the ChatGPT and Claude help pages on Projects.