What AI Can Actually Do With Your Farm's Own Records

On 22 September, Google and the Gates Foundation announced $100 million for organizations scaling AI climate and crop insights, aiming to reach 200 million smallholder farmers in Sub-Saharan Africa and South Asia. It is a commitment to expand access, rather than a claim that those farmers already use the tools.

Some farmers already use general AI tools in their work. The practical question is what evidence an answer uses, and what needs a human check before it becomes a real farm decision.

I think that caution is right, and I think it points at the actual problem. A general assistant can explain agronomy, but it does not know your farm unless you provide or connect that information. Without those records, it may answer a general version of your question, and the average field is not the one you are standing in.

This post uses four illustrative questions a working farmer might ask this month. For each, we look at what evidence a good answer needs and what remains uncertain even with that evidence. The field names and example answers are hypothetical.

If you are still choosing between tools, we covered that separately in how to choose agricultural AI software and our comparison of AI tools for agriculture.

1. “Can I spray North Field this week?”

Without your records, an assistant gives you a checklist: avoid rain, avoid wind, read the label. All true, and none of it answers the question.

The evidence a real answer needs:

• the exact location of North Field, so the forecast is selected for its location rather than the nearest city, while recognizing the model’s spatial resolution;
• hourly or daily rain and wind for the next several days;
• what has already been applied there this season, and when; and
• the crop and its stage.

Before deciding to apply a product, you also need its label, intended rate, target pest or disease, equipment and nozzle, and current wind and rain conditions at the field. Check label wind limits and rainfastness requirements, and avoid temperature inversions; calm air alone does not establish that spraying is safe.

What stays uncertain: forecast skill drops after the first few days, so a clean window on day six is a possibility, not a plan. The product label and your local rules still decide what is allowed, and no assistant should be treated as a substitute for them.

A hypothetical answer might say “Tuesday and Wednesday look unsuitable because rain could wash the product off; recheck Thursday against the label and field conditions” and show the forecast it used. A bad answer sounds just as confident and shows nothing.

2. “My soil test says phosphorus is 18. Do I need to apply any?”

Without your records, the assistant has to guess what “18” means. It may pick a threshold from whichever extension guide it saw most often in training.

The evidence a real answer needs:

• the lab method. The same number means different things under different extraction methods, which is one of the most common mistakes in reading a soil report;
• the units and the sampling depth;
• the date of the sample, because a five-year-old result is weak evidence;
• the crop you are planning; and
• what you applied to that field in the last few seasons.

What stays uncertain: thresholds are calibrated to regions and crops, and a table built for one country does not transfer cleanly to another. A good answer tells you which threshold it used and where it came from, so you or your agronomist can disagree with it. We went through this in detail in how to read a soil lab report for fertilizer decisions.

If the lab’s locally calibrated interpretation shows adequate phosphorus for the planned crop, your records may help you avoid an unnecessary application. The decision still depends on the crop, soil conditions and local agronomic guidance.

3. “What have we grown on the east block, and what went wrong last time?”

Without your records, the assistant cannot reliably reconstruct that field’s history. You would first have to supply the missing details.

The evidence a real answer needs:

• the crop cycles on that field, with sowing and harvest dates;
• the observations and problems recorded during those seasons; and
• the weather those seasons actually had.

What stays uncertain: the answer can only be as complete as what was written down. If a pest problem three years ago was never recorded, it does not exist as far as the assistant is concerned. A useful assistant tells you where the gaps are instead of quietly filling them with a plausible story.

This is also where records start to compound. The first season you keep them, they help a little. By the third season, you can ask whether the east block always struggles in dry autumns, and get an answer grounded in your own ground.

4. “What is still outstanding on the farm this week?”

Without your records, the assistant can only help you write a to-do list from scratch.

The evidence a real answer needs:

• open tasks and when they were due;
• recent observations that may need follow-up; and
• the forecast, because some jobs cannot happen in the weather that is coming.

What stays uncertain: whatever happened on the farm but never made it into the record. The assistant sees the log, not the field.

The value here is small per question and large over a season. It is the difference between remembering on Friday that something needed doing before Wednesday’s rain, and being reminded on Monday.

What every good answer has in common

Across all four questions, a trustworthy answer does the same things:

• It names its sources. Which forecast, which soil sample, which records.
• It says how fresh they are. A forecast from this morning and a satellite image from three weeks ago are not the same kind of evidence.
• It separates measured from modeled. An estimated soil value from a global map is not a lab test, and it should be labeled as such.
• It admits the gaps. “No soil test on record for this field” is a useful answer.
• It does not change anything without asking. Reading your records and writing to them should be separate permissions.

That list is also a fair test for any tool you are considering, including ours.

Start with the records you already have

You do not need a sensor network to begin. For one field, gather:

1. the field boundary;
2. the crop and sowing date for this season and, if you can, the last two;
3. your most recent soil test, including the lab method and date;
4. what you applied there this season;
5. recent observations and any problems worth following up; and
6. open tasks, due dates and completed work.

Those records give the assistant a useful starting point for the four questions above. It should still ask for any missing information that could change the decision.

Where ZarSage fits

I built ZarSage to be that farm record. It brings each field’s boundary, forecast, weather history for its location, recorded soil results, crop cycles, and work history together, with satellite signals on plans that include them. The historical weather comes from a gridded reanalysis archive, not measurements taken on your farm; uploaded lab results and your own observations are different kinds of evidence.

You can also connect it to ChatGPT or Claude, subject to the assistant’s account and workspace settings, so it can look up the right field before answering. Tool results carry source and timing information where available, and the assistant should explain known gaps. Read access and write access are separate, and you can disconnect whenever you choose. The setup page has the current steps for each assistant.

New subscribers start with a 7-day free trial on one field. Pick the field with a decision coming up, add what you know about it, and ask the question you would normally ask a neighbour or an agronomist.

If the answer is wrong, I want to hear about it. Write to me at musa.bello@zarsage.ai. It reaches me, not a support queue.