Most pages about a farm assistant open with what it can do. This one opens with what it cannot, because we built the thing and the list of what it refuses to do is the part we are most confident in. It is also the part a reader has the least reason to take on faith, so we are writing it down.
None of the items below are marketing caveats added after the fact. They are decisions made before the assistant read its first question, and each one closes off a way an AI feature could go wrong on a farm specifically — not in the abstract, but at the loading chute, on the spray record, in the number that decides next year’s planting.
An assistant is not a source of regulatory truth. Never take a withdrawal period, a re-entry interval, or a certification requirement from any AI, including ours. Those figures come from the product label, the certifier, and the veterinarian, and they vary by jurisdiction, formulation, and use. What follows is a list of what an assistant should be trusted for and, more importantly, what it should not.
It cannot create, edit, or delete a record
Start with the largest one. A read-only assistant cannot write to your farm at all — not a treatment, not a sale, not a note. It can only answer questions about records that already exist, using a connection that does not permit writing. This sounds like a modest restriction. It is closer to the whole safety story: a system that cannot write cannot be talked into writing something wrong, whether by a confused question, a bug, or someone typing something adversarial into a field it later reads. Every other item on this list is a detail. This one is the shape.
Ask it to log a treatment, close out a work order, or update a sale, and the correct behaviour is to say plainly that it cannot — not to attempt something adjacent, and not to quietly leave a note somewhere that looks like a record but isn’t one. There is no halfway version of this rule. A connection either includes a write path or it does not, and the assistant’s does not, for any request, phrased any way.
It cannot decide a regulated figure
It cannot tell you the correct withdrawal period for a drug, the correct re-entry interval for a chemical, or what your certifier requires this season. It can only quote what is already on the record — the interval you or your veterinarian entered, and the date the application calculated from it. If nothing was entered, it has nothing to quote, and it should say so rather than estimate. This is the subject of a whole page on its own: AI and regulatory decisions, because the failure mode here is serious enough to deserve its own treatment.
The reason this has to be a hard rule rather than a best effort is that a regulated figure is exactly the kind of question a fluent model will answer even when it should not — the shape of the question invites a specific-sounding number, and a model trained to continue plausible text will often supply one whether or not any record actually contains it.
It cannot cite a record — only the tool it ran
Ask it a question and the answer names which read-only tool it used to reach that answer — that it read your financial totals, say, or your livestock event history. It does not point at row 4,412 of a table or quote a specific record back at you as a citation. That distinction matters because a tool name is verifiable against the screen it corresponds to, and a fabricated citation is not always distinguishable from a real one. Treat every answer as a lead worth checking on the actual screen, not a footnote you can skip verifying.
This is a narrower claim than it might sound. Naming the tool tells you where to look; it does not tell you which specific row produced the answer, so an answer about, say, a livestock group still means opening that group’s history yourself to confirm the detail before acting on it. The assistant shortens the search. It does not remove the need to look.
It cannot do arithmetic, and it does not try
It does not add your expenses, average your yields, or compute a margin. It quotes the figure the application already computed — the same total your farm finances screens draw from the same records. A model that performs its own math will get a sum wrong occasionally, and it will sound exactly as confident when it is wrong as when it is right. Refusing to do the arithmetic at all is a narrower job than doing it, and a much safer one — the whole argument is laid out on why an assistant should not do arithmetic.
It cannot answer past a small number of tool calls, or without a quota
A single answer draws on a capped number of read-only tool calls, not an open-ended search of your entire history. That cap is a deliberate ceiling on what one question can touch, and it is a large part of why the assistant stays fast, cheap to run, and predictable about what it read. Layered on top of it is a monthly question allowance — 10 questions a month included on every account, trial included, with a $20/mo add-on raising that to 500 a month, billed separately from your plan. The Farm40 assistant is built to this list on purpose: it is scoped to your account at the database layer, it does no arithmetic, it names tools rather than citing rows, and it cannot write. The limit worth stating plainly: a capped, quoted, read-only assistant will sometimes say it does not know something a less-honest one would have guessed at, and that is the trade we chose.
What this leaves it good for
Subtract everything above and what remains is narrower than a pitch would suggest, and genuinely useful inside that range: finding a record you wrote but cannot place, and summarising across records you would otherwise reconcile by hand across several screens. Both are retrieval, not judgement, which is exactly the argument made in AI in farming as a whole and in more depth in AI for farm management. A list this short is not a marketing failure. It is the point. We would rather you finish this page knowing what an assistant refuses to do than believing a longer list of things it claims to do well, because the refusals are what you are actually trusting when you point a language model at records that can end up governing a sale or an audit.