A calculator is right or it is broken. Ask it to add two numbers and it either gives you the sum or it gives you nothing — a cracked screen, a dead battery, some visible sign that something is wrong. A language model asked to add two numbers has no equivalent failure state. It gives you an answer in either case, in the same font, the same tone, the same confident sentence. Most of the time the answer is correct. Some of the time it is not, and nothing in how the answer is delivered tells you which time this was.
That is the entire argument for why a farm assistant should not do its own arithmetic, and it is worth sitting with the mechanism before accepting the rule, because the rule sounds like a limitation until you understand what it prevents.
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. The arithmetic problem described on this page is about financial and production numbers, but the same logic applies with even less room for error to any figure with a legal consequence attached.
Language models are not calculators, and the fluency hides it
A calculator computes; a language model predicts the next plausible piece of text given everything it has read. For arithmetic, those two things often land on the same answer, because correct sums are common in training data and the pattern is learnable. But “often the same answer” is a probabilistic guarantee, not an architectural one, and the gap between the two shows up precisely on the harder cases — more line items, more digits, a calculation nested inside another calculation. The model does not know when it has crossed from the territory where it is reliable into the territory where it is not. It just keeps writing the next plausible token, with identical fluency on both sides of that line.
This is easy to underestimate because everyday experience with these tools is mostly reassuring — ask one to add three numbers and it almost always gets it right, which trains you to trust the next one too. The failures are rare enough to feel like an edge case and common enough, over hundreds of questions across a season, to eventually land on a number you actually rely on.
A wrong margin looks exactly like a right one
This is the property that makes the failure dangerous rather than merely embarrassing. A wrong sum from a person usually comes with a tell — a pause, a “let me check that,” a note that says “roughly.” A wrong sum from a model comes formatted identically to a right one: same number of decimal places, same declarative sentence, same absence of hedging. You would have no way to tell the two apart by reading the answer. The only way to catch it is to redo the arithmetic yourself — which is exactly the work the assistant was supposed to save you from doing, so in practice, nobody redoes it, and the wrong number gets used.
Notice, too, that the incentive runs the wrong way. Redoing the arithmetic every time defeats the purpose of asking in the first place, so the rational response to a tool that sometimes miscalculates is to stop checking it — which is precisely the point at which an occasional wrong answer stops being caught at all.
What this looks like with real-sounding stakes
Say your enterprise records show three expense categories for a field: seed, fertilizer, and a fuel allocation split across several fields by acreage. A model asked “what did this field cost me this season” has to pull three figures and combine them, one of which — the split fuel allocation — is not a plain lookup but a derived share. That is exactly the kind of multi-step, partially derived arithmetic where a model is most likely to drop a term, double-count a shared cost, or silently use the wrong denominator for the split. The output would still read as a single, clean dollar figure. Nothing about its presentation would hint that one of its three inputs was handled incorrectly.
The fix is not better checking. It is not computing at all
The tempting fix is to have the model show its work, so you can audit the steps. That helps a careful reader catch some errors, but it does not solve the underlying problem, because auditing a model’s arithmetic is itself an arithmetic task you now have to do — and you were asking the assistant precisely so you would not have to. The more durable fix is structural: the assistant does not perform the addition at all. It reads a total that a fixed piece of application code already produced — the same code, run the same way every time, that draws your cost of production screens — and it quotes that number back to you.
What quoting a precomputed figure actually buys you
It narrows the failure mode to something you can catch. The assistant can still misread which figure you wanted — it might quote gross revenue when you meant net margin, and that is a real error worth watching for. What it cannot do is silently invent a number that no part of the application ever produced. Every figure it states traces back to a fixed calculation you can independently pull up on the corresponding screen and compare. A wrong figure and a quoted figure are not the same kind of wrong. One is a math error hiding inside a confident sentence. The other is, at worst, a mismatch you can catch by opening one screen.
The trade is worth stating in the plainest terms possible: quoting gives up the convenience of asking about a number that has never been computed anywhere. It buys, in exchange, the property that every number the assistant does state is one you could reproduce by opening the same screen it read from. Between those two, only one keeps the assistant honest when it is wrong.
Where this still runs out
The honest limit: Farm40’s assistant quotes figures the application has already computed and performs no arithmetic of its own, which means it can answer “what was my gross margin” but not “what would my margin have been at last year’s fertilizer price” — that hypothetical requires a new calculation nobody has run, and quoting is all the assistant is built to do. For that kind of question, pull the underlying figures from an export and do the arithmetic yourself, where you can see every step. That is slower on purpose. It is also the version of this feature that will not quietly hand you a wrong number dressed as a right one — the same discipline described more broadly in what a good AI answer looks like and in AI for farm management.