'22 Is Open, 87 Is a Recip, Roll and Tack 14' — Why General AI Mangles Herd Check and a Cattle-Trained One Doesn't

Ask a general-purpose assistant what "she's a short-solid, gate-cut, eight-fifty on a two-slide over a five-fifty base with three percent pencil shrink" means, and watch what comes back. You will get something fluent, confident and wrong. Now ask it to write up a herd check: "22's open, 87's a recip, 103 fresh with an RP, 14's a DA — roll and tack her, and give 41 Excede, note the withhold." A general model will render "recip" as "recipe," turn "DA" into something medical and generic, and quietly drop the withhold. On a dairy, that dropped withhold is the most expensive word in the sentence.

This is not a complaint about AI being dumb. It is a description of what a trade sounds like from the outside. Cattle people compress an enormous amount of meaning into two-syllable shorthand, and the compression is the whole point when you are calling cows through a palpation rail at ninety head an hour.

The 2026 development: models trained on one industry's own language

Through 2026, vertically trained AI — models trained on a single industry's records, vocabulary and edge cases — became its own category rather than a niche. The reason is exactly the herd check problem. General models are trained on the whole written world, in which "open" means not shut and "fresh" means not stale. In this business, open means not pregnant and fresh means recently calved, and those two words drive the two most expensive decisions on the dairy.

A vertically trained model is one that has been taught a trade's vocabulary, units, abbreviations and edge cases, so that it writes what the veterinarian actually said instead of the nearest thing in ordinary English. Alongside it, speech recognition got dramatically better and faster at noisy, accented, jargon-heavy audio — NVIDIA's Nemotron Speech recognition released this year runs roughly ten times faster than the older generation, which is what makes a live barn microphone practical instead of a science project.

The specific words a general model gets wrong

It is worth listing them, because the list is the argument.

• Reproduction: open, bred, springer, recip, flush cow, carrying 60 days, resynch, DNB. A model that hears "recip" and writes "recipient" is fine; one that writes "recipe" costs you an embryo transfer record.
• Fresh cow problems: DA and LDA, RP, ketosis, milk fever, hardware. "Roll and tack" is a surgical instruction, not a construction one.
• Feed: as-fed versus dry matter is the single most dangerous unit confusion in the business. "Twelve pounds" of a wet byproduct means one thing as-fed and something wildly different on a dry matter basis, and mixing them up changes the ration for a whole pen. Add NDF, starch, thirty-hour digestibility, refusals, shrink.
• Milk: hundredweight, somatic cell count, days in milk, components, the deductions on a co-op statement. "SCC 210" and "210 days in milk" are the same three digits and completely different facts.
• Beef and marketing: expected progeny differences by trait, frame score, body condition score, short-solid and broken-mouth for age, pairs versus bred cows, preconditioned, pencil shrink, and the price slide — where the cents-per-hundredweight bid falls as the calves come in heavier than the base weight. Get a slide wrong on a load of sixty calves and you have miscounted several thousand dollars.
• Numbers that look alike: a herd tag, an official 840 electronic tag, a registration number and a lot number are four different identifiers for the same animal, and a general model will happily treat them as interchangeable.

flowchart TD
  A["Vet calls findings at the palpation rail"] --> B["Barn microphone picks up the audio"]
  B --> C["Cattle-trained model splits cow, finding, action"]
  C --> D["Pregnancy checks with days carried"]
  C --> E["Open cows onto the resynch list"]
  C --> F["Treatments with drug, dose and withhold"]
  C --> G["Recips and flush cows to the transfer sheet"]
  D --> H["Herdsman signs off on the tablet before it posts to DairyComp 305"]
  E --> H
  F --> H
  G --> H

Thursday morning at the rail, the way it goes now

Herd check is Thursday at 8. The vet works the rail with the herdsman, the breeder brings cows up, and somebody normally walks behind with a clipboard or a tablet trying to key entries while three people talk. That person is the bottleneck and the error source, and on a lot of dairies that person is also the herdsman, who is supposed to be looking at cows.

With a trained model listening, the walk-behind job changes from typing to confirming. The vet says "one-four-seven, thirty-five days, good"; it writes a pregnancy check at 35 days carried against cow 147. "Twenty-two open, put her on the resynch" becomes an open result and a list entry. "Fourteen's a DA, roll and tack, Banamine today, note the withhold" becomes a treatment with a drug, a date and a milk withhold date computed off the label. At the end, the herdsman reads a one-screen summary — thirty-one checked, twenty-two carrying, six open, two on the do-not-breed list, three treatments — and signs off before any of it posts to DairyComp 305.

The same tuning pays off in the parlor. If your milking crew speaks Spanish and your records are in English, the newest live speech translation covers seventy-plus languages nearly in real time, but the translation is only as good as the vocabulary underneath it — a general translator turns "cuarto trasero izquierdo" into a literal phrase, where a trained one writes it as left rear quarter into the mastitis record.

What the mis-writes actually cost, in days open

Take the herd check job on its own and price it the way a dairy prices everything: in days open.

Those are stated assumptions, not measured results, and the honest way to test them is to run one herd check both ways and compare the two lists. If your mis-write rate is really 1 percent already because your herdsman is meticulous, the case here is smaller and you should know that before you spend anything.

What training on the trade does not fix

It does not fix a vet who mumbles or a barn with a compressor running six feet away. Audio quality is still the limit, and no model recovers a cow number that nobody said clearly. In practice this means one microphone in the right place, and a habit of calling the cow number first and loudest — which good crews already do.

It does not make the record legal on its own. The drug log, the written protocols your veterinarian signed, and the Veterinary Feed Directive for anything in the feed are still documents with a licensed person behind them. A trained model is very good at capturing what was said and computing a withhold from the label. It is not the prescriber, and it should never be the one that clears a cow to go back in the tank.

And it will not know your farm's private shorthand on day one. Every dairy has pen names, protocol names and a couple of words that mean something only there — "the Poole barn," "protocol 4," a nickname for a specific bull. That is a short training conversation and a list you write once, not a reason to skip the whole thing. Budget a week of the herdsman correcting it, and correct it in the same place every time so the corrections stick.

Frequently asked questions

Do I need a special product, or will the big general assistants do this?

The general assistants are genuinely useful for office work — writing the letter to the co-op, summarising a lease, sorting the feed invoices. For anything that gets spoken at the rail or written on a treatment sheet, get something trained on cattle language, or at minimum give it your own glossary and event codes and test it on a recorded herd check before you rely on it.

How do I test whether it actually knows the vocabulary?

Read it ten real lines from last month's herd check and ten lines from an order buyer's bid, and grade the output yourself. If it gets slide, shrink, open, fresh, recip and withhold right, it has been trained on this trade. If it turns any one of those into ordinary English, it has not, and it will fail in the exact place that costs you.

Can it write into DairyComp 305 by itself?

Technically it can be set up to, and you should not let it at first. Run it for a month producing a list the herdsman confirms and imports. Once you have a month where the confirmed list matches what it proposed, then talk about letting it write directly — and even then, keep treatments and withhold dates behind a person's approval.

What about the beef side — will it handle an order buyer's bid?

A trained model handles the slide and the pencil shrink correctly, which is the part general models fail at, and it will write out what a load actually pays at three different weigh-up weights while you are still on the phone. Treat that as a calculator that checks your math, not as a negotiator. The negotiation is you.

Worth adding: the same shorthand shows up on the phone. The order buyer, the hauler, the nutritionist and the neighbor looking for hay all call the farm line and all speak this language, and most of those calls arrive while everybody is in the barn. CallSphere builds voice and chat agents that answer the line and take a message that actually holds the details — the cow number, the load weight, the delivery date — instead of a voicemail that says "call me back." The herd records still belong to your herdsman; the phone does not have to.