Before a Voice Agent Answers Bull Sale Calls, Build the Test Set Out of Last February's Phone Log

You tried this in 2024. Somebody sold you a phone answering thing before the spring sale, it told a buyer from Nebraska that Lot 42 had a birth weight expected progeny difference of negative four when the catalog said plus one point six, and you turned it off inside a week. That was the correct decision with the tools that existed then. What changed in 2026 is not that the agents got smarter — it is that you can now prove whether one is right before it talks to a customer, using calls that already happened on your own place.

Here is the problem it is aimed at. On a seedstock outfit, February is calving season and sale season at the same time. The owner is checking heifers at 2 a.m., the herdsman is in the calving barn, and the office is a kitchen table. The catalog went out three weeks ago and the phone rings all day: is Lot 42 still available, what is his scrotal, did he pass his breeding soundness exam, do you free-board until April 1, will you deliver to Torrington, what did his sire's daughters do. A large share of those calls go to voicemail, and a buyer working down a pile of four catalogs does not leave a second message.

Why "it sounds good in the demo" is not evidence

Every voice agent sounds good in a demo, because the demo asks it three easy questions. Cattle buyers do not ask easy questions. They ask compound ones — "what's the calving ease direct on the son of the 4038 bull, and is he PAP tested, because I'm at seventy-two hundred feet" — and the failure that costs you money is not a rude answer, it is a confidently wrong number read off the wrong lot.

Evaluating an agent means running it against a set of real calls whose correct answers you already know, scoring what came out, and reading back the steps it took to get there — before it is allowed anywhere near a live buyer. That is the 2026 development in one sentence, and it is unglamorous on purpose. Through 2025 you mostly had to take a vendor's word for it. Through 2026 the tooling to watch an agent work step by step, replay a call, see which page of your catalog it pulled a number from, and hold it to a pass rate on a fixed set of cases became the normal thing you do before turning one on.

The test set a cattle operation already owns

You do not have to invent test cases. You have a year of them. Pull the call log and the voicemail box from last February and March, plus the emails from the sale week, plus the question sheet your ring man scribbled on. Two hundred real calls is plenty. Then do the boring part: for each one, write the answer that is actually correct according to your own catalog and your own terms.

What lands in that set on a seedstock operation:

• Lot availability and whether a bull sold private treaty before the sale.
• Expected progeny difference questions by trait — calving ease direct, birth weight, weaning weight, yearling weight, milk, docility — and the index values buyers actually shop on.
• Registration numbers and pedigree questions where the caller has the sire's number, not the lot number.
• Breeding soundness exam status, scrotal measurement, and trichomoniasis testing for buyers in states that require it.
• PAP scores for high-country buyers in Colorado and Wyoming.
• Terms: free board until a date, first-breeding-season guarantee, sight-unseen buying, delivery radius, health papers and brand inspection for out-of-state loads.
• Bidder registration for the online platform, and the two or three questions old customers ask that are really "is the same family still running this place."

flowchart TD
  A["Pull 200 real calls from last February"] --> B["Write the correct answer for each one"]
  B --> C["Run the agent against all 200"]
  C --> D{"Right lot and right numbers 95 percent of the time?"}
  D -->|No| E["Fix the catalog file it reads from"]
  E --> C
  D -->|Yes| F["Live, but message-taking only"]
  F --> G{"Owner reads every transcript for one week"}
  G -->|Any wrong figure quoted| E
  G -->|Clean week| H["Widen it: bidder registration and farm visits"]

Grading it like you grade a bull

Set the bar before you run the test, not after. A workable standard for sale-season calls: any number it quotes must be exactly right or it must not quote a number at all. Score three things separately — did it identify the right lot, did it get the figures right, and did it correctly refuse. That third one matters most. An agent that says "I don't have that in front of me, I'll have Dale call you back in an hour, what's your number and how many bulls are you looking for" is worth more than one that guesses.

Then widen its authority in stages, and only on evidence. Stage one, it answers, identifies itself as an automated assistant, takes a message with the caller's name, county and how many head. Stage two, it reads catalog facts aloud. Stage three, it registers bidder numbers and books farm visits on your calendar. Stage four does not exist: it does not take deposits and it does not talk price. That last rule is not a technical limit, it is a business rule, and you should write it down.

One legal note, because it is cheap to comply with and expensive not to. Several state AI statutes took effect on 1 January 2026 — Texas and California among them, with Colorado, New York, Utah, Nevada, Maine and Illinois all having their own rules — and the common thread for a phone agent is disclosure. Have it say it is an automated assistant in the first sentence. Buyers do not mind. They mind being fooled.

The arithmetic on a sale week

Judge this on captured calls, not on minutes saved, because minutes saved is not the business you are in.

Every one of those numbers is an illustration, and yours will differ. But you can measure the top two lines this week from your own phone bill, and if 38 percent is anywhere near right, the test set is worth building.

Where a human has to stay on the line

Price is human. Any conversation that drifts toward what you would take for a bull, or a volume deal on three head, goes to you. Guarantees are human — a buyer calling in June about a bull that failed his first breeding season is not a customer service ticket, it is a relationship, and the answer decides whether he comes back next February.

Health and shipping paperwork is human. An interstate load needs a certificate of veterinary inspection and, for sexually intact cattle eighteen months and older crossing a state line, official electronic identification under the federal rule that took effect in November 2024. An agent can tell a buyer what is required and get the vet scheduled. It should not be the last word on whether a load is legal to move.

And keep listening after you turn it on. A test set built in February gets stale by the time the fall bred-heifer sale comes around, because the questions change with the calendar. Add the new calls to the set and re-run it before every sale.

Frequently asked questions

How many test calls is enough?

Two hundred is a good target and eighty is enough to start. What matters more than volume is coverage: make sure the hard cases are in there — the caller who has the sire's registration number instead of the lot number, the one asking about a bull that already sold, the one who wants semen instead of the bull.

Who does the scoring? I don't have anybody for that.

It is a two-evening job for whoever knows the catalog cold, usually the owner or the person who wrote the lot descriptions. You are not writing software, you are writing an answer key. After the first round, re-running the same two hundred takes minutes.

What happens when the catalog changes mid-sale-season?

That is the most common cause of a wrong answer, and it is not the agent's fault — it is that a bull sold private treaty on Thursday and nobody updated the file. Decide who owns that file and when it gets updated, the same way you decide who updates the sale board. Then re-run a short version of your test set after each change.

Will buyers hate talking to a machine?

Some will, and they should get to a person fast when they ask. What buyers hate more is a phone that rings out at eleven in the morning on the week of your sale. The version that works is short, honest about what it is, accurate on facts, and quick to say it will have you call back.

If you get to the point where the test set passes and you want the phone actually answered during calving, that is the piece CallSphere builds — voice and chat agents that pick up the ranch line at any hour, answer from the catalog and terms you gave them, take a proper message with the caller's county and head count, and book the farm visit on your calendar. The evaluation work above is what makes it safe to turn on, and it stays your work either way.