42 Employees, 9 Keyboards: Where Per-Seat AI Pricing Breaks Down for a Propane Marketer

The objection: "we looked at the free ones and they were toys"

You did, in 2024, probably at somebody's suggestion at a state propane association meeting, and you were right. The models you could download and run yourself were noticeably worse than the ones you rented by the seat, and nobody in a propane office had time to babysit a machine that gave a worse answer.

That gap closed a lot in 2026. Moonshot AI released Kimi K3, the largest open model in the world — a sparse mixture-of-experts design, which in owner's English means it is enormous but only wakes up the part it needs for each question, so it can be run without a data center's worth of hardware. It is not alone; the whole open tier moved up. The choice in front of you is no longer "good rented model versus bad free one." It is buy-per-seat versus run-your-own, and for a propane marketer the honest answer depends on a shape that is peculiar to this trade.

The definition worth keeping: an open-weight model is one you can download and run on your own hardware, in your own building, with no per-person license — you pay for the machine, the power and the person who looks after it instead of paying for seats.

The headcount shape: 42 employees, 9 keyboards

Take a single-plant marketer doing about 4.1 million gallons a year. Forty-two people on the payroll in February. Fourteen bobtail drivers with a CDL and hazmat and tanker endorsements. Six service techs with CETP cards. Two cylinder handlers on the fill house. A delivery manager, a plant manager, a safety director who is also the delivery manager two days a week. Five CSRs, two dispatchers, one credit and collections person, a controller, an office manager, two owners.

Count the people who sit in front of a keyboard for most of the day and you get about nine. Everybody else has a wheel or a wrench in their hands. That is the first thing that makes per-seat AI pricing look different here than it does at an insurance agency or a title company, where the seat count and the headcount are nearly the same number.

Nine seats is cheap. If seat pricing were the only question, you would buy nine and stop reading. But the work you actually want an AI to do in a propane operation does not live at those nine desks. It lives in the paper the other thirty-three people generate.

The volume shape: 180,000 delivery tickets nobody will pay per seat to read

In one heating season that same marketer produces roughly 180,000 delivery tickets, tens of thousands of call notes and voicemails, twelve years of service history in ADD Systems Energy Force or Cargas Energy, thousands of tank monitor readings a day from Otodata or Tank Utility, and a filing cabinet of leak check forms, tank set records, hose test records and cylinder requalification logs.

The jobs worth doing against that pile are bulk jobs. Score every will-call account for how likely it is to run out before its next planned delivery. Read twelve years of service notes on the 1,900 accounts with a wall furnace and find the ones with three regulator replacements. Read every delivery ticket where the driver wrote something in the notes field. Compare quoted prices to delivered prices across the summer-fill campaign. None of those are things one of your nine desk people sits down and does. They are things you run once, on everything, overnight.

flowchart TD
  A["One season of records: 180,000 delivery tickets, 12 years of service history"] --> B{"Desk work or bulk work?"}
  B -->|"Desk work, 9 people"| C["Per-seat plan such as Claude Cowork or ChatGPT Work"]
  B -->|"Bulk work, every ticket and every note"| D["Pay per use, or run an open model on your own box"]
  C --> E["Fixed monthly seat bill, flat in July and January"]
  D --> F["Cost rises and falls with gallons delivered"]
  E --> G{"Which line was bigger at the end of the season?"}
  F --> G
  G --> H["Re-decide once a year, in April, with the numbers in front of you"]

The arithmetic: seats, per-use, and a box in the closet

Illustration only, and the hardware figure in particular will be out of date within a year. Assume the 42-person marketer above, nine desk seats, one heating season of bulk work.

Read the third row again. The machine is not the expensive part; the six hours a month is. On these assumptions you would need somewhere north of $17,000 a year of rented use before owning the box pays — roughly twelve times the bulk volume in the second row. A single-plant marketer does not have that volume. A three-state group with 300 employees, eleven bulk plants and a shared back office might.

So the honest recommendation for most readers of this column: rent the seats, rent the bulk work per use, and revisit it every April when the season is closed and you can see a full year of what you actually spent. Put the number on the same page as your truck lease renewals so it gets looked at.

What "running your own" costs in people, not dollars

Nearly every propane marketer I know has one IT person, and it is usually a contractor who also keeps the server in the closet running Energy Force, sets up the bobtail tablets, and rebuilds the dispatcher's machine when it dies. Adding a model to that person's plate is not a purchase; it is a dependency.

Two questions decide it. First: when the box hangs at 5:10 a.m. on the coldest morning of the year and the will-call scoring did not run, who gets in the truck? Second: when that contractor retires or takes a full-time job, does anybody else in the building know how it was set up? I have watched a marketer lose a perfectly good system to nothing more dramatic than a change of contractor.

There is one genuine advantage to owning, and it is not cost. It is that the data never leaves the building. Recorded emergency calls, customer addresses, propane usage down to the gallon by household — some owners are not comfortable sending that anywhere, and some co-op boards will not allow it. That is a legitimate reason to own, and it is a better reason than the money.

Where this still needs a human, and where it should stay rented

A model of any kind, open or rented, does not know your yard, your routes or your drivers. Every one of those bulk jobs produces a list, and a list is a proposal, not a decision. The will-call run-out scores need your delivery manager to look at them, because he knows which of those addresses cannot take a bobtail after an ice storm and which customer will let it go to zero on purpose because he heats with wood until Christmas.

Two things should never be moved to a machine in the closet on a cost argument. Anything with a compliance signature on it — the leak check form, the bulk plant inspection, the cargo tank test record — stays with the certified person whose name is on it. And anything that touches an emergency response stays on the most reliable system you own, not the newest one. If you would not put your on-call rotation on the box, do not put your odor-call handling on it either.

The starting move on Monday is small: pick one bulk job, the will-call run-out list, and rent it. Run it against last season's data where you already know who ran out. If the list would have caught them, you have learned something real about your own operation for a couple hundred dollars, and you have not signed anything.

Frequently asked questions

Can I run one of these open models on the server already sitting in my office closet?

Not the big ones. The machine that runs your delivery software is built for a database, not for this kind of work, and putting both on it is how you end up with dispatch down on a Monday in January. If you go this route it is a separate box, on its own power, with its own backup.

If the open ones are free, why is anyone still paying per seat?

Because the seat price buys the part around the model — the interface your CSR actually uses, the connection to your files, the log of who asked what, and somebody to call when it breaks. Free covers the engine, not the truck.

What does the volume shape look like for a smaller marketer, say three bobtails?

Even more lopsided toward renting. Three or four desk seats, maybe 25,000 tickets a season. The bulk work is real but small, and there is no version of the arithmetic where buying a box makes sense. Spend the attention on the phones and the will-call list instead.

Does the EU AI Act deadline on 2 August 2026 touch a US propane distributor?

Almost certainly not, unless you have European customers or an affiliate over there. The rules that will touch you sooner are the state ones — Texas TRAIGA and California SB 53 both took effect on 1 January 2026, and several other states have their own. If you operate in those states, ask your attorney what your disclosure obligations are before you point anything at customers.

One place where renting is clearly the right call

The phone line. A customer calling at 6:40 a.m. in a cold snap does not care where the answer is hosted; they care that somebody picked up. CallSphere builds AI voice and chat agents that answer the business line and the website chat, take will-call orders and delivery questions, book appointments and pass anything urgent to a live person. It is a rented, per-use service by design, because the volume triples in January and falls off a cliff in May — which is exactly the shape you do not want to buy a box for.