Per-Seat AI Billing Punishes a Mover's Four-Month Peak - What Running Kimi K3 in the Warehouse Office Actually Costs

Twenty-nine seats. That is what one Midwestern agent I spoke with this spring is paying for across two AI subscriptions — coordinators, dispatch, three sales consultants, the claims desk, billing, the warehouse manager, the owner and his son. Thirty dollars a seat, two products, twelve months, no way out until renewal. Call it $20,880 a year.

In February, eleven of those twenty-nine seats were used more than twice in the month. The company had eleven office people in February. It has twenty-nine in July. The software does not care, because software is sold by the seat and a moving company is not staffed by the seat — it is staffed by the calendar.

Count the seats you paid for in January

Every mover reading this knows the shape. Roughly seven in ten household goods moves in the United States happen between Memorial Day and Labor Day. Leases turn on the first, so the last four days of every month in that window are triple-booked. You put on packers and helpers by the dozen, you promote two crew leaders to temporary coordinators, and you carry the whole office roster through a January where you are running twelve loads a week instead of seventy.

Per-seat licensing bills you for the July roster all twelve months. Worse, seats are the wrong meter entirely for what a moving company actually uses AI on. Your cost driver is not people typing. It is documents: descriptive inventories with rider codes, bills of lading, weight tickets, crew damage photos, packing lists, third-party lead forms coming in from Angi and Thumbtack and your own site, recorded sales calls. In July a busy ten-truck agent generates several hundred scanned pages and a few thousand photographs a week. None of that volume is measured in seats.

The meter is documents, not headcount

Here is the plain definition, since the industry press has made a mess of it: an open-weight model is one whose maker publishes the finished model itself, so you can run it on a machine you own or rent, instead of paying a subscription for each employee who is allowed to use somebody else's copy.

That distinction did not matter much when the open ones were noticeably worse. In 2026 it does. Moonshot AI released Kimi K3 with 2.8 trillion internal dials — the largest open model anyone has published — built so only a slice of it runs on any one request, which is why something that big can still be affordable to operate. More to the point for you, the whole open tier moved up behind it. The smaller open models now handle the jobs a mover actually needs done at a quality that would have required a paid subscription eighteen months ago.

The economics moved too. Frontier AI in the cloud is down roughly ten times from 2025 prices. And for high-volume repetitive work — reading ten thousand inventory pages, not composing ten thoughtful emails — running it on your own hardware costs roughly 90% less than sending every page to somebody's cloud.

flowchart TD
  A["Your July volume: 2,400 scanned pages + 9,000 crew photos"] --> B{"Is the work people-shaped or document-shaped?"}
  B -->|People-shaped: sales calls, claim letters| C["Buy seats for 6 to 9 office staff"]
  C --> D["Cancel nothing in January, seats sit idle"]
  B -->|Document-shaped: inventories, BOLs, photos| E["Run an open model on one box in the office"]
  E --> F["Flat monthly cost, volume triples in July for free"]
  F --> G["Military and corporate files never leave the building"]

Two ways to buy the same July, side by side

Illustration, not a quote — plug in your own numbers. Assume a mover with 29 office seats at peak and 11 in the trough, doing 3,100 jobs a year, of which 640 are interstate. Assume the document work you want done is inventory and bill-of-lading reading, crew photo sorting into claim files, and first-draft claim correspondence.

Read that table honestly: at 29 seats, running your own loses. It loses because of the last two lines, not the first two. The crossover in this trade is not a headcount number, it is a "do you already employ someone who patches servers" number. If you have a systems person on payroll — most 40-plus-truck agents and every van line agent running its own warehouse management do — the $21,600 line collapses to maybe $6,000 of that person's time, and running your own wins by roughly $8,000 a year and gets better every July as volume grows and the subscription bill does not shrink.

What running your own actually costs in people

Nobody sells you the person. The bill you do not see on the quote:

• Somebody has to own it. Not "the owner's nephew is good with computers" — somebody whose name is on it when it stops answering at 6:30 a.m. on a Saturday in June.
• Somebody has to update it. Open models improve every few months; the one you install in August is behind by Christmas, and staying current is a scheduled half-day, not an accident.
• Somebody has to keep the rules current. When you file a new intrastate tariff with the state, or the van line changes its inventory rider codes, the instructions you wrote have to change with them.
• Somebody has to say no. Left alone, staff will point it at everything, including things it should not touch, and you will find out in a claim file.

That is between a quarter and a half of one office person's year at a mid-size agent. If you cannot name the person right now, buy seats. There is no shame in that answer and it is the right answer for most three-to-fifteen-truck operations.

The military and corporate files that turn this from a price argument

There is a second reason movers self-host that has nothing to do with money. Look at what sits in your files during a normal week: PCS orders with service members' details on them under the government household goods program, corporate relocation files handled under an agreement with a relocation management company that specifies where the client's employee data may go, high-value inventories listing exactly what is in a house and exactly when that house will be empty.

Some corporate relocation contracts and some government work carry handling terms that make "we uploaded it to a consumer chat product" an uncomfortable sentence to say out loud in an audit. Running the model on a box in your own building makes the answer simple: the file never left. For agents chasing corporate accounts through a relocation management company's yearly bid, that is not a technical detail, it is a line you can answer cleanly on the questionnaire.

When buying seats is still the right call

Buy seats if your AI use is mostly people-shaped — sales consultants drafting follow-ups, the claims coordinator writing settlement letters, you rewriting the crew handbook. Those are dozens of requests a day, not thousands, and the subscription is cheap relative to a server that idles.

Buy seats if you are under about fifteen office staff year-round. Buy seats if your busiest month is only twice your slowest, rather than three times. Buy seats if the honest answer to "who patches it" is silence.

And whichever way you go, negotiate the seat count against your calendar. Ask for a peak-season tier: eleven seats November through April, twenty-nine May through October. Vendors selling into seasonal trades will do this more often than owners assume, because they would rather have your July than lose your year. The mover I opened with got that concession at renewal by asking once.

Frequently asked questions

Can I actually run Kimi K3 in my warehouse office?

Not the full-size one — that needs serious hardware and it is not what a moving company should be aiming at. The reason K3 matters to you is that it pulled the whole open tier upward. The realistic move is a smaller open model on one server, doing document reading and drafting, with the hard occasional job sent to a paid cloud product. Mixed is normal and it is fine.

Is my customer data safer on my own server?

It is more clearly yours, which is a different thing from safer. A server in your warehouse with no patching and one shared password is worse than a reputable cloud product. Self-hosting buys you a clean answer about where the data lives; it does not buy you security, and if you cannot commit to the patching, it actively costs you some.

What happens to my subscription seats when I lay off seasonal staff in October?

Read the contract before you sign it, because most annual agreements do not let you reduce seats mid-term. This is the single most common way seasonal businesses overpay. Ask for mid-term seat reduction rights, or for the tiered arrangement described above, and put it in writing at renewal rather than at signing time.

Does this change what I pay for my move management software?

No. Your move management system and your tariff are separate purchases and separate decisions. This argument is only about the AI layer sitting next to them. Do not let a software vendor bundle the two together into one number you cannot pull apart.

Where we fit, briefly: whichever way you land on the buy-versus-run question, the phone still rings all four months and the seat count never solved that. CallSphere builds AI voice and chat agents that answer inbound calls and website chat around the clock, book surveys, and capture leads with the details attached — priced against the calls handled rather than the number of people who happen to be on your roster in July.