Forty Contractor Advisors at Three Bookings a Month Each Is Where Per-Seat AI Pricing Breaks

You priced this out in 2025 and per-seat won, easily. The open models you could download were worse than the ones you rented, you would have needed a machine you did not have and a person you could not hire, and the whole exercise looked like a hobby. Fair. Two things changed since, and only one of them is the model.

The first is that the open tier closed most of the gap. Moonshot AI released Kimi K3 this year — the largest open model in the world, and open in the only sense an owner cares about: anyone can download the whole thing and run it on machines they control, no seat licence, no meter running. The second change matters more to a travel business: the number of people you are asked to buy seats for went up, and most of them barely use theirs.

Two shapes of travel business, and only one of them breaks per-seat

Almost every travel agency or tour operator I visit is one of two shapes, and per-seat answers differently for each.

Shape one is the host agency or affiliate-heavy retail agency: 40, 90, sometimes 300 independent contractor advisors, most booking three or four trips a month around a day job. They log in for wave season, go quiet in June, come back in November. Buy a seat for every one and you are buying 300 seats to serve the usage of about 45 people. That is a licensing problem, not an AI problem, and running your own model does not fix it — you would pay for a machine instead of seats.

Shape two is the operator with a small team pushing enormous volume: eleven people in reservations and product for a receptive or packaged-tour operation, producing final documents for 4,200 departures a season, rooming lists, manifests, guide briefings in three languages, and thousands of replies in the Viator and GetYourGuide message centers. Eleven seats is nothing; the work per seat is the story.

Per-seat pricing stops making sense when your headcount is large and your usage per head is small, or when your headcount is small and your usage per head is enormous — and only the second case is an argument for running your own model.

What Kimi K3 actually means for the line on your budget

Practically: for the first time, the thing you can run yourself is good enough that the decision is about cost and control rather than quality. K3 is enormous — large enough that it only wakes the portion of itself each job needs, which is how something that size runs at all — and it is downloadable — put it on machines you rent by the month, or a rack in your own office.

What it does not mean is "free." The model costs nothing; everything around it costs something, and in this sector that something is mostly people, not hardware. That is where most business cases quietly fall apart.

flowchart TD
  A["Count your seats, then count real jobs per seat per week"] --> B{"Many advisors barely touching it?"}
  B -->|Yes| C["Stay per-seat, and buy far fewer seats than you were quoted"]
  B -->|No| D{"Is a small team pushing thousands of documents a week?"}
  D -->|No| C
  D -->|Yes| E{"Must passport scans and card forms stay on your own machines?"}
  E -->|No| F["Pay per use on a hosted model, review the bill monthly"]
  E -->|Yes| G["Rent one machine, run an open model, name one person who owns it"]

The job nobody puts in the business case

If you run your own, you are not buying software. You are taking on a piece of operations, and it needs an owner with a name.

That person keeps the machine running, restarts it when it wedges, updates it, and — the part owners underestimate — is reachable on the Tuesday of wave season when it stops working and eight advisors have proposals due that evening. Nobody on a travel org chart has that job. Your operations manager is chasing a rooming list; your groups coordinator is on hold with a cruise line; your most technical person is the advisor who is quick with Sabre Red 360 shortcuts, which is a different skill.

Realistically you hire a contractor for four to eight hours a week, or a small managed provider. Budget it honestly — pretend otherwise and the machine becomes the owner's problem at 10 p.m., a second job taken on to save a licence fee.

Twelve months of arithmetic, both ways

Illustrative assumptions, host-agency shape: 40 contractor advisors, seats at $30 per person per month, a small managed setup for the alternative.

It is not close, and it does not get close by adding advisors — you would need roughly 160 people before the licence line caught the running line, and by then you need a bigger machine. The move for shape one is not to run your own. It is to stop buying 40 seats. Buy 12 for the full-time producers and review the list after wave season.

Now shape two, on volume rather than seats. Suppose the eleven-person operator generates about 90,000 pieces of work a season — documents, replies, translations, summaries — at roughly four cents each on hosted per-use pricing. That is $3,600, plus eleven seats for interactive work, say $3,960. Under $8,000 a year against $58,440 to run your own. To break even on cost alone you would need well over a million pieces of work a year, roughly 3,000 a day, seven days a week. Very few US travel businesses are that shape.

The case that has nothing to do with money

One argument for running your own survives the arithmetic, and in this trade it is not small: what is in the documents.

Your systems hold passport scans, dates of birth, the credit card authorization forms clients still fax and email you, dietary and medical notes, mobility requirements, and in group work entire school rosters of minors. You carry a merchant account the card networks treat as high risk, and your chargeback ratio is not somewhere you want surprises. Corporate and government clients — a school district, a university study-abroad office — will ask you in writing where that data goes and who else can see it.

If the honest answer must be "it never leaves machines we control," the cost comparison above is the wrong one, because an option does not qualify. That is a legitimate reason to run your own. Wanting a lower bill, at these volumes, is not.

Why almost all of you should not do this

Blunt, because the trade press is not. If you are a retail agency of six, or an operator running 40 departures a season, running your own model is a bad idea. Not risky — bad. You will spend more, have a worse tool, and introduce a failure point that lands on the owner during the four weeks you can least afford it.

Even where it does pencil out, keep three things off it. Anything where being subtly wrong is expensive — fare rules, penalty schedules, supplier contracts, group air deposit deadlines — runs on the best model available regardless of who hosts it. Anything guest-facing under your name gets a human read before it sends. And anything touching money movement stays with a person.

There is a middle path most owners never consider: run your own for the high-volume grind — document generation, translations of departure notes, first-pass review replies — and keep paying per use for the small amount of hard, high-stakes work. Data control where it matters, good judgement where it matters, without a full migration.

The honest first step

Before you evaluate anything, count. Pull your seat list and, for the last 90 days, count how many of those people used the tool in a week they were not in training. Then count the pieces of work your busiest team produced. Two numbers, one afternoon of your bookkeeper's time — and that afternoon changes the decision for about half the agencies sure they needed their own machine, usually by revealing they pay for three times the seats they use.

Frequently asked questions

Our host agency provides the AI tools. Does any of this affect us?

It affects what you push for. If your host bundles a seat into your monthly fee, ask whether you get the good model or the cheap default, and whether client data is handled under an agreement you have read. Ask where the documents go and what they hold on their own machines. Your name is on the Seller of Travel registration, not theirs.

Can we run something smaller than Kimi K3 and get most of the benefit?

For document generation, translation of departure notes and first-draft replies, yes — and the running cost drops sharply, because a smaller open model does not need the same machine. K3 matters mostly as proof of how far the open tier has come. The practical choice is usually a mid-sized open model on one rented machine, not the biggest in the world.

What actually happens if the machine goes down during wave season?

Your people go back to doing it by hand, which is survivable, and the owner spends the evening on the phone with a contractor, which is the real cost. Agree the response time in writing before you start, and keep a paid per-use account as a fallback so work can move over the same afternoon. That fallback is cheap insurance and most first-timers skip it.

Does running our own help with the EU rules if we sell to European travellers?

It helps with where data sits; it does not settle your obligations. The EU AI Act's high-risk and transparency duties carry a 2 August 2026 compliance date and can reach US companies whose systems affect users in Europe, and Texas TRAIGA and California SB 53 have been in force since 1 January 2026. Hosting your own changes none of that. Talk to counsel about which uses are in scope.

Whichever way you land, the one workload that rarely justifies its own machine is the phone. Inbound calls and website chat are bursty, they happen at night, and they must answer before the caller gives up. CallSphere builds AI voice and chat agents that answer the line and web chat 24/7, book appointments and capture inquiries, priced by conversation rather than per seat — the same argument this post has been making, applied to the one channel you cannot leave ringing.