Per-Seat AI Pricing Assumes Everyone Has a Desk. A Four-Unit Restaurant Group Has 240 W-2s and Nine Logins.

Nine. That is how many people in a four-restaurant group sit down at a keyboard on a normal Tuesday: the owner, the controller, the director of operations, four general managers, the executive chef, and the catering sales manager. The other 231 names on the payroll — servers, bartenders, hosts, bussers, food runners, line cooks, prep, dish — touch a screen all night, but it is a Toast terminal or a kitchen display screen, not a laptop.

Now look at how business AI is sold. It is sold by the seat, at roughly thirty dollars per person per month. Buy a seat for everybody on the payroll and a 240-employee group is looking at $7,200 a month, $86,400 a year. Against four rooms doing $14 million at a four to five percent net, that is most of one restaurant's annual profit spent on logins for people who clock in at a terminal and never open an inbox.

The volume work in a restaurant has nothing to do with headcount

Buy seats for only the nine and you have solved nothing, because the work that actually eats the week is not nine people typing. It is 380 vendor invoices a month across Sysco, a produce house, a fish purveyor, three beer and wine distributors, linen, and the grease hauler. It is ninety Google and Yelp reviews that deserve a reply. It is recipe costing that has to be redone every time butter or chicken breast moves five percent. It is prep sheets and line-check forms that have to exist in English and Spanish because half the back of house reads one and half reads the other.

None of that scales with how many people you employ. It scales with how many invoices arrive and how many covers you turn. That is the whole argument behind the change that happened this year: running your own model means the work is priced by the job — per invoice read, per review drafted, per prep sheet translated — instead of by the head, so the bill tracks your paperwork volume rather than your payroll count.

What actually changed in the open tier this year

Through 2024 and 2025, "open" models — the kind you can download and run on a machine you control instead of renting a login — were the budget option. They were fine for tidying up a paragraph and unreliable for anything you would put in front of your accountant. That gap largely closed over the last year. Moonshot AI released Kimi K3, now the largest open model in the world, and the open tier as a whole caught up to the point where the honest question stopped being "is it good enough" and became "buy per seat, or run your own."

Two practical notes before anyone gets excited. First, you are not putting the largest open model in the office closet next to the router. The big ones get rented by the hour on somebody else's machines — still priced per job, not per person, which is the point. The smaller open models are the ones that will genuinely run on a single machine in your commissary office, and for reading invoices and drafting review replies they are more than adequate. Second, frontier AI pricing fell roughly tenfold from 2025, so the per-seat products got cheaper too. This is not a story about open being free. It is a story about which meter your bill is attached to.

Which meter fits which job

flowchart TD
  A["Work lands on someone's desk"] --> B{"Does a named manager
have to think about it?"}
  B -->|"Yes: BEO for a 40-top,
fall menu rewrite"| C["Per-seat login:
GM, chef, controller"]
  B -->|"No: 380 invoices,
90 review replies"| D["Per-job run
on an open model"]
  C --> E["9 seats billed monthly"]
  D --> F["Billed per invoice
and per document"]
  E --> G["Same G&A line
on the P&L"]
  F --> G

The split is not ideological. Anything where a human is going to sit and argue with the machine — writing the banquet event order for a forty-person rehearsal dinner, reworking the fall menu around a bad tomato season, drafting a difficult conversation with a sous chef — wants a seat, and the nine people who do that work should have one. Anything that is the same shape 380 times a month wants a per-job arrangement.

The arithmetic, with the assumptions on the table

Here is the same volume work priced three ways. Every figure is an illustration, not a quote — run yours with your own invoice count off last month's file.

At four rooms the three columns are close enough that nobody should switch on price alone. At eight rooms they are not close at all, and the reason is simple. The first two bills double when you double the group. The third one barely moves, because the eighth restaurant does not double the number of distinct invoice layouts you handle — it just sends more of the same ones. Full-service groups add hourly headcount far faster than they add new kinds of paperwork. That is the crossover, and it usually shows up somewhere between the fifth and seventh room.

What running your own actually costs in people

This is where the pitch usually goes quiet, so let us be blunt. Nobody in a restaurant group is hiring a machine-learning person. What you are actually buying is time from the same outfit that shows up when the terminal at the service bar dies — your point-of-sale and IT vendor — or a fractional consultant who does this for a handful of hospitality clients. Budget twenty to thirty hours to get it standing up and connected to Restaurant365 or MarginEdge, then two to four hours a month after that.

The part that decides whether it survives is internal, and it is not technical. One person inside the company has to own it, and in a full-service group that is almost always the controller or the director of operations — someone who already looks at every invoice and would notice within a day if the thing started coding produce to the beer account. If no name is attached to it, it dies in ninety days, exactly the way the last tablet-based inventory count died.

Where this still needs a human

An open model reading a broadline invoice will be excellent on the line items and merely okay on the messy edges: the handwritten credit at the bottom of the delivery slip, the short case the driver initialed at 6:15 in the morning, the case of romaine you refused at the back door that will surface as a credit memo eleven days later. Someone still has to receive product against the invoice at the dock, temp the proteins, and sign. No software has ever done that part.

It also should not touch anything with a signature or a license attached. Liquor license renewal, your response to a health department inspection, ServSafe manager certification dates, allergen statements — those stay with a person, because the downside is not a mis-coded invoice, it is a suspended license or a guest in an ambulance. Your first Monday step should be embarrassingly small: pull twenty invoices out of last month's file, hand the same twenty to your current per-seat tool and to whoever would run an open model for you, and compare both against what your bookkeeper actually entered. Two hours of work, and it settles the argument with evidence instead of a slide deck.

Frequently asked questions

If I run my own, does my invoice and sales data leave the building?

That is the real reason most multi-unit operators look at this. With a per-seat plan, your invoices, recipe costs and sales mix sit on a vendor's servers under whatever their business terms say. With a small open model on a machine in your office, the file never leaves. With a big open model rented by the hour, it does leave, but you are renting machine time rather than handing a software company your P&L. Ask any vendor in writing whether your data is used to improve their product, and get the answer before you sign the annual.

We turn over most of our hourly staff every year. Does that change the math?

It changes it enormously, and it is why per-seat pricing fits restaurants poorly. Annual plans are sold on a seat count you commit to in January. In a full-service room, the person that seat was assigned to in January may be gone by March, and now license transfers are a side job for a GM who is already short two line cooks. A per-job arrangement has no seat to transfer and no January commitment to guess at.

Can this read invoices from our small local purveyors, not just the broadliners?

Usually yes, and that is the genuine improvement over two years ago. Sysco and US Foods invoices have been machine-readable for a long time. The ones that used to break everything are the fish purveyor's emailed PDF, the bread route's carbon-copy pad, and the farm that writes the weight in pen next to the case count. Current models handle those far better, but spot-check that category weekly for the first quarter rather than trusting it outright.

Do I have to pick one? Can I run both?

Most groups that get this right run both, deliberately. Nine seats for the people who think for a living, and a per-job arrangement underneath for the repetitive volume. The mistake is buying thirty-four seats to solve a problem that was really 380 invoices wearing a costume.

Where the phone sits in all this

One thing worth separating out: none of the above answers the line at the host stand, which in most full-service rooms is the loudest unstaffed job in the building. CallSphere builds AI voice and chat agents that pick up the restaurant's phone and web chat, take reservation and private-dining inquiries, and capture the caller's details around the clock — billed per call rather than per seat, which is the same shape of decision this whole piece is about. It does not read your invoices, and anyone telling you one product does both is selling.