1,800 New Item Setups a Year, 26 Desks: When a Food Distributor Should Run Its Own AI Instead of Paying Per Seat

"We looked at it. The quote came in per user, and we have 180 employees."

That objection is the right instinct pointed at the wrong number. A food distribution business does not have 180 people who need an AI seat. It has 26 desks and about 154 people who are selecting, loading, driving, receiving, and sanitizing — none of whom will ever open a chat window at work, and all of whom a per-seat quote quietly counts.

But the deeper problem with per-seat pricing in this trade is not the headcount. It is that the work does not come in people-shaped units. It comes in documents. Your item master has 14,000 active SKUs and churns roughly 1,800 new ones a year, and every one of them arrives as a pile: the manufacturer's spec sheet, a nutritional panel, an allergen statement, an ingredient declaration, a GTIN and case configuration with Ti-Hi and cube, shelf life, country of origin, and for anything going into a school district bid, a Child Nutrition label. That pile lands on one or two data maintenance clerks. Adding a seat does not add hours to their day.

The bid season that makes this obvious

February through May is when it hurts. School district and group purchasing bids come out with line-item specifications — "whole grain rich, 2.0 oz equivalent, no more than 480mg sodium, individually wrapped, 96 count" — and somebody has to walk each line against your item master and your suppliers' catalogs to decide what you can bid, at what pack, at what price. A mid-size distributor may work several hundred bid lines in a week during that stretch, on top of normal item setup.

The bid coordinator does it in a spreadsheet, opening PDF spec sheets one at a time. It is careful, boring, deadline-driven work, and it is the exact work that a model reads through in seconds. It is also work that scales with volume of paper, not with number of employees — which is why the licensing question suddenly matters.

Here is the clean way to think about it: pay per seat when the value comes from a specific person thinking better, and pay for capacity when the value comes from a large pile of documents being read the same way every time. Most distributors need both, and the mistake is buying one shape for both jobs.

What changed in the open tier this year

flowchart TD
  A["New item packet arrives from supplier"] --> B{"Sensitive or judgment work?"}
  B -->|"High volume, routine"| C["Own model reads spec sheet, panel, allergens"]
  B -->|"Bid strategy, pricing, customer email"| D["Licensed seat used by category manager"]
  C --> E["Draft item record: GTIN, pack, cube, shelf life"]
  D --> E
  E --> F["Data maintenance clerk reviews and posts to item master"]
  F --> G["Item live in catalog and on bid sheet"]

Until recently the free-to-download models were noticeably behind the ones you rent, and for reading a badly scanned spec sheet from a co-packer that gap showed. In 2026 it mostly closed. Moonshot AI released Kimi K3, the largest model in the world that anyone can download and run on their own hardware, and the open tier as a whole caught up enough that for reading documents and pulling fields out of them, the difference stopped being the thing that decides your project.

At the same time the rented side got roughly ten times cheaper than it was in 2025. So the choice in front of an owner is genuinely a choice now, not a compromise: rent capable models by the job, buy seats for the people who think for a living, or put a machine in your own server room and run the document work on it for what amounts to the electric bill.

The buy-versus-run arithmetic, for a house this size

Illustration with stated assumptions. A distributor with $48 million in sales, 180 employees, 26 office desks, 14,000 SKUs, and roughly 310,000 documents a year passing through the office — supplier spec sheets, order emails, signed delivery tickets, certificates of analysis, bid sheets.

At this volume renting wins by about $25,000, and it wins mostly on one line: the person. Hardware is cheap and getting cheaper. The IT time to keep a model running, patched, backed up, and answering at 3 AM when the overnight document run stalls is not cheap, and in a distribution business that IT person is already the one keeping the warehouse handhelds, the routing system, and the EDI connection alive during peak.

Run the same table at four times the document volume and it flips. When the metered line passes roughly $35,000 to $40,000 a year, owning the machine starts to pencil — and it pencils harder if you already run servers on site rather than everything hosted, because then you are adding a box, not starting a server room.

The two reasons to run your own that have nothing to do with money

First, contracts that say your data does not leave. If you do private-label or co-manufacturing work, or you distribute for a brand whose agreement restricts where their formulations and costs may be processed, "it goes to a vendor's cloud" can be a real problem. A model running on a machine in your building takes that argument off the table entirely.

Second, price certainty during your peak. Metered billing during the 100 days of summer, or during the Thanksgiving-through-New-Year run, is billing that spikes exactly when your volume spikes. An owned machine costs the same in July as in February. Some owners will pay a premium for a number they can put in a budget.

There is also a middle option most people miss. Because Kimi K3 and its peers are free to download, you do not have to be the one running it — plenty of hosting providers and ERP resellers will run an open model for you at a flat monthly rate. You escape per-seat pricing without hiring anyone. For most distributors between the two extremes, that is the honest answer.

Where per-seat is still the right buy

Do not talk yourself out of seats for the people who need them. Your category managers, your bid coordinator, your controller, and your VP of sales get real value from the best available model with their own files and email connected to it, and that is exactly what per-seat products like Claude Cowork and ChatGPT Work are built for. Twelve good seats beat twenty-six cheap ones.

And do not put anything with a customer's name on it behind your own model without a person reading it. Item master data is checkable — a wrong case cube shows up on the next pallet. A wrong allergen statement on a catalog record does not show up until somebody has a reaction, and that is not a place to save $25,000. Every allergen, ingredient, and Child Nutrition claim gets human eyes before it goes live, no matter which model read the spec sheet.

The other thing running your own does not fix: rotten source documents. A spec sheet faxed in 2011 and scanned crooked is hard for every model. The fix is asking suppliers for a current data sheet, which is a purchasing conversation, not a technology one.

Monday's step: count. Have someone tally how many documents actually pass through your office in a normal week and how many desks genuinely need a seat. Those two numbers decide this question, and almost nobody has them written down.

Frequently asked questions

How much hardware does running an open model actually take?

For reading documents and pulling fields out of them, a single well-specified server with a couple of graphics cards handles a mid-size distributor's daily volume comfortably. The largest open models want much more than that, which is why most distributors who go this route run a smaller open model on site for the bulk work and rent a frontier one for the hard thinking.

If I go self-hosted, what happens when the person who set it up leaves?

This is the question to ask before you sign anything, and it is why the people line dominates that table. If exactly one person understands it, you have built a single point of failure into your item setup process during bid season. Either get it documented and a second person trained, or pay someone else to host it.

Will an open model handle Spanish for our warehouse documents?

Yes — the open tier is strong on Spanish and this is one of the better uses for an on-site machine, since safety instructions, sanitation SOPs, and training material are high volume and never leave the building anyway.

Do I have to choose one?

No, and you shouldn't. The sensible pattern is a small number of licensed seats for the people who think for a living, plus capacity — rented or owned — for the document flood. What you want to avoid is buying 180 seats because the quote was written that way.

One adjacent note. Whichever way you go on licensing, the inbound calls do not change: order questions, will-call, delivery windows, and bid follow-ups still hit the same desk. CallSphere builds voice and chat agents that answer those phone lines and web chat around the clock, take details accurately, book time, and capture leads — a separate job from item setup, and one worth solving on its own terms.