You Hire 25 Seasonal Associates in October and Drop to Eight in February. Per-Seat AI Pricing Was Never Built for That.

What is a software seat worth in a business where a third of your staff is hired in October and gone by Valentine's Day?

That is not a rhetorical question. It is the arithmetic problem sitting under every AI tool a specialty or apparel retailer has been sold since 2024. The pricing was built for offices — a fixed number of desks, all year, each one a named person with a company email address. Your store does not look like that. You run eight people year-round, you hire fifteen to twenty-five seasonal associates for the stretch from the first week of October through the post-holiday return wave, you carry an alterations tailor two days a week, and by mid-February you are back to eight.

The seat count you are actually paying for

Take a three-store specialty chain doing $4.2 million. Year-round: two owners, a buyer, an e-commerce coordinator, three store managers, a bookkeeper. Seasonal: twenty-five associates and two extra stockroom hands from October to January. If a tool is $30 per user per month and you buy seats for everyone who touches it, that is $240 a month for nine months of the year and $1,050 a month for four — except nobody prorates cleanly, seasonal hires arrive in waves, and the person administering it is a store manager on the floor during the busiest quarter of the year. So what actually happens is one of two things: you buy seats for eight people and the seasonal staff never touch the tool, or you buy for everyone and forget to remove twenty-five seats in February.

Meanwhile the real AI workload in your business is not sitting on those desks at all.

Where the volume actually is: 11,000 new styles a year

Independent specialty retail generates an enormous amount of repetitive, low-stakes writing and sorting work, and almost none of it belongs to a person with a seat:

• Product copy. Between pre-book seasons, market pickups, and re-orders in new colorways, a store with 24 brands puts thousands of new style-color-size records into Shopify or Lightspeed every year. Brand-supplied copy is either missing, identical to every other retailer's (which hurts you in search), or written for a different country.
• Alt text and image tagging for every one of those images, which affects both search and your obligations on site accessibility.
• Size chart normalization. A French 38, an Italian 42, a US 6 and a brand that runs half a size small are four different facts about the same shoe, and the mismatch is the single biggest driver of online returns in apparel.
• Marketplace and feed tagging for Google Shopping, Faire, Instagram, and whatever else you list on — each with its own required category and attribute fields.
• Review and fit-comment summarizing, so the product page can say "customers report this runs long in the sleeve" without someone reading 90 reviews.

That is high-volume, repetitive, and forgiving — if a product description is slightly awkward, you fix it and nobody gets hurt. It is exactly the wrong work to buy per-seat, and exactly the right work to run on a model you host yourself.

What changed: the free tier got big enough to matter

An open-weight model is one you can download and run on a machine you control, paying for the machine instead of paying a monthly fee per employee. That option existed in 2024, but the free models were noticeably worse and you felt it in the output. In 2026 that gap largely closed. Moonshot AI's Kimi K3 is now the largest open model in the world, and the open tier as a whole moved close enough to the paid frontier that for writing product copy, tagging attributes and summarizing reviews, most owners cannot tell the difference in a blind read.

The cost picture underneath it also moved. Capable paid models are down roughly tenfold from 2025, to the point where running a model on every single style you receive costs less than what you spend on hangers. Running an open model on your own machine is roughly 90% cheaper again for this kind of high-volume, repetitive work. The decision is no longer quality; it is who does the work of keeping it running.

flowchart TD
  A["11,000 new style-color records a year"] --> B{"High volume and low stakes?"}
  B -->|Yes| C["Own model on the back-office machine"]
  B -->|No| D["Paid seats: buyer and e-commerce lead only"]
  C --> E["Product copy, alt text, size-chart cleanup, feed tagging"]
  D --> F["Assortment planning, vendor terms, promo copy"]
  E --> G["Merchant reviews a sample of 40 before publish"]
  F --> G
  G --> H["Live on the site and the Faire listing"]

What running your own actually costs — in people, not machines

This is where most write-ups lie to you. The hardware is the easy part; you can rent time on somebody else's machine running an open model and never buy anything, or put a single well-specified box in the back office. The real cost is that somebody has to own it.

Be realistic about who that is in a retail business. It is not your store managers. It is almost certainly your e-commerce coordinator, and the honest budget is three to four hours a week of that person's time plus an outside contractor on a small monthly retainer for the two or three days a year something breaks. If you do not have an e-commerce coordinator — if your website is run by the owner at 10 p.m. — then do not do this. Buy seats. That is the honest answer and it applies to most single-store retailers.

The second hidden cost is the setup: writing down what "good" looks like for your product copy. Your voice, your length, the fact that you never say "elevated" and always name the fabric. That is a day of the buyer's and the e-commerce coordinator's time, and it is the same day whether you buy or build.

The math on 11,000 styles

All figures illustrative — substitute your own headcount and style count.

Running your own totals about $10,442 against $5,880 for buying seats for everyone. On those numbers, buying loses only if the volume is much bigger or the coordinator time is already being spent. The break-even is roughly this: if the repetitive work would otherwise consume more than eight hours a week of paid staff time, or if your style count is north of about 20,000 records a year, running your own starts to win. Below that, per-seat is cheaper and considerably less trouble — and any consultant who tells you otherwise for a single boutique is selling a project.

What you should still buy per seat

Two categories, and neither is close.

The first is anything a customer talks to. If it answers your phone or your website chat, it is the front door of your store, and you want a vendor whose job is keeping it up on Small Business Saturday. Nobody wants their e-commerce coordinator paged at 11 a.m. on Black Friday because the chat widget stopped.

The second is the judgement work: open-to-buy planning, reading vendor terms, writing the email that goes to nine thousand customers. Buy the best two seats you can for the buyer and the owner and stop counting pennies there. The saving is in the eleven thousand product descriptions, not in the four decisions a week that determine your season.

Frequently asked questions

Do I need a server in the back room?

No. You can rent time on a machine running an open model and never own hardware, which is what most retailers doing this actually do. A physical box in the back office only makes sense if your volume is very high or you have a specific reason to keep everything in the building.

Will customers be able to tell the product copy was written by AI?

They will if you publish it unread. The stores where this works have a merchant reviewing a sample — forty descriptions out of a drop, not all four hundred — and a written style rule the model follows. Where it goes wrong is the fabric detail: it will happily invent "brushed cotton twill" if your brand data does not say what the fabric is.

What happens when my seasonal staff need it in November?

That is the point of running your own — there is no per-person charge, so a seasonal associate looking up "does this brand run small" costs you nothing extra. Give them access on their first shift and remove it with their point-of-sale login when the season ends.

Is the open model good enough for customer-facing answers?

For catalog work, yes. For anything where a wrong answer costs you a customer or a return — fit advice, return policy, whether an item is in stock — use a tool built for that job with your own approved information behind it, and keep it on a paid vendor with someone to call.

Where the phone fits into this

The seat-count problem shows up hardest on the phone. In December every line rings while every associate is on the floor, and in February the same line rings twice an hour with nobody free to answer it either. CallSphere builds AI voice and chat agents that answer the store line and the website chat 24/7, book fittings and appointments, and capture the name, number and reason of every caller — priced against your call volume rather than your seasonal headcount, which is the shape retail actually has.