At 240 Pickers and Three Minutes a Shift, Per-Seat AI Pricing Stops Adding Up. Here Is the Break-Even.

How many seats does a fulfillment center actually have?

Ask a software salesperson and watch them do the math they want to do. You have 240 hourly associates across two shifts and 14 salaried people. Per-seat AI pricing at $30 a head makes your bill $7,620 a month. Ninety-one thousand a year for a building where the median user opens the thing three minutes a shift, to ask what the pallet pattern is for a retail account.

That is the whole argument here. Warehousing has a headcount shape per-seat software was never designed for: a few heavy users and a great many people who need one correct answer, once, fast, in the middle of a pick path. Every seat-priced tool in this industry gets bought for the 14 and quietly never rolled out to the 240, which is where the value was.

Something changed in 2026 that makes the second option real, and it is worth doing the arithmetic properly rather than by instinct.

The question the floor asks sixty times a shift

Walk a pick module at 10 a.m. during peak and count the interruptions. A temp from the staffing agency, second day on the floor, has a client's fragile kit and does not know whether the insert goes under or over the tissue. A picker hits a location where the lot code on the shelf does not match the pick screen. Someone needs to know whether this client's returns get inspected before putaway or after.

Every one of those goes to the same place: the shift supervisor, or the one lead who has been there six years and knows all forty client SOPs by heart. He answers, walks back, gets stopped again. During peak, when half your headcount is seasonal, that lead becomes the slowest part of your building — and when he takes a Friday off, the whole floor's answer quality drops.

The SOP binder exists. It is on the shared drive, it is 40 folders deep, half of it is a scan of a document the client sent in 2021, and no picker with a handheld scanner is going to open it mid-path. What the floor needs is an answer desk that speaks the trade's language, works from your own client SOPs, and does not charge you thirty dollars a month for a person who asks it four questions a week.

What changed: the free-to-download tier got good enough for the floor

In 2024 the models you could download and run yourself were noticeably worse than the ones you rented. That gap has largely closed. Moonshot AI released Kimi K3 — the largest model in the world that anyone can download and run on their own machines — and the open tier broadly caught up on the kind of work a warehouse actually needs: read this SOP, answer in plain language, do it in Spanish and English, say "I don't know" when the answer is not in the document.

That last part matters more than benchmark scores. You are not asking it to do original thinking. You are asking it to read your own approved documents and quote them back correctly to a picker on a headset. So the decision is genuinely a purchasing one, not a technical one: rent per seat, or run your own on a machine in your building and pay nothing per person. Both are legitimate. The right answer depends on how many people would really use it, and how heavily.

flowchart TD
  A["Associate asks a question at the pick face"] --> B{"Is it covered in this client's SOP?"}
  B -->|Yes| C["Answer read back with the SOP section named"]
  B -->|No| D{"Is it safety, hazmat or lockout?"}
  D -->|Yes| E["Stop. Route to the safety coordinator"]
  D -->|No| F{"Is it a client-specific exception?"}
  F -->|Yes| G["Route to client services, logged for the SOP"]
  F -->|No| H["Route to the shift supervisor"]

Two ways to buy the same answer, side by side

Assumptions, illustrative: 240 hourly associates and 14 salaried staff; per-seat pricing of $30 per user per month; a self-hosted setup on one on-site server sized for a building this size; outside help at $145 an hour from the managed service provider who already looks after your warehouse system.

Now run it backwards, because the break-even is the number you actually want. Steady-state self-hosting costs about $24,700 a year regardless of how many people touch it. At $360 per seat per year, that equals roughly 69 seats. Under about 70 real users, pay per seat. Over about 70, running your own starts winning, and the more hourly associates you add the wider the gap gets.

That is why this is a warehousing question specifically. A 12-person freight brokerage never crosses the line. A fulfillment operation with 240 associates and a seasonal surge to 400 crosses it before the first pallet of holiday goods hits the dock.

What "run your own" costs in people, honestly

The hardware is the easy part. The part that sinks operators is that nobody owns it.

You need a named person, internal or at your service provider, who is responsible for three things: keeping the machine patched, keeping the client SOP documents current, and being reachable when it stops answering at 2 a.m. during peak. The middle one is the sleeper. An answer desk that quotes a client's 2023 pack-out instructions is worse than no answer desk, because associates will trust it. Somebody has to update those documents every time a client changes a kit, and that somebody is usually your client services rep, who now has a recurring task they did not have before.

Budget the support honestly. Six hours a month is realistic for a single-building setup that is not being extended; if you connect it to the warehouse system so it can answer "which lot is allocated to this order," multiply that. And accept that when it breaks, it is your outage. No vendor to call, no service credit — you have traded a monthly invoice for an operational responsibility, which some warehouse operators are perfectly comfortable with and some are not.

Where paying per seat still wins

For the 14 salaried people, keep the paid seats. Your VP of operations building a rate proposal, your client services rep drafting a QBR deck, your billing clerk working through a client agreement — heavy, varied work where the frontier models are still meaningfully better and $30 a month is nothing against the hours. Claude Cowork and ChatGPT Work are aimed squarely at those people.

Keep per-seat anywhere the work leaves your building too. Anything a client or a retailer sees, anything that goes into a bid, anything where being subtly wrong is expensive — use the best thing you can buy.

And there are two things you should never route to any of this, hosted or rented. First, safety and hazmat. If an associate asks whether a placarded pallet can be stored in a particular aisle, the answer comes from your safety coordinator and your written hazardous materials procedures, not from a machine reading a binder. Build that stop into the system deliberately, the way the chart above shows. Second, anything that touches a person's employment — write-ups, productivity scoring, scheduling discipline. Those go to your HR generalist, and putting them anywhere else is how you end up explaining yourself to a state labor board.

The 90-day test that costs you almost nothing

Do not buy a server yet. For ninety days, buy per-seat licenses for one shift's supervisors and leads only — call it eight people — and load your ten largest clients' SOPs. Have them log every question the floor brings them and whether the tool answered it correctly. At the end of ninety days you will have two numbers: how many questions per shift the floor actually asks, and what share of them were answerable from documents you already have.

If volume is high and the answer rate is above about eighty percent, you have the case for putting a machine in your building and giving it to all 240. If volume is low, you just saved $22,000 in hardware and a support contract, and you keep paying for the fourteen seats that were always going to earn their money.

Frequently asked questions

Do my associates need to be given a device for this?

Mostly not. The handhelds and voice headsets already on the floor — Zebra scanners, Honeywell voice units — are how the answer should reach them, and most operators start with the shared tablets at the pack stations and receiving desk. Personal phones are a bad first step in a warehouse for reasons that have nothing to do with AI.

Does an open model handle Spanish on the floor?

Yes, and that is one of the stronger reasons to do this in warehousing specifically. A bilingual floor with heavy seasonal turnover benefits enormously from an answer desk that responds in the language the associate asked in, and with a self-hosted setup you are not paying per temporary worker for the privilege.

What happens during peak when I go from 240 to 400 people?

Nothing, on the self-hosted side — that is precisely the advantage. Your cost does not move when you add 160 seasonal associates, which is when you most need every one of them to get a correct answer without stopping a lead. On per-seat pricing, peak is exactly when the bill spikes.

Is this something my WMS vendor will eventually just include?

Probably some of it, and if you are on Manhattan, Blue Yonder or a major platform it is worth asking their roadmap question before you spend. But WMS-included assistants tend to answer questions about the system, not about your client's pack-out instructions and your own written procedures. Those documents are yours, and whatever reads them should be under your control.

One more phone problem peak creates. Every seasonal hiring push generates calls and web enquiries from applicants at hours when nobody is at the front desk, and most are lost by the time someone calls back. CallSphere builds AI voice and chat agents that answer the line and website chat around the clock, ask the qualifying questions, and book the interview slot. It will not tell a picker what the pallet pattern is — but it makes sure the picker you are trying to hire in October gets through to somebody.