Pricing models
Per-seat pricing and usage pricing charge for different things
One model prices access. The other prices output. Almost every argument about software cost is really about which of those you are buying.
What it costs
Per-seat pricing charges for the number of people who can open a product, so the bill tracks headcount and every new teammate costs more whether or not they use it. Usage pricing charges for units of work the product delivered, so the bill tracks output and an unused month costs nothing. Polaris prices software at $0 and meters delivered work at about $2 per human-equivalent hour.
- Per-seat unit
- A person with access
- Usage unit
- Work delivered
- Polaris software
- $0, no seats
- Polaris meter
- ~$2 per human-hour delivered
A price is an answer to the question: what am I buying a unit of?
Per-seat software sells you a licence for a person. The product's job is to be somewhere work can be recorded, and the vendor's revenue grows when more people are allowed in. That is a coherent model and it built most of the software industry. It also means the invoice describes your org chart, not your output.
Usage pricing sells you a unit of something happening. Cloud infrastructure has worked this way for twenty years and nobody finds it strange: you pay for compute that ran, storage that was held, requests that were served. The awkward part has always been that the unit is a machine unit, so a finance team has to translate gigabyte-months into anything they care about.
What changes with AI workers is that the unit can finally be a work unit. Not tokens, not credits, not seats. An hour of the work a person would otherwise have done, estimated from what the job actually did, and itemised.
Per-seat logic breaks in two places. The first is tool sprawl: once a company runs four products across one headcount, it is paying four times for the same people, and each product's per-seat rate is individually reasonable while the total is not. A twenty-five person team in this cluster pays $551.88 per person per year across three tools, and none of those three tools ever did any work.
The second is AI. Charging per seat for an AI feature prices access to a thing whose whole value is variable output. A colleague who runs two prompts and a colleague who runs two hundred cost the same, which turns a productivity question into an allocation argument about who is allowed a licence. Companies of fifty are having that argument right now, and it exists only because of the pricing model.
The same company under both models
A twenty-five person team, ten of whom do meaningful work in the product and fifteen of whom mostly read.
| Question | Per-seat pricing | Usage pricing |
|---|---|---|
| What is the billable unit? | A person with access | Work that was delivered |
| What happens when you hire? | The bill rises immediately | The bill does not move |
| What happens in a quiet month? | You pay in full | You pay less, or nothing |
| What happens to the fifteen readers? | Billed the same as the ten doers | Cost nothing, because they produced nothing to bill |
| Can the invoice be tied to an outcome? | No line corresponds to anything produced | Every line names a job |
| How predictable is next month? | Exactly predictable | A forecast you have to make |
| What is the vendor rewarded for? | Getting more people licensed | Delivering more work that gets accepted |
The honest case for per-seat pricing
This is not a strawman we are about to knock over. Per-seat won for good reasons and still wins in specific situations.
Finance can plan it in one line
Headcount times rate times twelve. A CFO can forecast next year's software spend in a spreadsheet cell, and there is real value in a number that never surprises anybody.
Procurement is built for it
Annual seat contracts fit purchase orders, approval thresholds and renewal calendars. Usage pricing regularly fails a procurement process not because it costs more but because nobody knows what number to put on the form.
It aligns when the product really is a place
If what you are buying is genuinely a shared record that people read and write, then people is the right unit. A wiki priced by the word would be absurd.
It cannot run away from you
The worst case of a per-seat bill is known in advance. The worst case of an unbounded meter is not, and that fear is legitimate rather than irrational.
What each model is actually optimising
Per seat
- Vendor grows by increasing licensed headcount
- Customer saves by restricting who gets access
- The two incentives point in opposite directions
- Nobody in the loop is measured on work produced
Per unit of delivered work
- Vendor grows only when more delivered work is accepted
- Customer saves by assigning less, or by rejecting bad work
- Both sides are measured on the same thing
- Bad output is a revenue problem for the vendor, not just a support ticket
The three fears of a usage meter, and what answers them
Usage pricing fails on trust more often than on arithmetic. These are the objections in the order buyers raise them.
- 1
It will run up a bill while I am not looking
The unit has to be legible before anything else matters. In Polaris a single agent session is clamped to a maximum of eight human-equivalent hours, which at roughly two dollars an hour is about sixteen dollars for the largest job the system can bill. Work also arrives as a delivery you see, on a task you assigned, not as background consumption.
- 2
I will pay for work that is wrong
Agents deliver and humans close. An AI worker posts its result as a comment, ticks its own acceptance-criteria checklist, and stops. Nothing is ever marked done by the machine. You close and rate the task, and work you reject is work you can dispute.
- 3
I cannot predict what this costs per month
This one is real and we are not going to pretend it away. Polaris is in free public beta with no customers and no usage data, so there are no cohort ranges to quote. What exists instead is auditability: the formula is published, every job logs its inputs, and you can recompute any line yourself.
Where to go next
Per-seat pricing
The bill tracks headcount, which is only a proxy for value received.
Usage-based pricing
The bill follows consumption, which cuts both ways for buyer and vendor.
Human-equivalent hours
A billing unit denominated in the work replaced, not the compute consumed.
What an AI worker costs, and how the hours are counted
The whole formula is on this page, including the parts that make it an estimate rather than a measurement.
The cost of AI subscriptions for a team
Five figures a year at fifty people, and no shared record of what any of it produced.
Tool stack cost for a 25-person team
The size where paying monthly instead of annually costs $1,650 a year on its own.
Questions people ask
+What is the difference between per-seat and usage-based pricing?
Per-seat pricing charges a fixed rate for each person licensed to use a product, so the bill follows headcount. Usage-based pricing charges for units of something the product did, so the bill follows output. The first is predictable and unrelated to value delivered; the second is variable and directly related to it.
+Is usage pricing always cheaper?
No. At high volume a meter can cost as much as the subscriptions it replaced, and the fifty-person stack page in this cluster shows a volume where it does. The argument for usage pricing is not that it is always smaller. It is that every line corresponds to something that arrived.
+What is Polaris's billable unit?
A human-equivalent hour of work delivered by an AI worker, priced at roughly two dollars. The software itself is free with unlimited humans, tasks, workstreams and docs. Nothing delivered means nothing billed, in any month.
+How do I know the hours are not inflated?
The formula is published and mechanical: base pickup time, twelve minutes per web search, finished prose at about ninety characters a minute, eight minutes per acceptance-criteria item, ten minutes per comment addressed and per file produced, five per PDF, clamped between five minutes and eight hours per session. Every input is recorded on the job's work log, so any line can be recomputed and challenged.
+Why not just charge per token or per credit?
Because neither is a unit anyone can evaluate. A credit is an invented currency, and a token count tells a buyer nothing about whether the work was worth doing. An hour of human-equivalent effort is a unit a manager can compare against a salary, a contractor rate, or the time they would have spent themselves.
+What happens if an AI worker does bad work?
You close nothing and rate nothing. The delivery sits on the task as a comment, and the task stays open. Because a human closes every task in Polaris, the acceptance decision is always yours, and the work log shows exactly what the hour estimate was built from if you want to argue about the line.
Related
Per-seat pricing
The bill tracks headcount, which is only a proxy for value received.
Usage-based pricing
The bill follows consumption, which cuts both ways for buyer and vendor.
Human-equivalent hours
A billing unit denominated in the work replaced, not the compute consumed.
What an AI worker costs, and how the hours are counted
The whole formula is on this page, including the parts that make it an estimate rather than a measurement.
The cost of AI subscriptions for a team
Five figures a year at fifty people, and no shared record of what any of it produced.
Tool stack cost for a 50-person team
Three work tools and two AI subscriptions, which is where most companies of this size actually are.
Tool sprawl
The expensive part is not the licences. It is that nothing is authoritative any more.
Alternatives to the work tools you pay for by the seat
Sixteen pages about tools teams leave, what leaving actually costs, and when leaving is the wrong call.