Per-Seat or Usage-Based: AI Agent Pricing

Per-seat AI pricing charges for the right to have an account. Usage-based AI pricing charges for the work actually done. The difference rarely shows up first in the invoice total — it shows up in who on your team feels free to try something.

---

Most software a business buys is priced per seat, so that is the shape buyers expect when they evaluate an AI platform. It is a familiar, predictable model, and for tools where every user does roughly the same amount of work, it is a fair one.

AI agents are not that kind of tool. The person who runs one report a month and the person whose agents process every inbound support case are doing wildly different amounts of work behind the same login. This post covers what each pricing model actually charges for, what Insulin meters and what it deliberately does not, and how the choice changes adoption behaviour on a team.

What per-seat pricing charges for

Per-seat pricing is a licence fee for access — a fixed amount per named user per month, whether that person runs one query or ten thousand. The bill is a function of headcount, not of work.

That predictability is the model’s real strength: finance can forecast it a year out from a headcount plan. But it has a consistent side effect on adoption, and it is worth naming before you compare price sheets. When access costs money, every new user becomes a purchase decision. Three things tend to follow:

• Occasional users never get added. The analyst who would benefit twice a quarter cannot justify a full seat, so the tool never reaches the edges of the organization where a lot of the repetitive work actually sits.
• Logins get shared. When a seat is scarce, people route requests through whoever has one. That breaks attribution, and it breaks permissions.
• Renewals become seat audits. The annual conversation is about trimming inactive users rather than about what the tool produced.

None of that is an argument that per-seat pricing is wrong. It is an argument that the pricing model is also an adoption policy, and it is easy to buy one without noticing you bought the other.

What usage-based pricing charges for

Usage-based pricing is a charge for resources consumed, metered in the unit each resource is naturally measured in. Insulin’s documentation is explicit that “there is no seat count, no tier and no subscription” and that “an organization that uses nothing pays nothing.”

Insulin meters six resources:

Resource	Measured in
AI input tokens	tokens
AI output tokens	tokens
Sandbox processor	processor-seconds
Sandbox memory	GB-seconds
Sandbox storage	GB-seconds
Object storage	MiB-months

Two ways to pay for the models themselves. You can use Insulin’s hosted models and be billed per token with the rate shown in the console before you run it, or connect your own provider account, in which case that provider bills you at their rates and Insulin adds a flat fee per unit rather than a percentage. The full rate card is on the pricing page.

The adoption consequence is the mirror image of per-seat. Adding a colleague costs nothing until they do something, so the tool spreads to the people with occasional work instead of stopping at the people who could justify a licence.

Does metering punish exploration?

This is the fair objection to usage-based pricing, and it deserves a direct answer rather than reassurance. If every click costs, people click less — and a platform nobody explores produces nothing.

What is not metered matters as much as what is. Insulin’s documentation lists three categories that are never billed to you:

• The platform’s own AI. Features that call a model internally — summarization, search indexing, embeddings, reranking, field mapping — are Suger’s cost, not yours.
• Reading what you already have. Opening, listing or searching your stored files, and starting, stopping or connecting to an existing sandbox.
• Database storage.

So browsing your own knowledge bases, searching them, and reopening past work are free actions. What costs money is generating new tokens and running new compute — the work itself.

On top of that, every organization gets $100 of free credit each month, and no payment method is needed to spend it. For a team evaluating whether agents fit a workflow at all, that is usually the whole evaluation.

What happens when the credit runs out

A spend limit should pause work, not destroy it — and this is where usage-based platforms differ most from one another, so it is worth checking on any vendor you evaluate.

At a zero balance in Insulin, AI requests pause, and creating new sandboxes or uploading new files stops. What keeps working is everything that reads what you already have: opening, listing and searching stored files, and any sandbox already running. The documentation is unambiguous about the guarantee — “No data is deleted in any of these states. A running sandbox is never killed to collect a debt.” Adding credit resumes service about a minute later, with no manual step.

That distinction is the one to test in a trial. “What stops” is a pricing question; “what is destroyed” is a data question, and only one of them is reversible.

Choosing between them

Neither model is universally better. They optimize for different things, and the right question is which failure mode you would rather have.

	Per-seat	Usage-based
Bill tracks	headcount	work done
Easy to forecast	yes, from a headcount plan	needs a spend limit and a review cadence
Cost of adding an occasional user	a full seat	nothing until they use it
Heavy automation	effectively free once seats are bought	priced proportionally
Main risk	paying for inactive seats	an unwatched job running up consumption

Per-seat suits a team where usage is uniform and budget predictability outranks reach. Usage-based suits a team that wants the tool available to everyone and is willing to manage consumption — which is a real obligation, not a footnote. Set a monthly spend limit, and read enterprise AI spend governance for how to attribute cost by user, app and model once more than a couple of jobs are running.

Frequently asked questions

Does Insulin charge per seat? No. The billing documentation states there is no seat count, no tier and no subscription. Billing is consumption-based, so an organization that uses nothing pays nothing, and adding a colleague costs nothing until they do work.

What exactly does Insulin meter? Six resources: AI input tokens and AI output tokens, measured in tokens; sandbox processor, in processor-seconds; sandbox memory and sandbox storage, in GB-seconds; and object storage, in MiB-months.

Is searching my own documents billed? No. Reading what you already have is not metered — opening, listing or searching stored files, and starting, stopping or connecting to an existing sandbox. Platform AI calls, such as search indexing and embeddings, are also not billed to you.

What happens if the credit runs out? AI requests pause, and creating new sandboxes or uploading files stops. Reading, listing and searching stored files continues, and running sandboxes keep running. No data is deleted, and service resumes about a minute after credit is added.

Can I use my own model provider? Yes. Your provider bills you directly at their own rates, and Insulin adds a flat fee per unit rather than a percentage, the same whichever model you run.

Takeaways

• A pricing model is also an adoption policy. Per-seat pricing makes every additional user a purchase decision; usage-based pricing makes the work itself the thing you pay for.
• Check what is not metered. Reads of existing work and the platform’s own AI calls are not billed in Insulin, which is what keeps exploration from carrying a cost.
• Test what happens at zero. Pausing AI requests is a pricing behaviour; deleting data would be a different thing entirely. Insulin deletes nothing and never kills a running sandbox to collect a debt.
• Usage pricing comes with an obligation. Set a spend limit and review consumption on a cadence, or proportional cost becomes surprising cost.

Insulin is priced for the work, not the roster. See the pricing page for the rate card, or book a demo to walk through what a month of your workload would actually meter.