Why per-seat AI licenses hide the real cost of AI

A per-seat contract is the most reassuring shape AI spend can take. It has a price per person, a count of people, and a total that a finance team can put in a plan and be right about for twelve months. That reassurance is the problem. The seat price tells you what access costs. It tells you nothing about what the access is producing, and it is designed so that you never have to ask.

What a seat price actually covers

Enterprise AI contracts are converging on a shape: a per-seat fee for access, plus usage on top. Anthropic's enterprise plans and OpenAI's enterprise plans both report seat usage in their own units, and both meter the heavier features, agents, deep research, extended reasoning, against a pool of credits or a usage allowance that sits above the seat.

So a seat is two costs, not one. The fixed part is the seat price. The variable part is what the seat consumes, which is invisible on the invoice and visible only in the provider's analytics. A company that budgets the fixed part and ignores the variable part has budgeted the part that does not change.

Three things the seat count hides

Idle seats. In most deployments a meaningful fraction of seats are used rarely or never. The person left, or changed roles, or tried it twice and stopped. Each idle seat is a full seat price for nothing, and the provider has no incentive to point it out. The analytics surface will list every member, used or not; someone has to look.

Concentrated usage. Among the seats that are used, usage follows the same curve as everything else: most seats use a little, a few use a great deal. Per-seat pricing averages that out on the invoice and hides it in the contract. The people who are getting the most from the tool and the people getting nothing pay the same, and neither is visible.

The variable spend. The credits consumed by heavy features are the part of the bill that moves. On a credit-based plan, a small number of people running agents can draw down the whole pool, and the first sign is the pool running out. On a usage-billed plan, it is the invoice going up. Neither tells you who.

How to see through it

The seat providers do expose the data. What follows is what to read and how to use it.

List every member. Both enterprise surfaces report the member list independently of usage. Join it to usage by email, and every seat with no usage in the window is an idle seat with a name. Do this monthly and the count of idle seats becomes a number in the renewal conversation.

Read usage per person. Claude Enterprise reports per-user usage and cost by product and by model, daily. ChatGPT Enterprise publishes per-user credit usage by product and model, hourly, on a cost log. In both cases the figure is attributed to the person by the provider; there is no allocation to do. Store it per person per day, as reported.

Put the seat price on the same page. A person's true seat cost is the contract price for their seat, spread across the days of the billing period, plus what the seat consumed. Show both, summed, per person, and the concentration curve appears immediately.

Separate the commitment from the usage in the ledger. The contract is a commitment; the usage is a fact. Keep them in different places so that a renewal conversation can reason about the commitment and a productivity conversation can reason about the usage. Mix them and you get a number that answers neither question.

What to do at renewal

Renewal is where this pays for itself. Walk in with three numbers.

• Seats used versus seats paid for, with the list of idle seats. This is the direct saving.
• Usage concentration: the share of consumed credits attributable to the top ten percent of seats. This is the argument for a tiered structure, or for moving heavy users to a different plan, or simply for knowing who to protect.
• Spend per seat next to output per seat, for the roles that have a countable output. This is the argument for buying more seats for the people producing, not just for keeping the count flat.

Without these numbers, renewal is a negotiation about a percentage on a total. With them, it is a decision about people.

The general principle

Per-seat pricing turns a usage cost into a fixed cost on the invoice. That is a convenience for the buyer's finance team and a protection for the vendor's revenue, and it is fine as a commercial arrangement. It is not fine as the only view a company has of its AI spend, because it removes exactly the information that would let the company allocate the spend well.

The seat is the price of access. The usage is the price of the work. Budget the first, watch the second, and never confuse the total for an answer.

Arbitir reads the per-person usage from Claude Enterprise and ChatGPT Enterprise, lists every member so idle seats show as people with seats and no usage, keeps the contract itself as a license spread evenly by day, and shows the sum per person beside API spend and beside output.