Somewhere on your laptop is a spreadsheet with names in it. Two hundred rows, a column of checkboxes, and a running total at the bottom that you keep looking at. You’re deciding who gets an AI seat. Marketing wants twelve. The regional VP wants one for every director because it looks bad if he doesn’t ask. Finance wants the total to stop moving.
You’ll spend two weeks on that spreadsheet. The number in the bottom right is only one part of the bill.
The seat question had a real answer in 2024, when access was expensive and rationing was the whole job. The current answer has two parts: a flat license pool and a meter for agent and API work.
Seats and meters need different decisions
Here is the current shape. Flat per-seat business and workspace plans remain the default across the major vendors. Metering sits on top for agent workloads, APIs, credit packs, and Claude Enterprise.1 One spreadsheet cannot make both decisions.
Run that through the spreadsheet. A flat seat you hand to someone who never opens it is a real recurring cost. The discipline of saying no to the regional VP can still pay for itself.
Metered work behaves differently. An idle account still carries its access fee, but it does not accumulate consumption. The person who turns out to be your next power user adds a usage line that scales with the work they are producing.
The cost of a wrong “yes” and a wrong “no” depend on which part of the bill you are buying.
That split is the whole answer. Audit flat workspace licenses. Keep a close eye on named metered work, the output it buys, and the person who owns it.
Eighty dollars per employee is the number that holds up
For a knowledge-work organization, total AI spend should land between forty and a hundred and twenty dollars per employee per month, with the average closer to eighty. Below forty, you’re underfunding the people producing leverage. Above a hundred and twenty, you’re usually paying flat rates for seats nobody opens. Evaluating Spend has the full allocation model.
Notice that the benchmark is per employee, not per seat. Divide your AI spend by headcount and you get a number you can defend. Divide it by seat count and you only see the flat part of the bill.
So the seat rule is boring. Give a flat workspace seat to people with a recurring job for it, then audit the pool. Give metered access to the teams with a named workflow and an owner, then follow the usage against the output.
Then give the small fraction who use it hard everything they ask for: the better plan, API access, the consumption budget. The heaviest few percent of any role should usually have both a seat and API access, and the cost of that is small compared to what happens when you force a builder through a chat window. Recognizing Leverage is how you find those people. On a metered plan, sort the usage report by person. On a flat plan, use the activity report to find the people who have already made the seat worth keeping.
The flat-priced seats are where the audit still belongs
The old math holds completely in one place, and it’s probably your biggest line item.
Workspace AI is still sold on a flat seat. A Copilot seat bundled into your E5 costs thirty dollars whether the seat is opened or not.1 That is a real recurring cost, so an idle one is pure waste. If you have a hundred and fifty of them and a third are dead, you are burning about eighteen thousand dollars a year on nothing, and the rep will tell you it is already paid for.
This is where the quarterly audit still earns its keep, and it’s why killing the bottom thirty percent of seats is still the right advice. That advice is scoped to the flat-priced workspace SKUs. On a Copilot seat, a dead row costs you thirty dollars a month forever, so cutting it is free money. On metered agent or API work, the relevant audit is spend against the outcome the workflow bought. Same discipline, applied to the part of the bill that can become waste. The five-minute version in Evaluating Spend is the whole procedure; run the seat audit against your Copilot roster and use the meter for the rest.
Consumption is what you manage now
The seat count was only ever a proxy for consumption, and it was a decent one back when you had no way to see consumption directly. Now you can look at the thing itself.
Your bill is a usage report, so read it like one. Set per-team and per-user spend limits. Watch consumption by workspace rather than by headcount. Put one person on the dashboard monthly who can answer “why did this double” before your CFO asks it, and understand that the answer is often good: the Q3 briefing is blunt about the fact that a successful rollout produces a bigger invoice by design. The instinct to throttle the thing that’s working is the expensive mistake here, and most of the real savings live in plumbing your engineers already know about anyway.
The consumption that goes wrong is usually mechanical. An uncached script running ten thousand times against a workflow nobody profiled, burning tokens at the rate of a whole department. A heavy seat attached to a named person and a named output is leverage. A heavy seat attached to a job no one owns is an invoice. Both show up as consumption, and only the dashboard tells you which is which.
When the CFO asks how many seats you bought, you’ll have the number, and it won’t be what carries the meeting. The paragraph that works is the one in Defensible AI Spend: spend is up because the tools are metered and our use is up, it’s concentrated on people whose workflows we can name, and we audit it quarterly. That paragraph gets through a budget review without a seat count in it, because the seat count wasn’t answering the CFO’s question in the first place.
Something to carry
Go back to the spreadsheet. Find every row you unchecked on a metered plan and check it. That’s the whole decision, and it costs twenty dollars a head to be wrong about.
Then take the two weeks you were going to spend defending those checkboxes and spend them on the harder question, which is who in your organization is already producing something with the access they have. Those people are why the line item exists, and they’re the reason adoption stalls in the broad middle when you ration access to them. The org chart won’t tell you who they are. The usage report will.
Footnotes
-
Anthropic lists Claude Team at $20 per seat monthly on annual billing and $25 monthly; Claude Enterprise is access plus API-rate consumption. OpenAI lists ChatGPT Business at $20 per user monthly on annual billing and $25 monthly. Google lists Workspace at $7, $14, and $22 per user monthly, and Microsoft lists M365 Copilot as a $30 per-user add-on. Last verified July 29, 2026. ↩ ↩2