Back to blog
Last updated November 2026·Poyan Karimi

AI Spend Management: How to Track What AI Costs and Whether People Use It

AI spend management means knowing, every month, what the company pays for AI, who and what drove the bill, and whether the people you pay for actually use it. Cost and adoption belong in the same view. A seat nobody uses and a person spending ten times the average are both findings, and you only see them when the two numbers sit side by side.

Do you want to keep an eye on your AI spend and adoption? Below: where the money hides, the five numbers worth tracking, what each vendor lets an admin see, and a monthly routine that takes about an hour.

Where AI spend hides

Two years ago AI was one line on the software budget. Today it is spread over several, and the usual moment of discovery is a bill that is higher than last month’s with nobody able to say why. In a company of 20 to 1,000 people the spend sits in five places:

WhereHow it is billedWhy it is easy to miss
SeatsA fixed price per person per monthSeats are added when people join and rarely removed when they stop using the tool
Metered usage and APIPer token or per credit, on top of the seat or instead of itIt changes every month and grows as people get better at using AI
Coding tools and agentsUsually metered, or a higher-priced seatA few engineers can account for most of the bill
Overlapping toolsSeparate contracts bought by different teamsSales pays for one assistant, marketing for another, IT for a third, on three budgets
Personal subscriptionsExpense claims and company cardsThey appear as small card payments, not as software

The last row is a cost problem and a data problem at once: a personal account sits outside your company terms. That side is covered in what shadow AI is. For the third row, we have a separate guide on what Claude Code costs a team.

Which plan you are on decides how much of this is fixed. A seat with usage included gives a predictable bill. A seat with metered usage on top does not. The two models are compared in Claude Team vs Enterprise and ChatGPT Business vs Enterprise.

The five numbers to track each month

AI FinOps at this size is not a platform or a team. It is five numbers, written down on the same day each month, so that the trend is visible.

NumberWhat it tells you
Total AI spendEvery line from the table above added together: seats, usage, API, coding tools, expenses. Most companies have never seen this as one figure.
Cost per active userTotal spend divided by the people who actually used the tool that month. Cost per seat flatters you. This one does not.
Unused seatsSeats with no activity in the last 30 days. Money with nothing to show for it.
Top spendersThe five or ten people or keys that account for the most usage. On metered plans they often explain most of the change from one month to the next.
Spend per teamWho the money is working for. A team with high spend and a clear use case is a result. A team with seats and no activity is a rollout that did not land.
An example of cost per active user

Anthropic’s pricing page lists a Claude Team Standard seat at $20 a month on an annual plan (checked 17 November 2026). A company with 40 such seats pays $800 a month. If 22 people used Claude that month, the cost per active user is $800 divided by 22, about $36, not $20. This is an example with invented headcounts, to show the arithmetic.

These are money numbers. Whether the usage is changing how the company works is a different question, with its own five numbers, in how to measure AI adoption. Read the two together: low cost per user means little if nobody has a use case, and a high bill can be the best news in the report.

What each vendor lets an admin see

Each of the main assistants has an admin view for AI usage tracking. What follows is from the vendors’ own help pages, checked 17 November 2026. The views change often and differ by plan, so confirm in your own admin console.

ToolWhat an admin can see and limit
Claude Team and EnterpriseOwners can open usage analytics and export reports. Spend per individual member is described for Enterprise organisations on usage-based billing, not for Team. For extra usage beyond the seat, owners can set a monthly limit for the whole organisation, and limits per member where that option is offered.
ChatGPT Business and EnterpriseOpenAI documents workspace analytics for Enterprise and Edu, with active users over time and seats activated, and an admin console with usage and credits sections. We could not confirm from OpenAI’s pages exactly which of these views Business includes.
Microsoft 365 CopilotThe Copilot usage report in the Microsoft 365 admin centre, under Reports then Usage, shows enabled users against active users, with a breakdown per app. It is a usage report, so the cost side comes from your licence count.

The limit is the same in all three: each view covers only its own tool. None of them shows the expense-card subscriptions, the API keys in engineering, or the same person holding a seat in two assistants. The total is something you have to assemble yourself. If you want a wider frame for that work, the FinOps Foundation, the body behind cloud cost management practice, publishes guidance on FinOps for AI.

A monthly routine in five steps

One person, one hour, the same day every month. A spreadsheet is enough to start with.

  • Collect every bill. Vendor invoices, API and cloud bills, and a search of expenses and card statements for AI subscriptions. Add them into one total.
  • Export active users from each admin console and divide: cost per active user, per tool and overall.
  • List the seats with no activity in 30 days. Note who holds each one and which team they are in.
  • List the top spenders and write one line on what each of them is doing with it.
  • Decide and record. For every unused seat and every top spender, choose an action from the table below, and compare the total with last month.

What to do with a high spender and an unused seat

Neither finding is good or bad on its own. Ask what is behind it before you act.

FindingWhat is behind itWhat to do
High spenderReal work that returns more than it costsLeave them alone, give them the right seat or limit, and ask them to show the rest of the team how they work
High spenderWaste: the most expensive model for simple tasks, very long conversations, an agent left runningA short conversation and a sensible limit. Most people change once they see the number
Unused seatThe person never got startedOne hands-on session on a task from their own week, then check again next month
Unused seatThe role has no use for it, or they use another toolReassign the seat to someone on the waiting list, or remove it at renewal
Two tools doing the same jobTeams bought separatelyKeep both only if each has a use case the other cannot cover. Otherwise pick one at the next renewal

Cutting a high spender without asking is the costly mistake. The person using ten times the average is often the one who has worked out how to make AI do a real part of the job.

Where Deployed OS fits

The routine above works by hand. It gets slow when the cost figures sit in one console and the adoption figures in another, which is the gap Deployed OS closes for Claude.

Deployed OS

Deployed OS shows Claude tokens and cost per person and per team, over time, next to adoption: who is active, which teams use AI, which skills and connectors are used and which nobody touches. The Claude figures are imported from Claude once an admin adds the company’s Claude analytics key.

People are shown anonymously by default, as Person 1, Person 2 and so on, until an admin switches names on. That matters for companies that must consult a works council or union before looking at individual usage. Deployed OS also holds the company’s skills, connectors and rules in one place across Claude, ChatGPT and Copilot.

How Deployed OS works

Whichever way you collect them, the point is the same: put the cost next to the usage, look once a month, and act on what you find.

Frequently asked questions

What is AI spend management?

AI spend management is the monthly habit of knowing what a company pays for AI, who and what drove the bill, and whether the people it pays for actually use the tools. It covers seats, metered usage and API costs, coding tools and agents, overlapping tools bought by different teams, and personal subscriptions paid on expense cards. Cost and adoption are read together, because an unused seat and an unusually high spender are both findings.

How do you track AI costs across a company?

Collect every AI bill in one place once a month: vendor invoices, API and cloud bills, and AI subscriptions found in expense claims and card statements. Then export active users from the admin console of each tool and work out five numbers: total spend, cost per active user, unused seats, top spenders and spend per team. A spreadsheet is enough to start with, and the trend from month to month matters more than any single figure.

What is AI FinOps?

AI FinOps applies the practice of FinOps, which grew out of managing cloud bills, to AI costs: seeing who spends what, allocating it to teams, and deciding where the spend is worth it. The FinOps Foundation publishes guidance under the name FinOps for AI. For a company of 20 to 1,000 people it does not need a dedicated team, only a named owner and a monthly review of a handful of numbers.

How do you calculate AI cost per active user?

Divide the total AI spend for the month by the number of people who actually used the tool in that month, not by the number of seats. As an example, 40 seats at $20 a month cost $800. If 22 people were active, the cost per active user is about $36. The gap between the seat price and that figure shows how much of the spend goes to seats nobody uses.

What should we do with AI seats nobody uses?

First find out why. If the person never got started, one hands-on session on a task from their own week is usually cheaper than leaving the seat idle, and you check again the next month. If the role has no use for the tool, or the person uses a different one, reassign the seat to someone who is waiting or remove it at renewal. Review seats with no activity in 30 days every month so they do not pile up.

Found this useful? Send it to someone who needs it.

Deployed OS

See what your AI can reach.

Deployed OS puts your company's skills, systems and guardrails in your own company MCP. Each employee connects once, in the AI they already use, and you decide what each team can reach. Leave your email and we'll be in touch shortly.

Proof from the field

All cases

Keep reading