What Should a CFO Ask About the Company's AI Bill?

A CFO should ask six questions about the AI bill: what the total is and how it splits between seats and use, who holds a paid seat without using it, which roles account for the spend, whether the work in those roles suits AI, whether output moved where use is heaviest, and whether the model is larger than the task needs. In Deloitte’s State of AI in the Enterprise survey of 3,500 leaders, published October 29, 2026, 71% track their total spending on running AI models, which Deloitte calls inference and API spend, while 44% track the cost per unit underneath it.

A token is the unit AI vendors bill by, roughly a piece of a word. Seat-based products charge a fixed price for each licensed person.

1. What is the total AI bill, and which lines are seats and which are usage?

Why it matters. Seats and use behave differently. Microsoft lists Microsoft 365 Copilot at $30 per user per month, paid yearly. A use-billed line moves with volume. Deloitte’s CFO Signals survey of 200 North American CFOs for the third quarter of 2026 notes that AI providers more commonly charge per token than per seat and that AI bills “can spike quickly.” GitHub’s Copilot plans mix the two, with a monthly allowance of AI credits and extra credits at $0.01 each. Prices in this article are list prices read on November 13, 2026, and negotiated prices differ.

Where the answer comes from. Start with vendor invoices, card statements and the vendors’ cost reports, such as Anthropic’s Usage and Cost Analytics API and OpenAI’s usage and cost reports. Look past the IT budget: in a BCG survey of 1,330 executives published September 30, 2026, more than 80% of AI spending sat outside enterprise IT. Some use also lands on the cloud bill, since Anthropic’s models are sold through Amazon Bedrock and other cloud platforms.

What a good answer looks like. One monthly figure for the whole company, business-unit budgets and cards included, with seat lines (product, seats, price per seat) and use lines (model, tokens, cost) shown apart, each marked as list or negotiated price.

2. Who has a paid AI license and has not used it?

Why it matters. A seat costs the same whether or not its holder uses it, and finance can find unused Copilot licenses from billing and activity records alone.

Where the answer comes from. Microsoft’s Copilot usage report gives each user’s last activity date, overall and in each app, over 7, 30, 90 or 180 days. GitHub’s seat list records each user’s most recent Copilot activity, though code editor activity shows only when the editor’s usage reporting is on.

What a good answer looks like. Seats with no activity over a stated window, by role and team, with their monthly cost at the price paid. Our guide to finding unused AI licenses flags a seat after 30 days and takes it back only after a reminder and a further 30 days, and shows where each vendor keeps the date. Ask each holder before reclaiming a seat. In a role where much of the work suits AI, ask whether training or some other barrier explains the idle seat before canceling it.

3. Which roles account for the AI spend, and what is the cost per employee in each?

Why it matters. A total by vendor does not show where in the company the money goes. The FinOps Foundation’s State of FinOps 2026, released February 19, 2026, found that 98% of its 1,192 respondents manage AI spend, and lists allocating AI costs to business units as a top AI challenge. In Deloitte’s CFO Signals survey, 20% of CFOs reported advanced controls, such as tracing costs down to tokens or to individual pieces of work.

Where the answer comes from. Anthropic’s Usage and Cost Analytics API groups token use and cost by the user behind each request, and OpenAI’s usage report can group by user. Microsoft’s Copilot report gives activity dates and no token counts, so keep activity as a separate measure. Then match each named employee to a role and team in the HR system, and assign use by shared service accounts by a stated rule.

What a good answer looks like. AI cost per employee in each role over a quarter, meaning the role’s spend divided by its average headcount, with seats and use shown apart, as in the worked example in What does AI cost per employee in 2026?. Dividing matters, because a raw total mostly shows which teams are largest. Small groups are combined until no figure describes a handful of people.

4. Does the work in those roles suit AI?

Why it matters. High spend matters most where little of a role’s work suits AI, and that share differs widely. Anthropic’s labor market paper of March 5, 2026 measures coverage, a weighted share of an occupation’s tasks that show up in professional use of Claude. Computer programmers lead at 75%, while 30% of workers have zero coverage. A zero means little observed use, which is different from work AI cannot do.

Where the answer comes from. The answer comes from task data. O*NET, sponsored by the US Department of Labor, holds more than 19,000 occupation-specific task statements, and the Anthropic Economic Index maps the tasks people bring to Claude to the occupations that perform them. Each role’s task scores then need checking with its managers, since a company’s version of a job can differ from the national description. Is your AI spend in the right roles? sets out the comparison, and Which roles should use the most AI? applies it to ten roles.

What a good answer looks like. For each high-spend role, the share of its tasks that suit AI, and an explanation wherever spend is high and few tasks fit. A role where many tasks fit and spend is low belongs on the list too, as a likely miss. Ask what stands in the way there, and train first.

5. Did output move in the roles that use AI most?

Why it matters. In KPMG’s Global AI Pulse for the third quarter of 2026, a survey of 2,131 senior leaders in 20 countries released September 24, 61% review AI costs during approval, while 12% consistently assess AI value against cost across the organization. Amazon’s engineering leadership, in a memo of October 27, 2026, told leaders to check whether the teams with the heaviest AI use, at the 90th percentile, were delivering more to match.

Where the answer comes from. The measure the role’s managers already track, read from the system where the work happens: tickets resolved in the help desk, days to close the books, contracts reviewed, sales pipeline in the CRM system. How to measure AI value outside engineering covers the measure for each function.

What a good answer looks like. The measure for the same team and period as the spend, before and after, with other changes, such as staffing, named. A rise alongside heavy use is an association to test. Microsoft cautions that the differences its Copilot Business Impact report shows between usage groups “have not been evaluated for statistical significance.” In plain terms, the report does not show that AI caused the difference.

6. Is the model larger than the task needs?

Why it matters. Anthropic charges more per token for its larger models. Input is what is sent to the model and output is what it writes back. Anthropic’s price list gives Claude Fable 5.2 at $8 per million input tokens and $40 per million output tokens, Opus 5.5 at $4 and $20, Sonnet 5.5 at $2 and $10, and Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens. In Deloitte’s State of AI survey, 44% track per-unit inference cost, and 38% of the CFOs in its CFO Signals survey named limited timely insight into token usage and unit costs as a barrier.

Where the answer comes from. The same usage reports, separated by model. Anthropic’s report and OpenAI’s usage report both break token use down by the model used.

What a good answer looks like. Each role’s tokens by model. For roles whose work is mostly routine summaries, a smaller model is tested on the role’s own work against its quality bar before any switch.

Worked example: a fictional company’s bill

This company is invented, and its prices are list prices with no discounts. It pays for 600 Microsoft 365 Copilot seats at $30 per user per month, billed yearly, or $18,000 a month on top of the Microsoft 365 plans the seats require. Its support team also uses an Anthropic API account, which in one month processed 500 million input tokens and 100 million output tokens on Claude Opus 5.5 at $4 and $20 per million. That is 500 × $4 plus 100 × $20, or $4,000. The AI charges in the example come to $22,000 a month, about 82% of it seats.

Assume the Copilot usage report shows 90 seats with no activity in the last 60 days, the reminder included. At the $30 list price they cost 90 × $30, or $2,700 a month and $32,400 a year. Before canceling any, finance checks the roles behind them and asks the holders why.

The API work is mostly summaries of support tickets. Suppose half of it passes the team’s quality check on Sonnet 5.5, listed at $2 and $10 per million. Assuming the same token volumes, that half, 250 million input and 50 million output tokens, would cost $500 plus $500, or $1,000 instead of $2,000. With $2,000 left on Opus plus $1,000 on Sonnet, the month’s use line works out to $3,000.

Where should a finance team start?

Finance can answer the first two questions from invoices and admin reports. The others need HR records, task data or output measures, so finance and HR work from the same figures.

JobRoute reads the usage reports of a company’s AI tools, joins them to each role’s tasks and output measure, and shows spend by role and paid seats nobody is using, with dollar figures estimated from public list prices unless the company enters its own: see AI spend by role.