The CFO's guide to AI budgets

Every other software line in the budget was approved before it was spent. Someone signed a contract, seats were provisioned, and the monthly cost was known before the month began. AI broke that. An API key issued in January can cost a hundred dollars in February and twenty thousand in March, with no purchase order in between. The budget did not fail. It was never asked.

This is a guide to budgeting for that, written for the finance leader who has to put a number in the plan and defend it in the review.

The three shapes of AI spend

Budget by shape, because each behaves differently.

Committed. Enterprise seat contracts: a Claude Enterprise or ChatGPT Enterprise agreement, a coding assistant for the engineering org. Priced per seat, paid annually or monthly, known in advance. Budget it the way you budget any license: seats times price, with a line for growth.

Metered. API usage billed by the token, and the usage credits that sit on top of some enterprise plans. This is the shape that surprises people. It scales with what people do, not with how many of them there are, and it can move by an order of magnitude in a week when someone ships an agent.

Incidental. Personal plans expensed on cards, tools individuals sign up for, trials that became subscriptions. Small individually, meaningful in aggregate, invisible until someone reconciles the card program.

Most companies have a number for the first, a guess for the second and nothing for the third.

Who should own the budget

Not the CFO, and not IT. The owner of an AI budget should be the leader whose team's output it is supposed to improve, at the level where a person can actually change behavior.

That usually means the team level. A sales leader owns the sales team's AI budget because they can see which reps are spending and what those reps are closing. An engineering manager owns theirs because they can see which agents are running and which services are calling models in production. A company-wide AI budget owned by finance produces one number that nobody can act on.

The finance team's role is to set the envelope, split it into team budgets with the owners, and see everything.

The right period is a month

Annual AI budgets are planning documents. The operating unit is a month, for two reasons.

Metered spend changes too fast for a quarter. A team that discovers a new workflow in week two will have spent the quarter's allocation by week six, and the budget will have been a fact on a slide rather than a control.

And a month lines up with the provider's invoice, which is what the number gets reconciled against. A budget you can check against a bill is a budget people trust.

What a budget is for

Here is the part that gets it wrong most often. An AI budget is not a cap, and the first control you reach for should not be a spend limit.

Providers offer hard caps. Turn one on and a rep in the middle of a deal, or an agent halfway through a run, stops working with no warning. The cost of the cap is invisible and the cost of the overspend is visible, so the cap looks like discipline. It is usually a way of losing output to save a number nobody had put next to that output.

A budget is a threshold that starts a conversation. When a team crosses it, the owner should be told, in the place they already work, with the figure and the people behind it. Then the owner decides: this is the best money we spend and the budget was wrong, or this is one person running something they should not, or this is a shared key everyone is using and we did not know. Each of those is a different action, and none of them is "block it".

Record the decision. A budget exceeded, a notice sent, and a note of what the owner decided is an audit trail that will matter when the annual plan is built from it.

The number to start with

If you have never budgeted AI before, do not start from a target. Start from the last three months of actual spend, by team, by shape.

For committed spend, the contract is the budget. For metered spend, take the monthly run rate per team, add the growth you expect from the workflows you know are coming, and set the threshold there. For incidental spend, reconcile the card program once, count the plans, and either bring them under a contract or set a policy.

Then, and this matters, put the budget next to what the team produced in the same months. A sales team that spent forty thousand dollars on AI and closed eight million dollars has a ratio. That ratio, not the forty thousand, is what tells you whether the budget should go up.

What to ask for from the tooling

Whatever system holds the budget has to be able to do four things, or the budget is decorative.

• Show spend per person and per team, every day, across every shape, with the unattributed portion visible rather than spread.
• Compare the month so far against the budget every night and tell the owner when it crosses, where the owner already is.
• Include shared keys in the team's figure, so a budget cannot be routed around.
• Record that it was exceeded and what was decided, so the history exists when the next plan is written.

Anything that only shows the invoice total is a report, not a budget.

Arbitir sets budgets at the organization, team or person level, checks them nightly against spend across API keys and enterprise seats with shared keys included, sends the notice to Slack, and records the event. It does not block anything, and that is deliberate.