Closing the month on AI

Start from the invoices

In its April guide for CFOs, Deloitte describes a health care company whose token use grew 8 to 10% a month. Over six months that came to about a trillion extra tokens and more than $6 million a year of cost no one had planned for. The cost had already risen before the finance team could see what was driving it. [1]

Growth that fast can simply mean the tools are useful. What went wrong was the delay. For six months the cost kept rising on invoices, and finance had no view of what was driving it.

A month-end close exists to catch that kind of thing. Deloitte has a name for what follows when token cost spreads across software contracts, cloud bills and services invoices: margin leakage. [1]

We think AI cost should go through the close like any other cost. The method is familiar; what is new is the way AI is billed.

Allocate and reconcile

Begin with the invoices. An AI gateway measures requests and prices them at list or contracted rates. The invoice is what the company owes. The two will differ, because of cache discounts, credits, commitments and usage that never went through the gateway. Treat the invoice as the figure of record and keep the gateway's reading beside it, labeled, without adding the two together. Some cost cannot be sized yet, such as AI inside a SaaS product or a team paying on a card. Write each of those down with a reason and an owner, and leave it unsized until it can be measured.

Allocation is where AI differs most. The FinOps Foundation's guidance lists the problems: token prices are inconsistent, new SKUs arrive often and without tags, one model serves many interfaces, and there is no accepted way yet to split the cost of work done by several agents. [4] We suggest rules that carry each API key, application, agent and person to a cost center and a GL account, with an owner who revises them, the same way a cost-center hierarchy has one. Showback, where every department sees its line before anyone is charged, is the model the FinOps Foundation describes; we would start there. [4] Let a cost-center owner dispute a line, and require a reason with each dispute and with each answer.

Keep savings in separate tiers. A saving in a business case is a forecast. A saving the gateway measured is provisional. A saving finance has matched to the ledger is confirmed. Programs that add the three together can end up reporting returns no one can find in the accounts.

Then set the month's AI movement against the GL accounts it landed in, and ask of each line whether an approved business case claims it. The unclaimed lines deserve the most attention, because they are spend that went ahead with no recorded decision. Where a case did claim the spend, compare the month's actual cost per unit of work with the ceiling the case set before it started, and record the result, including a loss.

Finish with two people

Finish with two people. One proposes the close, a different person signs it, and both are recorded where an auditor can check them without asking the team that produced the numbers.

Finance leaders know where this gets hard. In Deloitte's Finance Trends 2027 survey of 1,434 finance leaders, uncertain regulatory or compliance requirements came first among the AI cost challenges Deloitte lists, at 20%. Complex cloud and vendor billing for AI compute followed at 19%, and integrating usage data with ERP and finance systems at 15%. The second and third are exactly what a close has to absorb. [2]

The approval side has moved faster than the close. KPMG's third-quarter survey found that 74% of US organizations with at least $1 billion in revenue now include cost reviews in AI approvals, up from 61% the previous quarter. A review at approval sets the expected cost. The month-end check is how anyone learns whether that cost is what landed. [3]

The billing format is changing as well. According to SiliconANGLE’s June 8, 2026 reporting from FinOps X, FOCUS version 1.5 is expected to include features that break AI spend down by token type and workload. [5]

A finance team can start small: one provider invoice and one month. Allocate it, set it against the ledger, and list the lines no business case claims. That list is a good agenda for the next AI budget review.