A controller reviewing an accrual support packet still has to trace the amount to source records, check the period and decide whether the evidence supports the proposed treatment. Generating the explanation is one step in that chain. If an AI workflow produces three drafts before the reviewer can accept one, the team has paid for all three, plus the time spent finding what was wrong. A cheap model call can be part of an expensive close. The invoice from the model provider gives the CFO only a small part of the operating picture.

The FinOps Foundation's unit-economics guidance distinguishes resource measures, including cost per token, from business measures such as cost per case resolved. For finance, apply that principle to a deliverable someone can accept: divide total workflow operating costs by the number of unique outputs meeting a defined acceptance standard. Include unsuccessful attempts, human review and rework in those costs, along with licenses, retrieval, tool calls, infrastructure, monitoring and ongoing support. Allocate shared charges consistently and count each expense once. This is an operating measure within the broader AI ROI calculation, with quality and turnaround reported alongside it.

Before running the pilot, the controller should approve the acceptance checklist. For an accrual packet, that might require the amount to tie to a frozen schedule for the correct entity and period, links to purchase orders and evidence of services received, explicit assumptions where an invoice hasn't arrived, and the designated reviewer's sign-off. The output is the reviewed support packet; journal approval and posting would need their own controls and costs if included. Save the checklist version with the result so next month's team can reproduce the test. A fluent explanation with an unsupported number stays in the review queue. Eloquence doesn't clear the close checklist.

Give each packet a stable identifier that survives retries, model changes and reviewer edits. Three attempts at the same packet can produce one accepted output. Keep the costs of rejected or abandoned work in the numerator, even when it contributes nothing to the denominator. Measure a defined intake cohort through a fixed cutoff, then show accepted, rejected and still-open items, including how long they've been waiting. Apply the same observation window to each approach. Excluding open cases and the work spent on them can make an easy subset look cheaper than the actual workflow. If nothing passes, report the spend and zero acceptances; a zero unit cost would be fiction.

Consider an explicitly hypothetical close cohort of 1,000 packets. The workflow makes 1,200 model calls costing $240, incurs $360 in allocated platform and support costs, and uses 40 hours of review and correction valued at a loaded $75 an hour. Total operating cost is $3,600. By the cutoff, reviewers have accepted 900 unique packets; 100 remain rejected or open. That's $4 per accepted packet, alongside a 90% acceptance rate. The model calls averaged 20 cents each. Those figures answer different operating questions, and the $3,000 assigned to employee time is an economic cost estimate, even if payroll doesn't change.

Compare that result with manual and rules-based workflows using the same acceptance checklist, cost scope and deadline. Group comparable work together: a routine recurring accrual with complete records belongs in a different band from a disputed service estimate spanning several entities. Hold the case mix steady or report each band separately. The same discipline applies to intercompany exception investigations. In the July 2024 preprint AI Agents That Matter, Princeton researchers argued for evaluating cost alongside accuracy. Their benchmark research supports that evaluation principle; it doesn't establish finance-department savings. A low blended average can reflect an easier workload while the underlying process hasn't improved.

Keep implementation spending visible in the investment case: initial integration, data cleanup, control design, testing and training. Report ongoing unit cost separately; if setup spending is also allocated across outputs, disclose the period and volume assumption. That gives the CFO a way to distinguish a process getting cheaper to run from a project still recovering its initial investment. Recovered capacity needs another line. Hours released for analysis can be useful without reducing expenditure, while realized cash savings require evidence such as lower contractor invoices or overtime payments. Counting the same salaried hours as both capacity and cash savings would give the budget committee a generous interpretation of arithmetic.

Next quarter's workflow review can then focus on specific changes: better retrieval to reduce unsupported amounts, deterministic checks before human review, or routing routine packets through a simpler process. Anthropic's December 2024 engineering guidance recommends adding complexity when the task requires it and recognizes the cost and latency tradeoff. The tooling has changed since publication, as the page now warns. Test each proposed change against the same accepted-output standard and workload bands, and watch what happens to review time, unresolved cases and cost. The next funding decision can follow the finished work all the way through the queue.