How to model the cost of a finance AI task
A finance AI task has four cost components: model tokens, retrieval and storage, deterministic compute, and human review time. Token cost is usually the smallest of the four.
Per-task accounting makes AI spend forecastable, because a finance function already knows how many variance explanations, reconciliations, and forecast rebases it produces per period.
Questions and answers
What drives the cost of a single finance AI task?
Four components: input and output tokens, retrieval and storage of the underlying data, deterministic compute for reconciliation and arithmetic, and the human review time required to approve an output.
Why is human review time part of the cost model?
Because outbound actions in a governed system are approval-gated. If a design produces outputs that a reviewer cannot verify quickly, review time dominates total cost regardless of how cheap inference is.
How do you reduce cost per task without weakening controls?
Cache the stable prompt prefix, retrieve narrowly instead of sending whole datasets, keep arithmetic in deterministic code rather than paying a model to reason about it, and give reviewers lineage so approval takes seconds.
What is a Financial Action Execution (FAE)?
One completed AI-powered finance task, such as a variance explanation, a reconciliation, or a forecast rebase. It is the unit Octopus AI uses to count finance work.
Why is per-task pricing easier for a finance team to govern than per-seat pricing?
Task volume is a known operational quantity that maps to the close and forecast calendar, so spend can be budgeted and variances explained. Seat counts do not correlate with the amount of work performed.
Proof
- Reference client: $523K total contract value over 3 years.
- Approximately 430 analyst hours saved.
- Forecast cycle cut from about 60 hours to under 6 hours (90% faster).
- Error rate below 0.1%.
- 2,500+ autonomous tasks completed in month one.