Before automating finance, define what “correct” means.

Financial work contains different kinds of correctness. A calculation can be exact while its explanation is misleading. A summary can be fluent while omitting an important exception. Evaluating an AI-assisted workflow means treating these as separate questions.

Separate calculations from interpretation

Use explicit, testable calculations for numerical operations. Where AI helps extract information or draft commentary, preserve the connection to the underlying records. Reviewers should be able to identify the source of a figure and reproduce its calculation.

Build an evaluation set

Choose authorised examples that represent normal work and difficult cases: missing entries, mismatched periods, conflicting documents and unusual transactions. Have qualified reviewers define the expected result before comparing approaches. Protect sensitive information and limit access to the people who need it.

Measure the cost of review

A draft produced quickly may still require substantial checking. Track total preparation time, corrections, unsupported statements and reviewer effort. Compare the complete process with the existing workflow rather than measuring generation speed alone.

Define when the system should stop

Missing evidence should trigger a request for clarification or human review. Establish which decisions remain with the finance team and what information must accompany an output before it can be approved.

This is a proposed evaluation framework, not a report of a completed Bolderank study or a validated financial product.